The field moves.
This book moves with it.

On September 20 Alibaba's Qwen team released Qwen-Image-2.1, one model for text-to-image and editing with 7 billion parameters in its visual generator, native transparent (RGBA) output and editing, subject extraction from photographs, and up to ten reference images, with local edits marked by circles, painted annotations, or masks. ComfyUI supported it on day one and noted that no other major open model outputs alpha, and the ostris ai-toolkit added LoRA training for it, including training with an alpha channel, the same day. On the Artificial Analysis text-to-image arena it is now the top open-weights model at 1042 Elo, eighteenth overall and about 150 points behind OpenAI's leader. The catch is in the LICENSE file. Every earlier Qwen-Image release shipped under Apache 2.0; this one ships under the Qwen Research License, for non-commercial purposes only, with commercial use by separate agreement. The weights are downloadable. The permission to sell what they make is not included.

Ch. 28→Ch. 143→Ch. 69→Ch. 76→Ch. 38→

On September 23 Google opened video generation in Google Vids to anyone with a personal Google or Workspace account at no cost, running Gemini Omni 1.1 Flash, the model that shares the top of the Artificial Analysis text-to-video board. Free generations are 1080p, scenes can be extended and clip durations set, every clip carries a digital watermark, and paid Google AI and Workspace plans get larger allowances; Google did not publish the free limits. The same model is sold per second through the Gemini API. Nothing about the model changed. What changed is that the price of a frontier-quality clip to a consumer is now zero, set by a company whose business is not selling clips.

Ch. 41→Ch. 129→Ch. 130→Ch. 93→Ch. 51→

On the Artificial Analysis with-audio text-to-video board as read on September 25, Google's Gemini Omni Flash sits first at 1233 Elo, Alibaba's Wan 3.0 at 1229, fal's post-trained MiniMax H3 Max at 1227, and MiniMax H3 at 1220, with confidence intervals of seven to nine points on each. A week earlier Wan 3.0 led at 1243. The order flipped, but the three leaders overlap inside their own error bars, so this is a tie that happens to be printed in a particular order. The one new name is Sand.ai's MAGI-2 Preview (0912) in sixth at 1156, listed with no API pricing yet. No new frontier video model was released in the window.

Ch. 69→Ch. 145→Ch. 65→Ch. 50→

Two routers shipped news in the same two days. On September 23 Comfy launched Comfy Router, one SDK and one key for Seedance 2.5, MiniMax H3, Nano Banana Pro, GPT Image 2, Kling, and Black Forest Labs models, where a model hosted on more than one of fal, Runware, WaveSpeed, and Higgsfield takes the provider as an explicit parameter, with no silent substitution and every job reporting which provider ran it. On September 24 Runway published a blind evaluation of its own Model Router on about 250 image-to-video prompts. Sending everything to Seedance 2.5 gave 78 percent usable clips at $1.80 each; the quality router gave 77 percent at $1.28; adding a one-dollar cap gave 74 percent at $0.61, 66 percent cheaper, almost entirely on Gemini Omni Flash; and the pure cost router gave 38 percent at $0.30 on Gen-4 Turbo. Runway says 64 percent of its router builders optimize for cost. These are vendor numbers, but they are the first published curve of usable output against spend across models, and the router sends most of the traffic to other labs.

Ch. 99→Ch. 43→Ch. 129→Ch. 88→Ch. 77→

On September 23 Runway released a free plugin for DaVinci Resolve Studio 19 and later on macOS and Windows. It generates inside Resolve, restyles the clip under the playhead with Aleph 2.0 using up to five keyframes while keeping the source duration, drops results straight into the Media Pool and timeline, and shows each generation's credit cost before you commit. With the Premiere Pro and After Effects plugins from September 8, Runway now has a panel inside Premiere Pro, After Effects, and Resolve. On September 18 it made credits fungible between its web app and its API, one pool on one invoice. Adobe's September Firefly update, posted the same week, added whole-clip video background removal and more aspect ratios and reference inputs for the partner models it hosts. The generator keeps moving to where the edit already is.

Ch. 98→Ch. 43→Ch. 91→Ch. 109→

ComfyUI v0.37.0 on September 21 added native Qwen-Image-2.1, turns on fast-disk loading automatically when it detects a fast drive, and added quantized and graph-captured paths for Qwen text encoders; v0.37.1 and v0.37.2 added partner nodes for Tencent's HY Image 3.5 Preview and Recraft V4.1 Flash. On September 22 Comfy published a fused encoder kernel and an int8 decoder for the MiniMax H3 video VAE that take a 1344 by 768, 129-frame encode and decode round trip on an RTX 5090 from 24.3 to 12.7 seconds. And on September 25 the ostris ai-toolkit added LoRA training directly on the 8-step distilled Fast H3 v2 through a training adapter, and switched all MiniMax adapters to ones that no longer need contrastive guidance. That is a trainer-side answer to the problem the book described last week, adapters trained on the full step schedule degrading on a distilled model. It is commit-level evidence, with no release post.

Ch. 76→Ch. 78→Ch. 38→Ch. 136→

On September 23 Recraft launched V4.1 Flash, a raster-only 1K tier with no vector or 2K variant, at $0.007 an image in its API, which Recraft calls the lowest price it offers, and about 1.3 seconds end to end. Recraft markets it as the fastest model on the market, though its own latency table shows FLUX.2 klein 9B at the same 1.4-second median, so read it as tied for fastest. A day earlier Tencent released the Hy Image 3.5 preview, one model for generation and multi-turn editing at native 2K that ComfyUI wired in as a partner node with up to five reference images. Its predecessor Hy Image 3.0 shipped open weights; as of September 25 no 3.5 weights have appeared, and Tencent's quality comparisons against Seedream and Nano Banana are its own. Both releases point the same way: the cheap tier gets cheaper, and the new tier ships as an API first.

Ch. 25→Ch. 31→Ch. 129→Ch. 64→

On September 23 Google released Gemini 3.8 Flash TTS and Flash-Lite TTS in the Gemini API and AI Studio, with Flash-Lite also going into Google Vids. Flash designs an original voice from a natural-language description, takes line-by-line stage directions for pacing, tone, dialect, and backchanneling, and covers more than 100 languages. It can replicate a voice from a 30-second sample, but only after the user supplies a verbal consent recording from the voice owner that matches the reference speaker, and every clip is watermarked with SynthID. Voice replication in AI Studio is not available in Illinois, Texas, the EEA, the UK, Switzerland, or India. Google cites first place on Hume AI's Voice Design Benchmark at 71.4. The consent step is the notable design choice; the geography list reads like a map of where voice and biometric law already bites.

Ch. 54→Ch. 66→Ch. 59→Ch. 87→Ch. 107→

At Made on YouTube on September 23, YouTube said its likeness-detection tool, which scans uploads for enrolled creators' faces, will start integrating speaking-voice detection, and that enrollment and match alerts are coming to mobile. Around September 24 Qobuz began tagging tracks its system identifies as AI-generated and removing them from editorial and algorithmic recommendation, and published the numbers behind it: 0.38 percent of its streams come from such tracks, and it flags 60 percent of the streams on them as fraudulent. And the 2027 Sanremo festival rules, published September 22, require entries to be an expression of human creativity, ask artists to sign a declaration, and let the broadcaster use technical means to check. Three different gates, one direction: the unit of protection is widening from the face to the voice and from the file to the chart.

Ch. 125→Ch. 59→Ch. 93→Ch. 135→

Answering the second UMG and Sony suit on September 22, Suno called the claims fundamentally flawed and said v6 was trained on content licensed from its partners plus interactions including creations and preference signals from its community, which places users' own generated songs inside the training data of a model the labels say descends from unlicensed ones. The same day, lawyers for Michael Smith, who admitted collecting more than $8 million by bot-streaming AI-generated songs, asked for probation, arguing no individual artist suffered perceptible harm. On September 24 sound designers and effects-library companies launched the Professional Sound Alliance against the scraping of sound libraries into commercial models, and on September 23 Sony Music became the first music company in the ARIAM policy coalition alongside Disney and the BBC. In the studios' case against Midjourney, the judge on September 23 moved the hearing on the pending motion for judgment on the pleadings to November 23.

Ch. 59→Ch. 55→Ch. 131→Ch. 118→Ch. 57→

Three papers treat the forgetting world model as the problem to solve. WorldCrafter (September 21) gives a video world model a camera-queryable implicit 3D memory trained jointly with the generator; GAE (September 21) swaps in a geometry-native latent with the generator held fixed and cuts FVD by 12.7 percent on RealEstate10K and 23.1 percent on DL3DV while halving camera-trajectory error on RealEstate10K; and a September 23 paper releases a 1.5 million sample corpus and a 300-question exam for object permanence, with a 16B world model trained on AWS Trainium2. On joint generation, a September 23 paper finds audio attends strongly to video while video barely attends back, and a regularizer that strengthens the weak direction cuts audio-video desynchronization from 0.804 to 0.752; AV-GRPO (September 24) trains one modality at a time with the other frozen. A September 20 study traces physics failures in video diffusion to RoPE, which decays spatial attention so fast that early motion locks into implausible positions. And ViRDM (September 24) distills a few-step causal video model with no teacher or critic in 16 A100 GPU-hours.

Ch. 140→Ch. 141→Ch. 58→Ch. 136→Ch. 16→

As of September 18 a FLUX.3 Edit Video [FAST] endpoint is live on fal, described as Black Forest Labs' frontier video model, editing an existing video from natural-language instructions and applying targeted changes while preserving the rest of the scene, at 3 cents per second of generated 720p video. Black Forest Labs' own release notes end at the September 10 entry adding 2K and 4K output to FLUX 3 Video, and still carry the earlier line saying that video editing and Omni Reference with images and videos will be available soon. Neither company has published a dated launch post, so treat the ship date as unpinned. What is verifiable is the present state, and the present state is that the capability is purchasable on an aggregator before the lab's own changelog acknowledges it exists. For anyone tracking releases, that inversion matters: the marketplace is now a faster release feed than the lab.

Ch. 60→Ch. 21→Ch. 99→

On September 17 Suno updated Suno Studio with MIDI-to-audio transcription that is faster, carries more musical context and is more accurate to the original, the ability to ask the chat bar to create new audio stems from a MIDI clip, more predictable results when prompting MIDI from scratch, and clip renaming individually or in bulk. On September 14 ElevenLabs' changelog recorded that its music endpoints and SDKs now accept music_v2_5, adding the model announced three days earlier to video-to-music, composition planning, streaming and fine-tuning, with composition plan chunks raised to 6,132 characters. Both moves point the same direction. The unit of control in music generation is sliding away from the finished track and toward the editable part, which is the same road image and video tools walked when they stopped being one-shot generators and grew layers, masks and reference inputs.

Ch. 55→Ch. 53→Ch. 107→Ch. 92→

An omni-modal evaluation of MiniMax H3, arXiv 2609.18323, posted September 16, builds four scenarios in which no single modality carries enough evidence, so the model has to combine cues to infer what is happening and what happens next. Across 517 instances it reports an overall success rate of 41.97 percent, with video-based decision reasoning strongest at 56.00 percent and audio-based disambiguation reasoning weakest at 27.40 percent. That gap is the finding worth keeping: a model with one unified audio-visual latent is not yet reasoning across the two, it is mostly reasoning from the pictures. Two more papers fill in the audio evaluation picture. TTM-Bench, arXiv 2609.18585, scores text-to-music systems on musical content alignment and on computational cost as separate axes, reporting that higher alignment does not systematically coincide with lower cost, which is why it refuses to publish one blended number. SongCraft, arXiv 2609.16315, folds generation and editing into a single latent flow-matching model trained to reconstruct audio from a varying number of interpretable conditions, so editing lyrics, vocal melody, beats or singer identity becomes a matter of changing one condition and holding the others fixed.

Ch. 87→Ch. 142→Ch. 58→Ch. 68→

Three world-model items arrived in four days. On September 15 Odyssey published Odyssey-3, an autoregressive diffusion transformer it describes as a foundation world model trained on a vast collection of visual observations, which a small learned action decoder adapts to control robot arms, humanoids, cars, drones and game agents; the company cites 20 hours of simulated driving data for the driving case and promises a public release in the coming weeks, with no weights or API yet. Zing-0.5, arXiv 2609.17909, also September 15, is a 5B model taking joint keyboard and live text control, running 832 by 480 at 24 FPS via four-step generation and context-preserving streaming, and it does something the field has mostly avoided: it publishes a serving cost, roughly nine tenths of a cent per stream minute. ModAR, arXiv 2609.17524, is the contrarian note, finding that world-action models benefit from predicting point tracks, DINO features and depth maps while additionally predicting future RGB gives no consistent benefit, and reporting a higher average success rate than a video-model-initialized baseline using roughly 20 times fewer training FLOPs and no pretraining. Read together: one company claims generality, one paper prices it, and one paper suggests the expensive pixel-prediction pretraining may not be the right foundation at all.

Ch. 141→Ch. 140→Ch. 94→

Two papers in the same week take apart the cost of fast video. DART, arXiv 2609.20051, posted September 17, starts from a problem the LoRA chapters should have seen coming: step distillation reduces the cost of video generation, but a LoRA trained against the longer trajectory can behave differently or degrade quality when reused on the shortened schedule. The fix is training free, combining low-rank coordinate transport with target-schedule response calibration using only forward evaluations, and on a four-step target it moves the joint quality score from 0.9029 to 0.9227 and macro functional retention from minus 0.4644 to plus 0.1349. The authors are careful to say this is not universal recovery: some adapters gain function and others are strongly attenuated, so reuse has to be judged adapter by adapter. Video DeltaNet, arXiv 2609.20744, also September 17, attacks the same cost from the other side, pairing local softmax attention with a bidirectional linear memory inside MiniMax H3 and reporting that it denoises a 14.3 second 768p video in 6.70 seconds on eight B200s, a 14.5 times speedup over the 50-step dense baseline.

Ch. 136→Ch. 12→Ch. 39→

fal's 75 percent launch discount on H3 Max ended on September 14 exactly as its page said it would. The standing rates are now 5 cents a second at 480p, 8 cents at 768p and 16 cents at 1080p, with H3 Max Turbo billing at half of that at every resolution, and five free generations a day for signed-in users. Two days later fal published the build log. The post training ran on a cluster of interconnected GB200 NVL72 nodes on fal Compute, the final run took on the order of a week to ten days with considerably more time going into the experiments that shaped it, and no latency optimization shipped unless the model held its rank in fal's quality evaluations. An inference host describing its own training substrate is a plain statement about where the margin is: serving other people's weights is a commodity, and post-training them is not.

Ch. 129→Ch. 101→Ch. 99→Ch. 65→

The two pieces of open tooling this book treats as load bearing both widened in the same window. ComfyUI shipped v0.36.0 on September 15, adding Marigold v2, which estimates depth, surface normals and albedo from a single image, a Generic Loops implementation that lets a graph iterate over heterogeneous lists natively, a Video Concatenate node, and support for the Yue2 music model. Meanwhile ostris' ai-toolkit, the default LoRA trainer for FLUX and Qwen-Image, committed support for training VLLMs, omni models and LLMs on September 15 and Qwen 2.5 Omni with LoRA loading for captioning models on September 16. Where the rest of its week went is the striking part. Nearly every other commit in the window is music work: experimental Yue2 support, sheet music conditioning applied half the time so the model learns to work with and without it, a MOSS music captioner, and vocal separation. The canonical image LoRA trainer spent a week becoming a music trainer, which is what a pillar looks like from inside a repository rather than from a press release.

Ch. 14→Ch. 76→Ch. 75→Ch. 38→Ch. 92→

On September 16 Governor Newsom signed SB 1050, chaptered the same day, which makes it unlawful to publish a video or audio advertisement prominently featuring a synthetic performer without a clear and conspicuous disclosure worded substantially like this performance features a synthetic performer, or no human performer is depicted. The statute defines a synthetic performer as a digital figure, voice, or representation created in whole or in part using generative AI that creates the realistic impression of a human performance by someone not recognizable as any identifiable natural person, and it exempts advertisements for expressive works such as films, television and video games, along with AI used solely for translation or accessibility. Two design choices are worth noting. The duty attaches to advertising, which is the narrowest and most commercial slice of synthetic media and therefore the easiest to police. And it covers voice as squarely as face, which pulls synthetic speech into a disclosure regime that the likeness laws written for on-camera performers had largely left alone.

Ch. 125→Ch. 131→Ch. 135→Ch. 118→Ch. 59→

Three things happened in four days, and all of them landed on distribution rather than on generation. On September 15 UMG Recordings and affiliated labels sued the distributor DistroKid in the District of Delaware, case 1:26-cv-01156, docketed as a copyright action; trade coverage of the complaint describes claims under Delaware's deceptive trade practices act alongside direct and vicarious infringement, and reports the filing alleging that DistroKid floods streaming services with AI-generated tracks and keeps distributing recordings it has already been told are infringing. On September 17 404 Media published a reporter's account of generating a song in Udio and getting it onto a real Brooklyn band's verified Spotify, Apple Music, Tidal and Amazon Music pages without the band's knowledge, by signing up with a digital distributor and claiming to be the artist. On September 18 UMG, Sony and affiliated labels filed a second complaint against Suno in the District of Massachusetts, case 1:26-cv-14275. The pattern is the one this book keeps finding. The model is hard to sue and impossible to switch off, while the distributor has a name, an address, an ingestion pipeline, and a duty it can be held to.

Ch. 59→Ch. 93→Ch. 125→Ch. 131→Ch. 55→

On September 15 Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, native-audio conversational models available in the Gemini API and Google AI Studio. Google says the Extended Thinking variant reasons and speaks at the same time, takes the top overall spot on Artificial Analysis' Speech to Speech Quality Index at 82.6, scores 97.7 percent on Big Bench Audio and 68.6 percent on the tau-Voice agentic task benchmark, and detects and switches between 97 languages mid conversation. Both models can execute API and tool calls in the background while continuing to stream audio, at five tenths of a cent per minute of audio input and 1.8 cents per minute of output. The dating is the interesting part. OpenAI made GPT-Live 1 generally available on September 10, and Google shipped the same architectural idea five days later. The conversational layer is collapsing into one always-listening model, with reasoning demoted to something that happens behind it, and two labs arrived there in the same week.

Ch. 54→Ch. 66→Ch. 94→Ch. 142→

Reve is shutting its image generation service down. Its help center now carries a sunset notice giving users until September 27 to create and until October 31 to download work they already made. Reve was the layout-first lab, the one that treated an image as something to plan like code before rendering it, and its Reve 2.1 sat second or third on the Artificial Analysis text-to-image board through the summer. In July OpenAI invested in the company and members of its AI research team joined OpenAI to help lead multimodal research, with Reve saying at the time that it would continue to operate independently. The shutdown lands three days after the same thing happens to Sora, whose Videos API and every Sora 2 alias are removed on September 24. Two products this book has repeatedly cited near the top of their boards are being switched off in the same week, neither of them for lack of quality. Krea had already lost Reve's models on August 31, which is the practical version of the same lesson: an aggregator's catalogue is only as durable as the labs inside it.

Ch. 31→Ch. 130→Ch. 32→Ch. 40→Ch. 30→

On September 14 the Manhattan District Attorney's office seized the domains of twelve websites used to publish and sell non-consensual sexual deepfakes, which the office describes as the largest known seizure of AI-generated celebrity deepfake sites to date. It says people under investigation used AI image and video tools to turn existing photographs and video of roughly 1,200 real people into what appear to be hyper-realistic sexual images and videos, that the victims are overwhelmingly women and mostly public-facing, including actors, politicians, athletes, musicians, and advocates, and that the seizure was carried out under a court order with the investigation continuing and the office monitoring for the sites to return under different domains. Note where the intervention landed. Nothing here attempts to stop a model from generating this material. The enforceable point was the domain name, the one link in the chain with an owner, a registrar, and a jurisdiction, which is the same reason provenance and distribution rules keep proving more actionable than generation rules.

Ch. 125→Ch. 135→Ch. 131→

On September 9 Amazon turned on lip-sync dubs for Prime Video, starting with both existing seasons of Maxton Hall, globally, in English, with the final season to include it on release in December. The mechanism runs opposite to what the phrase suggests. The dub is still performed by human voice actors, and AI and visual effects alter the on-screen actor's mouth in post-production to match that human recording, under what Amazon describes as creative oversight. Amazon says it plans to extend the technique to more titles. The technical capability is not new, since lipsync has been good enough for advertising and social video for years. What is new is the venue: a tier one streamer applying it to a flagship title makes reshaping a performer's face a normal step in localization, which puts a consent question the dubbing and performance industries have not settled onto a very large distribution surface.

Ch. 61→Ch. 93→Ch. 123→Ch. 125→

No frontier video model was released between September 5 and September 14. What moved was commercial. fal's promotional rate on H3 Max ran at 75 percent off through September 14, with 768p at a cent a second, and its own page states that prices quadruple after September 15. OpenAI's deprecations page confirms the Videos API and every Sora 2 alias and snapshot are removed on September 24 with no recommended replacement listed, which means OpenAI will sell no video generation at all after that date. Runway announced on September 8 that the team from the Paris lab Kinetix is joining it, pointed at world models for robotics. The standings barely moved: Wan 3.0 leads text-to-video at 1243 Elo, Gemini Omni Flash is at 1238, H3 Max at 1231, while on image-to-video H3 Max leads at 1206 and Veo 3.1 has slipped to eleventh at 1088. Three claims circulating about this window are false and worth naming: there is no Veo 3.2 on DeepMind's model page, the Wan-AI repository has published no 3.0 open weights, and a press-wire item dating a Kling 3.0 launch to September 11 is a republication of a February 2026 announcement.

Ch. 40→Ch. 69→Ch. 50→Ch. 129→Ch. 51→

Four papers in this window matter for people who generate for a living. Mask Forcing, posted September 8, argues that the over-saturated and over-smoothed look of distilled autoregressive video is not a tuning problem but a consequence of the reverse KL objective being mode seeking, and fixes it by perturbing the student's own rollout with random spatial and temporal masks, reporting gains across several distillation methods with no real video data and no extra post-training stage. Marigold V2, also September 8, repurposes pretrained flow-matching image transformers into single-step depth, normal, and albedo predictors, reporting 16 to 26 percent better AbsRel than the previous best on KITTI and ETH3D and resolving hair-thin structure. SenseNova-U1.5, September 10, is an 8 billion parameter unified model that understands and generates with no vision encoder and no VAE, working directly in pixel space. Programmable World Model, September 9, compiles instructions into executable programs that hold entity state and hands the video model 3D oriented bounding boxes, so a program decides what is true and the generator only decides what it looks like. The meta-observation is the delivery time: Marigold V2 was merged into ComfyUI three days after it posted, and SenseNova U1.5 support shipped before the paper did.

Ch. 136→Ch. 140→Ch. 18→Ch. 16→Ch. 62→

ComfyUI v0.35.0, released September 9, put a compiler under the graph. The Comfy Compiler pairs an ahead-of-time memory compiler with the CUDA graph capture added in August, so VRAM allocation calls collapse to a one-time setup and the physical peak matches the logical peak instead of being inflated by the allocator caching freed blocks, which matters most on the consumer cards where cached allocations decide whether a workflow runs at all. The same release added a sparse attention node, parallel-decoding-distillation LoRA support for MiniMax H3, pixel-space generation for SenseNova U1.5, AVIF output, HDR conversion, and attention and FP8 support for AMD RDNA3.5. Mainline then merged the open YuE2 music model and Marigold V2 depth within days. On the training side, ostris merged a resident inference engine into AI Toolkit between September 10 and 12, a generation server that runs as a toolkit job, serves every registered image, video, and audio architecture, and hot-swaps models per request while keeping shared components resident, with a token-authenticated local HTTP API. Both projects are converging from opposite directions on the same thing, a machine you leave running rather than an application you open.

Ch. 76→Ch. 77→Ch. 78→Ch. 38→Ch. 99→

On September 10 OpenAI made GPT-Live 1 generally available in its API at five cents per minute of voice session, billed by the second, two months after the model started powering ChatGPT Voice. The model is full duplex, listening and speaking at the same time rather than taking turns, and it can hold the conversation while a separate backend model or agent does the reasoning and the tool calls, with backend and tool usage billed separately. That is an architectural change rather than a quality bump. The standard voice agent has been three models in a chain, speech to text into a language model into text to speech, and most of the awkwardness people notice, the pause before a reply and the trampled interruption, came from the handoffs rather than from any single component. Putting the conversational layer in one always-listening model and demoting reasoning to something that happens behind it is the same separation of concerns that routing products are making elsewhere in the stack.

Ch. 54→Ch. 66→Ch. 94→Ch. 142→

Two court documents surfaced in the same fortnight as the licensing deals and complicate them. On September 1 Suno answered the amended complaint in the Massachusetts case brought by UMG Recordings, Capitol Records, and Sony Music, admitting that it obtained audio data from YouTube for use as training data and naming yt-dlp as the tool, while pleading as affirmative defenses that the plaintiffs lack both constitutional and statutory standing to bring the circumvention claim. The admission is narrow and each instance is paired with a denial of the surrounding legal conclusions, but yt-dlp is the utility the complaint identifies as defeating YouTube's protections. Separately, the American Federation of Musicians is suing Warner Records, Atlantic, and Universal, not an AI company, under section 301 of the Labor Management Relations Act, arguing the labels licensed recordings made by union musicians to Suno and Udio without compensating them or disclosing the terms required by the collective bargaining agreement. The union filed its opposition to the motions to dismiss on September 4 and the labels replied on September 11. Together they show the fight relocating: once a label licenses rather than litigates, the disputed money becomes revenue it already holds, and the argument moves from copyright to labor contract, with performers opposite their own labels.

Ch. 59→Ch. 118→Ch. 55→Ch. 131→Ch. 122→

On September 8 Believe and its self-release platform TuneCore signed an opt-in global partnership with Suno covering participating artists and labels for a revenue share, running in both directions, so tracks made with Suno's licensed model become eligible for distribution through Believe and TuneCore. Believe had refused to distribute Suno-made music earlier in the year. On September 9 Suno released the v6 family, the flagship v6, an exploratory v6-wild, and a free v6-mini, built with Warner Music Group, BMG, and Believe, saying v6 was not trained on the data behind earlier versions, that rights holders are paid when a user generates with it, and that every previous model will be retired. On September 10 Universal Music Group signed a multi-year agreement with ElevenLabs, its first with a major voice AI company, beginning with a licensed platform where fans remix and reinterpret the music of artists who opt in. ElevenLabs then shipped Music v2.5 on September 11 as its default music model, a separate and unlicensed-repertoire product. The pattern across the four days is that the majors have stopped trying to remove AI music from the market and started metering it, and that the one deal struck from strength rather than settlement went to the company they had never sued.

Ch. 59→Ch. 55→Ch. 53→Ch. 57→Ch. 93→Ch. 66→

Four products shipped in one week that all move generation out of its own website. On September 8 Adobe put Generate Video on the Premiere Pro timeline, filling gaps using the surrounding project as context, alongside Generate Sound Effects and beta Generate Soundscape and Generate Music, and gave After Effects an AI Assistant, with generations routable to Veo, Kling, Runway, or Luma rather than Firefly alone. The same day Runway shipped plugins that run its tools in a panel inside Premiere Pro and After Effects, and Blackmagic released DaVinci Resolve 21.1. Also on September 8, Krea released an agent that takes a brief and picks among more than 150 image, video, and audio models on the user's behalf, billed against ordinary credits. On September 14 ElevenLabs put voice, music, image, and video generation behind its MCP connector, drawing on more than fifty models from inside Claude, ChatGPT, and Cursor. The common thread is that the model is becoming something a layer above it selects, inside a tool you were already using, which is a different competitive game than being the best model.

Ch. 98→Ch. 88→Ch. 91→Ch. 43→Ch. 30→Ch. 53→

On September 8 OpenAI released ChatGPT Images 2.5 to every ChatGPT tier including the free one, and split the API line into gpt-image-2.5-flare, the faster default, and gpt-image-2.5-sunburst, the slower and more precise option. It adds a Sketch tool that turns a drawing made in ChatGPT into the compositional template for a render, and claims up to half the latency of Images 2.0 along with better subject preservation from reference photos. The commercially significant change is the quality ladder: the API now exposes five settings instead of three, with xhigh and max above the previous high ceiling at unchanged token rates of 8 dollars per million image input tokens and 30 dollars per million output tokens, which puts the top of the ladder at several times the old top price. Within days both variants held first and second on the Arena text-to-image board and on Artificial Analysis, displacing Microsoft's MAI-Image-2.6, which had claimed the number two slot four days earlier. Adobe shipped it in Firefly the same day and Recraft added it the next, which is its own signal: the gap between a frontier model and the products that resell it is now measured in hours.

Ch. 23→Ch. 69→Ch. 64→Ch. 129→Ch. 145→

On September 3 Nvidia confirmed it will acquire Hugging Face for 12.93 billion dollars. Hugging Face hosts roughly three million models and half a million datasets for more than eighteen million developers on about 150 million dollars of annualized revenue, and Jensen Huang said the platform stays open, with developers choosing their own models, frameworks, clouds, and inference providers, and Nvidia compute not required to build or deploy. Take that commitment at face value and the structural fact stands anyway: nearly every open model this book covers, the FLUX dev checkpoints, the Stable Diffusion lineage, Wan, LTX, MiniMax H3 and Music 3, Ideogram 4.0, and the whole fleet of community quantizations and LoRAs built on them, is distributed through one company, and that company now belongs to the vendor selling the hardware they run on. The open-weights tier always had a single point of failure in distribution; this is the week it became a strategically interesting one.

Ch. 100→Ch. 143→Ch. 130→Ch. 93→

On September 1 the United States filed in the consolidated OpenAI copyright litigation in the Southern District of New York arguing that broad copyright liability for model training would be legally incorrect. Read its scope before reading it as relief: the brief is written about large language models and written works and never analyzes image, video, or audio training. Meanwhile the plaintiffs changed shape. Sony Music Publishing sued Anthropic in the Northern District of California on August 28 (5:26-cv-09217). On August 31 Gerencia 360 Music sued Suno in the same district (3:26-cv-09248), again naming the scraping vendor Bright Data, the second suit in a fortnight to put the collection supply chain in the caption. The same day four recording artists led by David Lowery and Jason Isbell sued Suno in the District of Massachusetts (1:26-cv-14005), docketed under Other Statutory Actions rather than Copyright, a case built on identity and voice that routes around the fair-use fight entirely. A third Suno action followed on September 3, and on September 2 Canada's SOCAN became the third collecting society to sue, on a claim aimed at 150 specific outputs rather than at training. Separately, the Carl Sagan estate sued Luma AI over the advertising for a generative video model, xAI drew a cluster of civil actions over alleged generation of sexual imagery of identifiable minors, and the studio case against Midjourney produced a discovery order requiring production of samples of private stealth-mode subscriber prompts.

Ch. 131→Ch. 59→Ch. 125→Ch. 135→Ch. 118→

AI-native video reached a scheduled television shelf. Fairground Entertainment, founded by an Xumo co-founder and funded with about four million dollars led by Viant Technology, launched what is described as the first free ad-supported channel built entirely from AI-generated content on The Roku Channel's live guide, programmed like a traditional schedule from more than a hundred AI creators, with AI-generated commercials in the breaks; the reviews were hostile, but a connected-TV slot is a different and more sellable shelf than an algorithmic feed. On governance, 404 Media obtained Moonbug Entertainment's internal AI guidance for the Cocomelon and Blippi studios, which bars AI from originating characters, core storylines, and song lyrics while permitting ideation, storyboarding, design, and background elements, and requires per-tool legal approval, prompt logging, and separate file management for AI and human assets. Utopai Studios and Japan's DeNA began an evaluation-phase collaboration putting Utopai's PAI system into DeNA animation workflows with DeNA keeping creative authority. And Suno pulled an advertisement that made Mary J. Blige appear to endorse it after learning it had done the deal with someone falsely claiming to be her representative, which sets the failed-consent case directly against Netflix's consented Gene Wilder voice recreation from the fortnight before.

Ch. 93→Ch. 100→Ch. 117→Ch. 125→Ch. 59→

Two papers a fortnight apart push on-policy methods into image distillation. DiffusionOPSD converts image-level reward signals into explicit targets for clean-output predictions at sampled queries against a frozen behavior policy, and Self-OPD removes the teacher entirely for flow-matching models by branching each timestep into stochastic candidates and scoring them against a deterministic self-reference, so the student's own exploration becomes the supervision. The same vocabulary is already in the tooling, with ostris's ai-toolkit adding a D-OPSD bleed loss on August 29. On controllability, The Missing Temporal Link diagnoses why script-driven audio-video generation stays internally in sync while drifting off the script's timeline, and reports shot-boundary error falling from 1.11 seconds to 0.042 seconds and dialogue accuracy within half a second rising from 28.3 to 84.1 percent. On evaluation, HarnessEval-W and SemComp-Bench both attack the scalar preference score, the first with an agentified pipeline that returns an inspectable reasoning chain and the second by scoring whether a generation achieved the intended outcome while staying semantically grounded.

Ch. 136→Ch. 12→Ch. 138→Ch. 68→Ch. 17→

On September 4 Google moved Lyria 3.5 out of Flow Music exclusivity into the Gemini app on web and mobile, the Gemini API, AI Studio, and Google Vids, with genre and vocal-or-instrumental selection and a choice of track length; the model did not change, the distribution did. On August 26 Google folded speech-to-text into the Gemini line with Gemini 3.5 Transcribe, in recorded and streaming variants, which resolves self-corrections and drops filler words rather than transcribing them literally, covers more than eighty-five languages, diarizes up to three speakers, and reports a 2.6 percent word error rate non-streaming and 4.0 percent streaming as measured by Artificial Analysis. A ComfyUI community challenge that ran from August 20 to September 1, requiring audio and video to come from the same MiniMax H3 generation, produced the first substantial craft report on native sync sound: strong on in-shot sound effects, weak on dialogue and singing, lip sync that generalizes past humans, and drift across cuts controlled by chaining segments against a recorded reference track. ElevenLabs shipped no model and instead added asynchronous Flows APIs, workspace asset APIs, and a CLI. Mureka released V9.5 on vendor-reported numbers only.

Ch. 57→Ch. 54→Ch. 58→Ch. 53→Ch. 66→Ch. 107→

ComfyUI's own kernel library, comfy-kitchen, is now a pinned requirement of the base install alongside the comfy-aimdo offloading allocator, with a node that lets you pick the attention backend per model; it carries eager, CUDA, Triton, and AMD HIP paths and quantization from FP8 and NVFP4 down to INT8 and four-bit, which is why fast paths for new video models now land in core rather than in third-party wrappers. On August 18 Comfy open-sourced Comfy MCP for local machines, so an agent reads your actual GPU and your actual installed nodes before deciding whether to run locally or in the cloud. On August 27 Comfy became the first official reseller of MiniMax H3 commercial-use licenses, which turns the commercial right to an open-weights model into a product sold through the community distribution channel. And on August 31 Civitai retired the flat 25 percent creator share on generations and let creators set their own per-image licensing fees, with fine-tune fees stacking on base-model fees.

Ch. 76→Ch. 14→Ch. 77→Ch. 143→Ch. 38→Ch. 95→Ch. 99→

A refresh for the first days of September 2026. On standings: Alibaba's Wan 3.0 and Google's Gemini Omni Flash are tied at 1238 Elo at the top of the Artificial Analysis with-audio text-to-video board, with fal's post-trained MiniMax H3 Max at 1235 and MiniMax H3 at 1228, while H3 Max leads the with-audio image-to-video board at 1200. On images, OpenAI's GPT Image 2 still leads text-to-image at 1178 and Microsoft's MAI-Image-2.6 has taken the editing board at 1122, the same week Microsoft put it into Microsoft Foundry in public preview alongside a faster Flash tier. The durable point is a new one: open weights sit ten Elo points from the video frontier and about a hundred and sixty behind on still images, a split that tracks who benefits from publishing rather than how much compute a medium takes. Elsewhere in images, Midjourney opened its first V8.2 edit model to everyone, Black Forest Labs shipped no image release at all, and Krea gained a ChatGPT plugin while losing the Reve models and the older Veo tiers. Model standings move weekly; treat any single ranking as a snapshot.

Ch. 69→Ch. 64→Ch. 31→Ch. 26→Ch. 21→Ch. 30→Ch. 24→Ch. 43→Ch. 143→

A refresh for the second half of August 2026. On models: Alibaba shipped a real Wan 3.0, hosted and closed, that takes documents and web pages as generation inputs and went to number one in the Artificial Analysis with-audio video bracket, while Google's Gemini Omni 1.1 Flash added forty-second scene extension and first-and-last-frame control. On the open tier: MiniMax H3 became the checkpoint the field built on, distilled to four steps by FastVideo and post-trained by fal into a variant that outranks the original, but its community license excludes the EU, the UK, South Korea, and the United States, and the derivatives inherit that. On rights: Round Hill sued Suno and Anthropic and named a scraping vendor as a defendant, a Boston judge let stream-ripping claims proceed, and all three major labels took equity in Stability AI in the same fortnight. Spotify, Apple, ARIA, and Luminate all began labeling AI music, each targeting something different. Google made its visible watermark optional while keeping SynthID and C2PA. Google's Imagen line shut down for good on August 17. Model standings move weekly; treat any single ranking as a snapshot.

Ch. 45→Ch. 51→Ch. 44→Ch. 41→Ch. 42→Ch. 43→Ch. 69→Ch. 22→Ch. 32→Ch. 25→Ch. 55→Ch. 57→Ch. 54→Ch. 59→Ch. 76→Ch. 136→Ch. 140→Ch. 143→Ch. 131→Ch. 135→Ch. 123→Ch. 125→

Four papers from the week are worth a practitioner's attention. An empirical study of pixel-space text-to-image diffusion proposes learning generative priors in latent space and then transitioning to pixel space during post-training, reporting models that match or beat their latent-space counterparts while running 3.18 to 4.75 times faster end to end. V-RAE rebuilds the video latent space on frozen vision foundation model features rather than pixel reconstruction, converging up to roughly six times faster than conventional tokenizers. Together they reopen the VAE as a live design choice rather than settled plumbing. On the character side, LiveAnimate streams human animation at about 19.63 frames per second on two H100s from a 14B model while holding identity nearly constant across a three-minute sequence, and UniSwap swaps face and voice together with one audio-visual diffusion transformer, cutting sampling from thirty denoising steps to three per block. Context-Matched Distillation argues that distilling autoregressive video students from bidirectional teachers misaligns supervision with what the student could actually see, and fixes it with Prefix Scoring.

Ch. 07→Ch. 136→Ch. 137→

The Motion Picture Association and ByteDance announced a global agreement on August 17, 2026 to protect intellectual property on ByteDance's AI video and image models, the first such deal between Hollywood and a major model developer, following an MPA cease-and-desist in February 2026. It governs what the models output and does not settle the separate question of what they trained on. The same day Round Hill Music sued Suno in the Northern District of California, pleading copyright and DMCA claims and, unusually, naming the data-scraping vendor Bright Data as a contributory infringer for supplying the proxy networks and scraping tools, which extends AI training litigation into the collection supply chain for the first time. Disney, Warner Bros., and NBCUniversal opposed Midjourney's motion to narrow their case on August 12. On provenance, Google open-sourced Credentio for validating C2PA Content Credentials locally, then made the visible watermark optional in Gemini and Flow while keeping SynthID and Content Credentials intact. Licensed remixing kept expanding, with Spotify signing Kobalt and Universal Music Group licensing its catalog to the remix app Hook.

Ch. 123→Ch. 118→Ch. 131→Ch. 135→Ch. 59→

A weekly refresh for mid August 2026. On access: MiniMax open-sourced its 33B omni-modal H3 with native stereo audio and day-one ComfyUI support, Lightricks released open-weights LTX-2.5 that turns a still into a 10-second clip in about 6.8 seconds, and ByteDance opened the Seedance 2.5 API on Volcano Engine (native 30-second single-take video, up to 50 reference inputs). On standings: Gemini Omni Flash leads the Artificial Analysis Video Arena with MiniMax H3 at number two, while OpenAI's gpt-image-2 still tops the image arena and xAI's Grok Imagine Image 2.0 claims number two after an editing-first release. On rights: Suno signed a global licensing deal with BMG and committed to tamper-resistant watermarking and download caps, the clearest sign yet that AI music is trading its Wild West era for licensed catalogs and provenance. Model standings move weekly; treat any single ranking as a snapshot.

Ch. 44→Ch. 46→Ch. 49→Ch. 59→Ch. 60→Ch. 76→

A weekly refresh for early August 2026. On the law: the Munich Regional Court found Suno liable in GEMA's copyright suit (July 31), and the EU AI Act's Article 50 synthetic-content transparency rules took effect (August 2). On the models: MiniMax's omni-modal Hailuo H3 generates 2K video with native stereo sound in one pass, xAI's Grok Imagine 1.5 added references, voice, and 1080p, Google DeepMind's Lyria 3.5 pushed the music-generation race, and IFPI set chart-eligibility rules for AI music. Google also launched and then pulled a Nano Banana image feature in Google Earth within a day, a reliability lesson.

Ch. 131→Ch. 44→Ch. 49→Ch. 57→Ch. 59→Ch. 22→

The hybrid filmmaking guide gained its on-stage half and its finishing craft. New material covers virtual production (the LED volume that The Mandalorian made mainstream with ILM StageCraft) and performance capture (the Avatar lineage), both now fed by generative AI as in the Wonder Project and Luma venture Innovative Dreams; the mechanical reason a performance resists generation (no physical weight, texture too perfect, temporal degradation that worsens with shot length); and a Techniques that sell the composite section: keep the synthetic insert short and sandwich it between real anchors, grow a wide shot from a real anchor plate, bridge two real frames with a generated camera move, and degrade the clean AI image back toward reality with grain, halation, and an aggressive grade. Cost pressure is now grounded in FilmLA's 2025 figures (Los Angeles down to 19,694 shoot days, television down nearly sixty percent from its 2021 peak).

Ch. 116→

Concrete claims throughout the book now link to their source inline, right where the claim is made, rather than only in the updates provenance list. Two hundred and forty five citations were threaded across eighty four sections, each anchored to a distinctive proper noun, dated event, or specific figure and opening the primary source in a new tab. The reader can now verify a model launch, a funding round, a benchmark result, a research paper, or a legal ruling at the point of reading.

The hybrid filmmaking section is now a full instructional guide to running a live-action production with AI at the edges. It covers the division of labor across the frame (shoot the performance, generate the front and back of the pipeline), previs as planning not commitment, an on-set capture checklist (clean plates, lighting and lens reference, tracking markers, a real-time reference as on Here, and consent capture on the day), the post-production effects loop with tools mapped to tasks (Metaphysic and MARZ for faces, Runway Aleph for in-plate edits, Autodesk Flow Studio for CG doubles, Beeble for relighting) and the reminder that the human finish is where the quality lives, finishing and localization (Topaz, ElevenLabs, Flawless), and a clearance workflow for consent, chain of title, and E and O insurance that makes a hybrid film ownable where a fully generated one is not. Closes on common mistakes and a prep-to-delivery checklist.

Ch. 116→

A new AI Filmmaking section covers the hybrid process in depth: how generative AI enters a real live-action production at specific handoffs rather than making the whole film. It walks the stages (previs, on-set real-time de-aging on Zemeckis's Here, the deep post and effects zone with Metaphysic now in DNEG's Brahma, MARZ Vanity AI, Runway Aleph, Autodesk Flow Studio and Beeble, finishing with Topaz, and audio dubbing plus visual dubbing with ElevenLabs and Flawless), and argues that hybrid dominates because it is the only version that is cheaper, copyrightable, and insurable at once, where a fully generated film fails the human-authorship and errors-and-omissions chain-of-title tests. It ends on labor: SAG-AFTRA's consent regime made de-aging and voice refinement of an actor's own performance workable, the backlash concentrated on edge-of-frame uses that displaced effects and title artists, and training data remains the unresolved fight.

Ch. 116→

A freshness sweep added Midjourney V8.2, Microsoft's MAI-Image-2.5-Pro and MAI-Voice-2-Flash preview, Meta's Muse Image and its consent controversy, and the Delhi High Court's ANI v. OpenAI fair-dealing ruling. The microdrama vertical got a deeper treatment: it monetizes like a free-to-play mobile game, its short vertical format fits exactly what AI video does well, AI arrived first through dubbing and localization, and distribution scale (ReelShort and DramaBox reportedly hold about seventy percent of revenue) is the moat, not the model. The ecosystem Microdramas directory was expanded accordingly.

Ch. 26→Ch. 31→Ch. 131→Ch. 93→

A QA sweep reconciled internal contradictions introduced during the recent expansion (stale model hedges versus shipped notes, date and figure drift, Wan open-versus-closed, the Sora timeline in the Hollywood chapter). A freshness check against the leveled-up sources added dated notes: GPT Image 2 and Gemini Omni Flash now top the Artificial Analysis boards; the music industry launched a voluntary AI-Generated versus AI-Assisted labeling standard; Germany's GEMA v. Suno ruling is set for July 31, 2026; Udio added BuyDRM export enforcement; Netflix acquired Ben Affleck's InterPositive for about 587 million dollars; and Google DeepMind took a research stake in A24.

Ch. 69→Ch. 59→Ch. 56→Ch. 117→Ch. 46→

Expanded the flagship AI Filmmaking part with new sections on consistency and control (image-first, reference conditioning, identity LoRAs, camera and keyframe control, performance transfer), sound and post tied to the audio pillar, where AI wins by format (ads, music video, the microdrama boom, previs, VFX), the studios and the money (Promise, Asteria and Moonvalley, Primordial Soup, Staircase, Toonstar, Critterz), festivals labor and the law (SAG-AFTRA consent, the Copyright Office report, the Thaler Supreme Court denial), and the real-time world-model frontier. Deepened the platform taxonomy and the verified filmmaker case studies.

Ch. 111→Ch. 113→Ch. 115→Ch. 117→Ch. 118→Ch. 120→

Closed the editorial pass with structure and naming. The audio part is renamed Audio Models: Voice, Music, and Sound so its title reflects that audio is voice, music, and sound design, not voice alone. The audio-frontier section now sits with the other frontier topics. And the Research Frontier is reframed as the book's forward-looking capstone spanning image, video, and audio, resolving the jump from the industry chapters back into research methods.

Ch. 142→

Deepened two of the hottest specialized areas. Editing now covers the current instruction-based models (Gemini 2.5 Flash Image / Nano Banana, Qwen-Image-Edit, FLUX.1 Kontext, ByteDance SeedEdit and Seedream) and why surgical, multi-turn, consistent-character editing anchors real workflows. Lipsync was rebuilt from the Wav2Lip research lineage through Sync.so, Hedra Character-3, HeyGen, Runway Act-Two, and lip-sync now landing natively inside video models.

Ch. 60→Ch. 61→

Moved the Audio part closer to parity with image and video. Added a standalone section on native audio in video (Veo 3, Sora 2, Seedance, Wan, Kling, Grok) and the up-market pressure it creates. Deepened Udio (its fidelity edge, the October 2025 UMG settlement, and the download-lockout backlash) and the music and sound landscape (Lyria across Google's ecosystem, Google's acquisition of Riffusion via ProducerAI, Stable Audio's licensed-data play, AudioCraft's non-commercial weights, and Tencent and ACE-Step on the open frontier).

Ch. 58→Ch. 56→Ch. 57→

Consolidated and deepened the flagship AI Filmmaking part. New arc: vibe directing as the new posture, the platforms and the agentic production loop, how an AI film actually gets made (with sound as a first-class stage tied to the audio pillar), the craft of technique-volume-and-the-edit, the films and filmmakers and where AI wins first, and the team of one and the author's chair. Verified examples include the Kalshi Veo 3 ad, Neural Viz, OpenArt's Director and vibe directing, Runway Gen-4/Aleph/Act-Two, the Runway AI Film Festival and Total Pixel Space, Critterz, the Here de-aging work, and Aronofsky's Primordial Soup.

Ch. 108→Ch. 109→Ch. 110→Ch. 112→Ch. 114→Ch. 119→

Editorial pass stitching the audio pillar into the chapters that previously covered only image and video. Comparative Analysis now has an audio model landscape (voice latency, cloning, and the music structure and licensing axes). Evaluating Quality now has an audio quality section (naturalness, artifacts, consistency, controllability, intelligibility, provenance, weighted by use case). The Craft now covers directing voice, music, and sound, including the two-consents rule for cloning. The Operator's Playbook now models audio unit economics (speech meters on volume, music sells a cap), a hosted-vs-self-hosted decision, and the licensed-data-plus-indemnification moat. The Foundations gains a one-page audio-mechanics orientation.

Ch. 66→Ch. 87→Ch. 107→Ch. 09→Ch. 129→Ch. 142→

Added two instrument-panel figures to the audio foundations chapter: the neural codec as the audio VAE (waveform to residual-vector-quantized tokens and back), and the two text-to-speech families (autoregressive codec language models versus flow matching).

Ch. 52→

Final audio drop, completing the part. Added the music and sound landscape (Lyria, Stable Audio, Meta AudioCraft, Riffusion, Tencent, sound effects and foley, and the native-audio-in-video bifurcation that pushes standalone tools up-market) and a closing reckoning chapter on licensing and voice likeness: the turn toward licensed data, the ELVIS and NO FAKES laws and SAG-AFTRA consent regime, provenance via watermarking and C2PA, and the through-line that the ability to generate a voice does not settle whether it is lawful or right.

Ch. 57→Ch. 59→

Second audio drop. Added the voice landscape beyond ElevenLabs (the real-time lane with Cartesia, OpenAI, Google, Hume, Sesame and the Scarlett Johansson Sky episode; the open lane with Kokoro, Chatterbox, Fish Speech and Microsoft declining to release VALL-E 2; the editing and enterprise lane; and voice in filmmaking, where speech-to-speech conversion beats text-to-speech), plus full deep dives on Suno and Udio, including their mirror-image settlement paths with the record labels.

Ch. 54→Ch. 55→Ch. 56→

Added a new Audio and Voice part alongside image and video, opening with a foundations chapter on how audio generation works (neural codecs as the audio VAE, the two TTS families and voice cloning, why music is harder than speech, sound effects, and native audio in video) and a full ElevenLabs deep dive covering the company, the model line and product surface, getting the best output and holding a character voice across a film, and the likeness and consent problem. Voice, music, and the licensing reckoning follow in subsequent drops.

Ch. 52→Ch. 53→

Answered the missing case in the will-my-prompts-train-the-next-model FAQ: hosted company UIs and APIs run on the providers servers and can log and train on your inputs, while running open-weight models locally through something like ComfyUI keeps everything on your machine, private by construction rather than by promise.

Ch. 15→

Rewrote the ComfyUI workflow walkthrough to follow one concrete image (a cinematic astronaut in a wheat field, made with an SDXL checkpoint like Juggernaut XL) through every node, with specific prompts, a 25-step DPM++ 2M Karras KSampler setup, and named variations (image-to-image, LoRA, upscale, ControlNet, and swapping in a Wan or Hunyuan video loader), so the abstract graph is grounded in an actual use case.

Ch. 14→

Added a subsection to the ComfyUI infrastructure chapter that meets a beginner's overwhelm head on, with analogies (the airplane cockpit, the microwave versus the professional kitchen, the graph as a language) and the core why: ComfyUI does not add complexity, it reveals the complexity that was always there, and every workflow is the same six-node skeleton with small insertions.

Ch. 14→

Deepened the Krea chapter with its honest dual nature as a real-time canvas and aggregation layer that also trains its own models, the flag that Krea 1 rode the FLUX ecosystem and Krea Realtime 14B distilled Wan while only Krea 2 is from scratch, its real-time and aesthetic-post-training edge, and practical access and workflow guidance.

Ch. 30→

Deepened the Alibaba Qwen-Image chapter with the 20B MMDiT on a frozen Qwen2.5-VL encoder, the Apache 2.0 open weights and standout Chinese text rendering, the honest 2026 drift as Qwen-Image-3.0 shipped closed with no weights or report, and practical open-weight, quantized-local, and API access guidance.

Ch. 28→

Deepened the ByteDance Seedream chapter with the unified Seedream 4.0 (one diffusion transformer for generation, editing, and composition at 4K in about 1.8 seconds), the launch arena double-top over Nano Banana that later eroded, the secondary-sourced 4.5 and 5.0 flags, and practical access across Dreamina, Doubao, and BytePlus with the US caveat.

Ch. 29→

Deepened the OpenAI image chapter with the March 2025 shift from standalone DALL-E to native generation inside GPT-4o, the viral Studio Ghibli moment and the melting-GPUs demand, the gpt-image-1 through gpt-image-2 lineage now topping the arena, and practical access, token pricing, and the honest weaknesses (speed, the warm tint, moderation).

Ch. 23→

Expanded the thin Stability chapter into a full profile: the SD3 license backlash that drove the community to FLUX, the two architectural eras (latent diffusion to MMDiT rectified flow), honest strengths and weaknesses, the near-bankruptcy and the Akkaraju, Parker, and Cameron reset, practical open-weight and licensing guidance, and the honest flag that the Stable Diffusion 4 reports are SEO blog spam.

Ch. 27→

Deepened the Midjourney chapter with the Disney, Universal, and Warner Bros. copyright lawsuits and the training-data opacity behind them, practical access and parameter prompting guidance (--ar, --stylize, --sref, RAW), and the bootstrapped, profitable independence funding a long arc toward video, 3D, and world simulation.

Ch. 26→

Deepened the Recraft chapter with its design-and-brand positioning, the red panda stealth arena win with V3, the editable-vector and typography differentiators, the V4 and V4.1 design-taste rebuilds, and practical guidance on when to choose it over general image models.

Ch. 25→

Deepened the Ideogram chapter with the honest read that its text-rendering moat narrowed as DALL-E, Imagen, FLUX, Recraft, and GPT-native generation caught up, the gap between its own and neutral benchmark rankings, the unverified Ideogram 4.0 open-weight reports, and practical prompting for legible in-image text.

Ch. 24→

Deepened the Google image chapter with the consolidation onto the Gemini-native path (Imagen 4 retirement, migration to Nano Banana 2), the confusing Pro-versus-2 tier naming, the mid-2026 arena standing behind GPT Image 2, and practical guidance on the Pro, 2, and 2 Lite ladder plus the identity-consistency and conversational-editing strengths.

Ch. 22→

Deepened the Black Forest Labs FLUX chapter with the 2026 multimodal turn (FLUX.2, the July FLUX 3 unified image, video, audio, and action model, the FLUX mimic robotics model and Audi test, and the 300-million-dollar Series B) plus practical open-versus-pro access and licensing guidance.

Ch. 21→

Added practical access, pricing, and prompting guidance to the xAI Grok Imagine chapter, plus an honest, sourced account of the permissive-moderation stance and the nonconsensual-deepfake controversy and regulatory response that followed.

Ch. 49→

Deepened the Pika chapter with the Guo and Meng founding and their diffusion-research pedigree, the deliberate choice to own the fun, social, effects-driven corner of AI video rather than the frontier, honest strengths and weaknesses, and practical guidance on using the effects and ingredient features.

Ch. 48→

Deepened the Luma chapter with its pivot from video generation to unified world models (Ray3 reasoning, Uni-1 and Luma Agents), the 900-million-dollar HUMAIN Series C and the Project Halo supercluster, honest strengths and weaknesses, and practical access and HDR-pipeline guidance.

Ch. 47→

Expanded the Alibaba Wan chapter with the crucial 2026 finding that its open weights stop at Wan 2.2 while 2.5, 2.6, and likely 2.7 are closed API-only products (an open-core split, not full openness), plus honest strengths and weaknesses and open-versus-hosted access guidance.

Ch. 45→

Added an honest strengths-and-weaknesses read to the MiniMax Hailuo chapter (the stale mid-2025 arena ranking, the 456B parameter figure that is actually the M1 language model, and the Disney, Universal, and Warner lawsuit) plus practical access and director-camera prompting guidance.

Ch. 44→

Added practical access guidance (Dreamina, Jimeng, Doubao, CapCut, BytePlus, and the US-availability caveat), the honest resolution-spec inconsistency, and the unified-multimodal direction including Seedance 2.0's native audio, to the ByteDance Seedance chapter.

Ch. 46→

Rewrote the densest ComfyUI passages for a general reader. Each hard section now opens with an everyday analogy before the machinery: the installed tool as a small power plant you run rather than a storefront you visit, the typed wires as plumbing that only connects when the gauges match, the six-node default workflow as a darkroom assembly line where the painter works on a compressed draft (which is why latents, samplers, and a VAE decode each exist), the denoise dial as how much of the canvas you repaint, the API export as turning a hand-drawn recipe into a callable button, and owning versus renting a GPU as owning versus renting a car. New figures render the six typed data wires, the model-to-workflow-to-agent climb, and the four ways to run, and the canonical graph and latent-pipeline diagrams now sit inside the manual itself.

Ch. 70→Ch. 71→Ch. 72→Ch. 73→Ch. 77→Ch. 79→

Noted the July 23 launch of FLUX 3, a unified image, video, audio, and action model with native synchronized audio and a companion robotics model (FLUX mimic), in the Black Forest Labs chapter. This is the first finding promoted out of the weekly Field Reports into the standing book.

Ch. 21→

Added a Higgsfield profile to the ecosystem aggregator layer: the cinematic camera-control differentiator, the DoP and Soul models wrapping frontier engines like Kling, Veo, and Sora, the steep 2026 revenue and funding ramp toward a reported 1.3 billion valuation, and an honest read on the wrapped-model quality ceiling and the February 2026 Forbes controversy over fake demos, deepfake content, and billing.

Ch. 96→

Deepened the Runway chapter with the company's NYU ITP founding and its co-authorship of Stable Diffusion, the funding arc and the AI Film Festival, an honest read on where its editing-first control beats raw generation and where it trails the frontier, and a practical guide to plans, pricing, and treating Runway as an editing and orchestration hub rather than a clip slot machine.

Ch. 43→

Deepened the Google Veo chapter with the DeepMind organization and the vertical-integration moat (in-house TPUs, the YouTube corpus, and distribution across Flow, Gemini, Vertex AI, and the API), the reported seventy-five million dollar A24 partnership, and a practical guide to the four access paths, per-second and credit pricing, and how to prompt for camera, lens, and native audio.

Ch. 41→

Rebuilt the OpenAI Sora chapter into a full profile: Tim Brooks and Bill Peebles and the DiT paper Sora was built on, the spacetime-patch world-simulator thesis read honestly against what actually emerged, Sora 2's native audio and Cameos, and the economics and cultural legacy behind the 2026 app shutdown and API sunset.

Ch. 40→

Expanded the Kling chapter with Kuaishou's business and the July 2026 spin-off (a ~$2.8B raise near an $18B valuation, Kuaishou keeping ~68%, Tencent/Alibaba/Baidu backing), an honest architecture read from the Kling-Omni and Kling-Foley papers, its mid-2026 arena standing (~#6, behind Gemini Omni Flash, Seedance, and Wan), and practical guidance on modes, access, pricing, and prompting.

Ch. 42→

Comfy Org shipped an official hosted Comfy Cloud MCP server (public beta, June 2026) that lets an AI assistant search, assemble, run, and iterate on ComfyUI workflows, alongside a growing set of local community MCP servers (Artokun's control plane, Joe Norton's lightweight server) that can also edit the live graph and install nodes and models.

Ch. 77→

ComfyUI added int4/int8 quantization, native SeedVR2 upscaling, Depth Anything 3, SCAIL-2 multi-reference character replacement, and PixelDiT support through mid-2026, while ai-toolkit added LoRA training for FLUX.2 and Z-Image.

Ch. 78→

A judge let the studios' suit against MiniMax's Hailuo proceed (May 26), the New York Times sought sanctions against OpenAI over discovery (July), and the EU AI Act's enforcement powers over general-purpose model providers take effect August 2, 2026 with fines up to 15M euros or 3% of global turnover.

Ch. 10→Ch. 135→

Mid-2026 saw rights-holders take stakes and sign licensing deals: Google DeepMind invested $75M in A24 to co-develop Veo tools, Lionsgate took an equity stake in Runway (with a Bertelsmann partnership following), and Getty Images signed a display-only deal with OpenAI for ChatGPT.

Ch. 123→

Google shipped Nano Banana 2 (Gemini 3.1 Flash Image) in February 2026 across Gemini, Search, and Ads, in Lite, standard, and Pro tiers, with several variants ranking in the image-arena top ten.

Ch. 22→

Ideogram released Ideogram 4.0 on June 3, 2026: a 9.3B single-stream diffusion Transformer with a Qwen3-VL text encoder, native 2K resolution, and JSON prompting, its first model with open weights alongside API access.

Ch. 24→

FLUX 3 launched July 23, 2026 as a multimodal frontier model generating images plus video with synchronized audio (clips up to roughly 20 seconds), in limited early access, extending Black Forest Labs beyond still images.

Ch. 21→

OpenAI released GPT Image 2 (ChatGPT Images 2.0) on April 21, 2026, retiring DALL-E 3 and GPT-Image-1.5 with near-perfect text rendering, higher resolution, and a thinking mode. It currently ranks first on the Artificial Analysis text-to-image arena.

Ch. 23→Ch. 69→

World models became the fastest-moving frontier category in 2026: Reactor emerged from stealth with a $59M round to build a developer platform for real-time AI worlds, alongside World Labs' Marble, Google DeepMind's Genie 3, and Tencent's open real-time HY-World.

Ch. 97→

The compute layer that serves generative media drew huge rounds in mid-2026: fal named AWS its preferred cloud at roughly 2.5 million developers, Baseten raised a $1.5B Series F near a $13B valuation, and Together AI raised $800M at $8.3B, all on rapidly growing inference demand.

Ch. 99→

Kuaishou disclosed in a May 12 exchange filing that it is evaluating a restructuring of the Kling AI business, with reports pointing to a pre-IPO round near a twenty-billion-dollar valuation as Kling's revenue run rate climbed toward five hundred million dollars.

Ch. 42→

Luma released Ray3.2 (June 9), adding frame-by-frame directability and HDR generation with paired EXR outputs for professional pipelines, followed by 'Skills' repeatable agent workflows (June 16).

Ch. 47→

ByteDance unveiled Seedance 2.5 at its Volcano Engine FORCE conference (June 23), claiming native single-shot video up to thirty seconds and up to roughly fifty reference inputs, with public API access via BytePlus opening July 16.

Ch. 46→

Runway shipped Aleph 2.0 and Edit Studio (May 21), then a Media Router that selects across image, video, and audio models (July 23), while taking a Lionsgate equity stake (June 11) and announcing a Bertelsmann partnership (July 1). The pattern is a bet on owning workflow and distribution rather than only the base model.

Ch. 43→

Google shipped no Veo 4 at I/O 2026 and instead launched Gemini Omni Flash (June 30), a fast video generation and conversational-editing model that now ranks first on both the text-to-video and image-to-video arenas. Veo 3.1 remains the standalone flagship.

Ch. 41→Ch. 51→

The ComfyUI manual gains a mastery arc: where it runs (local vs rented GPUs, RunPod, serverless, Comfy Cloud), the pitfalls that cost beginners a weekend, the 80/20 path to getting good, how professionals actually build with it, and what 'senior at ComfyUI' means.

Ch. 79→Ch. 80→Ch. 81→Ch. 82→Ch. 83→

Two chapters on process: deep profiles of how PJ Accetturo, Neural Viz, Dave Clark, Paul Trillo, and the Dor Brothers actually make their work, and the patterns that recur across all of them (the common stack, generate-then-cull ratios, model-mixing, and editing discipline).

Ch. 112→

A new chapter uses fal, the developer inference platform hosting 1,000+ generative models, as a vantage point on the whole medium: a guided tour of the interesting and less-mainstream models by category, profiling Bria (licensed-data image), Reve, Sonilo (video-to-music), Decart (real-time world models), and Lightricks' open LTX-2, among many others.

Ch. 101→

Verified the mid-2026 landscape against the live leaderboards and updated the stragglers: Midjourney V8.1 is now the default, ByteDance shipped Seedream 5.0 Pro, and the current arena order (GPT Image 2 on top, Reve 2.1 and Microsoft MAI-Image-2.5 behind; Gemini Omni Flash leading video) is confirmed.

Ch. 26→Ch. 29→Ch. 32→Ch. 51→

A thorough hands-on manual for ComfyUI in The Machine: install and setup, reading the node graph, the default text-to-image workflow node by node, the core recipes (img2img, inpaint, outpaint, upscale, LoRA, ControlNet, IPAdapter), native video and audio, the Manager and registry, the headless/agentic API, and practical memory and reproducibility.

Ch. 70→Ch. 72→Ch. 73→Ch. 74→Ch. 75→Ch. 77→Ch. 78→

Three more chapters on the craft: how the agentic loop actually runs inside Director, Flow, LTX Studio, and Showrunner; where AI filmmaking is already winning (ads, music videos, microdramas, previs, VFX) and where it still loses; and the real team shape and economics of an AI production, from Critterz to the solo creator.

Ch. 109→Ch. 114→Ch. 119→

The book is now organized as two movements, The Machine (how generative models work) and The Medium (what is being built on them). A new chapter walks the entire creative stack in prose, phase by phase, naming the companies doing the work in each box.

Ch. 95→Ch. 96→Ch. 97→Ch. 98→Ch. 99→Ch. 100→

The AI Filmmaking part gains three craft chapters: how an AI film actually gets made end to end, the techniques that separate finished work from slop, and the real films and filmmakers driving the moment (PJ Accetturo, Neural Viz, Curious Refuge, Runway's festival, and the Sora-made feature Critterz).

Ch. 110→Ch. 112→Ch. 114→

The native AI-creator canon (Will Smith spaghetti, the Ghibli trend, Neural Viz, PJ Ace, Curious Refuge, Runway's AIFF) and the two backlash fights, over slop (Willison's definition) and over consent (Glaze/Nightshade, Tilly Norwood, ScarJo vs Sky, likeness vaults, the jobs debate).

Ch. 124→Ch. 125→

What the studios actually did: Lionsgate-Runway's underdelivery and restructuring, Netflix's ~300 AI-assisted 2026 titles, the Disney-OpenAI Sora deal and its March 2026 collapse plus Disney's lawsuits, and the Sphere's $370M+ AI Wizard of Oz.

Ch. 123→

A new Industry act opens with the directors' split (Scorsese joining Black Forest Labs, Cameron, Lucas, Jackson vs del Toro, Villeneuve, Spielberg, Nolan as DGA president) and the institutional rules (WGA/SAG-AFTRA/DGA deals, the Academy's AI eligibility rules, the White House letter).

Ch. 121→Ch. 122→

A new part covers the turn from generating clips to directing films: vibe directing (OpenArt's June 2026 coinage, the film analogue of vibe coding), the agentic platforms led by Runway and OpenArt, the broader movement, and the authorship debate that dominated Cannes 2026.

Ch. 108→Ch. 109→Ch. 119→

The contents are regrouped into named acts (Foundations, The Big Picture, The Engines, The Craft, The Business, The Frontier, Reference), the glossary moves to the back, and the title broadens to reflect that the story is now the whole stack and AI filmmaking, not only the models.

A different mode from batch: realtime canvases (Krea, Decart) and world models generate as fast as you move, collapsing the pipeline and folding finishing and distribution into the moment of generation.

Ch. 94→

Distribution is where the money lands. The vertical microdrama market (~$3.8B in-app in 2025, forecast to more than double in 2026) rivals Netflix for US mobile time; today it is mostly live action with AI as accelerant, but it is the format most exposed to full generation, with AI-native apps like Popshort and studios like Fairground emerging.

Ch. 93→

Audio (voice, dubbing, lip-sync, music, SFX) is nearly half the production phase and had been underweighted; a video pipeline without an audio plan is half a pipeline.

Ch. 92→

A finished AI-native piece is a chain of specialized jobs (write, board, generate, hold consistency, voice, lip-sync, edit, grade, mark, distribute); generation is one link and not the hardest.

Ch. 91→

Creators increasingly work in multi-model canvases (Krea, Higgsfield, Flora, ComfyUI) rather than going model-direct; the workflow, not the model, owns the user, and agents are starting to drive the workflow.

Ch. 90→

A new anchoring part reframes the field from the point of view of the whole creative stack rather than individual models: as generation commoditizes, value migrates up to workflows and down to distribution. Concrete backing: the application layer now captures the majority of enterprise gen-AI spend, and app-layer players like Higgsfield reached a ~$500M run rate in about a year.

Ch. 88→Ch. 89→

A new Foundations chapter on reasoning-augmented image generation: GPT-Image-2's Thinking mode plans, web-searches, and self-checks before rendering (architecture undisclosed), alongside a genuine 2026 research revival of autoregressive image models.

Ch. 18→

A new Foundations chapter on native audio-visual generation: Veo, Kling 3.0's Omni Native Audio, and open-weight LTX-2 generate synchronized sound in the same pass as the picture; open models like Foley-Omni handle video-to-audio.

Ch. 19→

A new Operator's Playbook section on the EU AI Act Article 50 transparency obligations that apply from August 2, 2026 (deepfake disclosure and machine-readable marking), and the C2PA plus SynthID standards converging to meet them.

Ch. 135→

A new section promoting world models from research to product: DeepMind's Genie 3 (real-time interactive worlds, ~1-minute memory, research prototype) and World Labs' Marble (a commercial paid product that exports Gaussian splats and meshes).

Ch. 141→

Chapter 7 gains a subsection on Comfy MCP (launched June 30, 2026), the official bridge that lets AI agents assemble and run ComfyUI workflows on Comfy Cloud GPUs, plus in-canvas copilots.

Ch. 14→

The economics chapter gains a hardware-and-pricing subsection: B200/GB200 rental ranges, Blackwell's inference-cost cut over Hopper, and current API reference prices (FLUX.2 from official docs; per-second video prices marked approximate).

Ch. 13→

Chapter 7 gains a full treatment of the KSampler: what a denoising step does and where the returns flatten, how CFG scale trades prompt adherence against image quality, and how to pick a sampler and scheduler. The through-line is that settings must match the model class.

Ch. 14→

The Sora app went dark in April and the API will be discontinued September 24, 2026, amid roughly a million dollars a day in cost, a dead Disney deal, and copyright problems. Sora continues only as internal world-model research.

Ch. 40→Ch. 51→

Reve 2.1 and Microsoft's MAI-Image-2.5 now rank #2 and #3 on the Artificial Analysis image arena, and xAI's Grok Imagine Video 1.5 (May 31) briefly topped the image-to-video arena with native audio. All three were absent from earlier editions.

Ch. 32→Ch. 51→

Meta Superintelligence Labs launched Muse Image, its first image model, an agentic system that plans a layout before rendering, shipping free inside the Meta AI app, Instagram, and WhatsApp.

Ch. 32→

Google began sunsetting the Imagen family, with API endpoints shutting down through August 2026 and the migration path pointing to Nano Banana. Nano Banana 2 Lite became the fastest, cheapest tier at ~4s and ~3 cents per image.

Ch. 22→Ch. 32→

Kling 3.0 Turbo added multi-shot prompting and native audio as Kuaishou raised at a reported ~$18B valuation; ByteDance's Seedance 2.5 generates native single-pass 30-second 4K clips; Alibaba's HappyHorse topped the open video arenas with single-pass joint audio and video.

Ch. 42→Ch. 46→Ch. 51→

Three serious open-weight image models shipped in five weeks, closing the gap on the closed flagships: Ideogram 4.0 with structured-JSON prompting, Krea 2 with a two-second Turbo variant, and the MIT-licensed pixel-native HiDream-O1.

Ch. 32→

Recraft V4.1 briefly led outside Google and OpenAI in May, but by late July had been passed by Reve 2.1 and Microsoft's MAI-Image-2.5 on the Artificial Analysis arena.

Ch. 32→

Rather than a Veo 4, Google I/O introduced Gemini Omni, an any-to-any multimodal family. Gemini Omni Flash shipped the same day and sat at #1 across the major video arenas by July.

Ch. 41→Ch. 51→