Excellence in fashion ecom and film/TV, today and in the months ahead

Fashion e-commerce has become the most commercially-mature application of generative imagery in 2026, specifically because it combines four things that align perfectly with what current models are good at: high volume (thousands of SKUs per brand), high sensitivity to cost per image (every dollar saved on photography flows to margin), strong need for consistency (brand look and feel across the entire catalog), and tolerance for imperfection on non-hero shots (a product thumbnail does not need to be indistinguishable from studio photography). The industry has responded with a tier of specialist tools that all essentially solve the same problem stack: turn a flat-layFlat-layA product photo taken from directly above with the item laid flat, a common input that AI tools convert into on-model imagery. or ghost-mannequin photo into a professional-looking on-modelOn-modelProduct imagery that shows a garment worn by a real or synthetic model rather than laid flat or on a mannequin. image, generate lifestyle backgrounds for any product, maintain visual consistency across a catalog, and do it at a cost per image roughly one to two orders of magnitude below traditional photo shoots.

The working consensus among operators in early 2026 is that AI handles roughly 70 to 80 percent of typical catalog photography needs, with the remaining 20 to 30 percent (hero product pages, luxury goods, items where materiality and texture drive the purchase decision) still better served by traditional photography. Excellence in fashion e-commerce today means getting the 70 to 80 percent right at volume and knowing which assets need to stay on the traditional pipeline. The operators who are winning are not the ones who try to replace traditional photography entirely, they are the ones who draw the line correctly and then execute flawlessly on both sides of it.

The tooling landscape as of mid-2026

The fashion-specific tool landscape has consolidated into roughly seven first-class options, each with a distinct position. Claid is the leader for high-volume catalog operations, with batch automation, background removal, AI scene generation, and virtual model features that handle enterprise catalogs of more than 500 SKUs with consistent output quality. Photoroom is the leader for speed and simplicity, with template-driven workflows for marketplace listings, Amazon-ready aspect ratios, and the fastest time-to-listing for small catalogs. Rawshot is the leader for volume production with compliance requirements, offering more than 600 synthetic models, 150 camera styles, and 1,500 backgrounds, plus C2PAC2PA (Content Credentials)An open provenance standard, backed by Adobe, OpenAI, and Microsoft, that attaches cryptographically signed metadata proving a file's origin, though the metadata strips easily when a file is re-encoded. content authentication which is increasingly relevant for brands selling in the EU under the August 2026 AI Act requirements. SellerPic combines fashion model swaps with AI video generation and has become popular with Shopify and direct-to-consumer sellers for social-selling workflows. Botika is the Shopify-native option for quick flat-lay-to-on-model conversion with minimal setup. BetterStudio targets brands that want premium studio-grade imagery with strong ethical positioning around digital-twin models. WearView is the newest entrant, positioning itself as a complete fashion workflow platform covering virtual try-onVirtual try-on (VTO)Placing a garment or product onto a model or person in an image, so a catalog can be produced without a physical photoshoot., AI model creation, product-to-model, and video in one platform.

Behind these specialist platforms sit the foundation models that actually produce the pixels. Most of the fashion specialists are running on top of FLUX.2 [dev] or a fine-tuned FLUX variant, with some using Qwen-Image-Edit for specific editing tasks and some using Nano Banana Pro for hero shots that need maximum polish. The specialist platforms add the fashion-specific workflow layer: virtual try-on, garment fitting, model diversity controls, fabric physics simulation, pose libraries, brand kits, catalog management, and the integrations into Shopify and marketplace systems that operators actually need. This is the clearest case study in the document of the a16z/fal finding that the unit of work is a workflow not a model, fashion operators rarely talk about which base model they are using, they talk about which workflow platform handles their catalog.

What mastery actually looks like in fashion e-commerce

Mastery in fashion generative imagery today has five components, in rough order of importance. First, brand consistency at scale, meaning your AI-generated images look like they come from the same brand across hundreds or thousands of SKUs without requiring manual curation of each one. This is almost entirely a matter of training a custom LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. on your existing brand photography (the fal/a16z report calls this 'customizability' and flags it as the reason open-weight models are winning enterprise), plus building the catalog management infrastructure to apply that LoRA consistently. Second, garment physics that survive close inspection, meaning the fabric drapes correctly, the seams align where they should, the model's body proportions are appropriate for the garment, and the lighting preserves the texture and material differences that distinguish one fabric from another. Third, model diversity done well, meaning your virtual models represent the actual customers you want to sell to without falling into the defaults-are-white-and-thin trap covered in the bias discussion in the common-questions chapter of the foundations. Fourth, workflow integration with your product information management system so that new SKUs flow through the generation pipeline automatically rather than requiring a human to kick off each batch. Fifth, the judgment to know which assets still need traditional photography and not to try to AI-generate those.

Where fashion e-commerce is moving in the next 6 to 12 months

The direction of travel is toward video and live-interaction use cases. As of mid-2026, most fashion AI work is still-image generation, but the frontier is moving fast toward motion. Virtual try-on is evolving from still-image fitting into short-form video that shows the garment moving on the model as they walk or turn, which is dramatically more convincing for shoppers than a still image and significantly increases conversion rates in early tests. Seedance 2.0's character consistencyCharacter consistencyThe problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work. advantages (covered in the ByteDance Seedance chapter) map directly onto fashion use cases because you can feed it a single model image and have that model appear consistently across dozens of generated videos wearing your full catalog. Expect the specialist tools (Claid, Rawshot, WearView, SellerPic) to ship video features through Q2 and Q3 2026 that will make garment-in-motion the standard catalog format by the end of the year.

The other frontier is real-time interaction. Live virtual try-on, where a customer sees themselves wearing a garment in real time as they move a camera, is currently technically possible but not yet widely deployed because the quality and latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput. trade-offs are hard to hit simultaneously. As the distilled fast variants of FLUX.2 (FLUX.2 [klein]) and the SANA-Sprint class of one-step models mature, expect real-time virtual try-on to become a feature of major fashion apps by late 2026 or early 2027. The operators who figure out how to integrate live interaction with their existing catalog generation pipelines will have a significant early-mover advantage. The ones who wait for the frontier to stabilize will arrive too late.

Beyond 2027, expect the fashion e-commerce stack to converge toward something that looks more like a unified creative operating system than a collection of specialist tools. A single platform that handles catalog generation, video, live try-on, model training, brand consistency, and marketplace distribution, with the underlying models treated as swappable infrastructure components. Whether this platform comes from one of the current specialists (Claid and WearView are the most plausible candidates) or from a Shopify-native competitor (Shopify itself has been investing heavily in AI tooling) or from an entirely new entrant is one of the open strategic questions. What is clear is that the current fragmentation is temporary, and operators should build their pipelines with the assumption that the tooling will consolidate.

Film and TV is the opposite of fashion e-commerce in almost every way. Where fashion is high-volume, cost-sensitive, and forgiving on non-hero shots, film and TV is low-volume relative to the total frames produced, quality-sensitive at every frame, and intolerant of the kinds of inconsistencies that AI models still produce. Excellence in film and TV generative imagery today is not about replacing traditional production, it is about augmenting specific parts of the pipeline where AI happens to be good enough. The operators who are winning in this vertical are the ones who have correctly identified which parts those are and built workflows around them.

The specific parts of film and TV production where AI generative tools have meaningfully landed in 2026 are: concept art and visual development (pre-production ideation, storyboarding, mood boards, character design), virtual productionVirtual productionShooting live action against LED walls or in-camera digital backgrounds instead of a green screen, often using AI-generated environments. assets (background plates, set extensions, environment mattes, props), pre-visualization (pre-viz for shots that will eventually be filmed traditionally, to plan camera moves and blocking), specific VFXVFX (Visual Effects)Imagery created or altered beyond what was captured on camera, now increasingly generated. tasks (simple rotoscopingRotoscopingIsolating a subject from its background frame by frame, traditionally by hand and increasingly automated by AI., background replacement, object removal, upscalingUpscalingIncreasing an image or video's resolution after generation, often with a model that invents plausible fine detail rather than simply stretching pixels. legacy footage), and short-form content adjacent to film and TV (title sequences, promotional trailers, social media cutdowns). What has not landed, despite three years of hype, is full scenes in finished productions. Even the most aggressive AI-first production houses use generative video as one tool among many, and the majority of screen-time in finished work still comes from traditional cinematography.

The tooling landscape for film and TV in mid-2026

The dominant tool in this vertical is Runway, which has positioned Gen-4 and Gen-4.5 specifically for film and television production workflows with character consistencyCharacter consistencyThe problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work. features, precise camera controlCamera controlPrompting or conditioning that directs virtual camera moves such as pan, orbit, dolly, and zoom in a generated video., and integration with post-production pipelines. Runway is not the highest-quality video model (Veo 3.1 beats it on photorealism, Kling 3.0 beats it on multi-shot storyboarding, Seedance 2.0 beats it on character consistency), but Runway has the best combination of quality, pipeline integration, and professional tooling to serve the specific needs of film and TV teams. The Gen-4.5 release added scene and character consistency features that make narrative continuity workable for the first time, which was the thing that was holding back serious film adoption of text-to-videoText-to-videoGenerating video from text alone, with no starting image. models.

Veo 3.1 is the quality leader for hero shots where budget allows. Per-second pricing in the range of $0.40 to $1.50 makes Veo too expensive for the majority of production work but entirely viable for specific hero shots where the quality ceiling matters more than the cost. Most serious film-production workflows in 2026 route hero shots to Veo 3.1 and use Runway for the bulk of generative content. Kling 3.0 sits alongside these two as the preferred choice for multi-shot narrative sequences. Seedance 2.0 is used for character-consistent work where a specific designed character has to appear across multiple generations. Open-weight models like Wan 2.6 serve the high-volume lower-budget end of the pipeline where generation volume matters more than per-shot quality. For specific specialized tasks, Sync Labs' LipsyncLipsyncSynchronizing the mouth movements of a person in video with an audio track of speech. 2 Pro handles dialogue synchronization at quality higher than what the native video models produce (covered in the specialized-models part), and ElevenLabs handles voice generation.

The essential complement to the video models is the workflow layer. Professional film and TV production cannot operate out of a collection of separate web tools, the pipeline integration overhead kills productivity. The platforms that have emerged to serve this need include Krea (for rapid iteration and multi-model routingModel routingAutomatically sending each request to the best model for it, by quality, speed, or cost, instead of committing to one model for everything.), MindStudio (for unified workspace across models plus built-in tools like clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. merging, face swapFace swapReplacing a face in an image or video with another identity, often used as a final correction pass to enforce character consistency., background removal, subtitle generation, and upscaling), and specialist post-production tools like DaVinci Resolve which have added AI generation features directly into existing professional workflows. The operators who succeed in this vertical almost always use multiple platforms simultaneously, the question is not which platform to standardize on but how to chain them together efficiently.

What mastery actually looks like in film and TV

Mastery in film and TV generative imagery today has six components. First, shot-by-shot selection, meaning you make a deliberate decision about which shots in a production will use AI generation and which will use traditional cinematography, based on cost, quality requirements, creative intent, and risk tolerance. The operators who try to AI-generate everything produce worse films than the operators who AI-generate thoughtfully. Second, character consistency infrastructure, meaning you have trained a LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. on each major character in your production and have systematic workflows for maintaining visual identity across every shot that character appears in. This is currently the biggest technical challenge in narrative AI video and the one that distinguishes serious operators from amateur ones. Third, cinematographic language in your prompts, meaning you write prompts using professional film vocabulary (specific lens focal lengths, lighting setups, camera movements, aspect ratios) rather than casual descriptive language, which produces dramatically better results on Veo, Runway, and Kling specifically because those models were trained on cinematographic data with that vocabulary.

Fourth, audio-first thinking, meaning you plan the audio design (dialogue, sound effects, music, ambient) alongside the visual generation rather than as a post-hoc addition. Veo 3.1 supports joint audio-video generation and produces dramatically better results (Sora 2 did too, before OpenAI wound it down in 2026) when you use that capability than when you generate silent video and add audio later. For dialogue specifically, using Sync Labs' Lipsync 2 Pro on top of either natively-generated audio or ElevenLabs voice synthesis is the current best practice for getting mouth movements right. Fifth, pipeline integration with your existing post-production tools, meaning the AI-generated content flows into DaVinci Resolve, Avid Media Composer, or Premiere Pro cleanly rather than requiring manual re-encoding and format conversion at every step. Sixth, and most importantly, the judgment to know when AI is the wrong tool entirely and traditional production is required. This last one sounds obvious but the operators who lack it produce work that looks exactly like every other AI-generated film, and the operators who have it produce work that feels like a real film that happened to use AI for specific shots.

Where film and TV is moving in the next 6 to 12 months

The biggest near-term change is long-form coherence. As of mid-2026 the current single-shot video limit is roughly 30 seconds of coherent output, which is enough for a single shot but not enough for a scene. By late 2026, expect the top models to reach coherent two-to-three-minute generation through some combination of hierarchical generation, improved character consistency, and better long-range attention mechanisms. This is the capability that unlocks genuine narrative scene generation rather than the current shot-by-shot assembly approach, and it will meaningfully change the cost structure and the creative possibilities of AI-assisted production when it arrives. Operators who are paying attention should be building their pipelines today in a way that is ready to absorb three-minute coherent scenes when they become available.

The second major change is audio-video tight coupling. The current generation of video models that support audio (Veo 3.1, Sora 2, Kling 3.0, Seedance 1.5 Pro) treat audio as a joint generation target but the audio quality and synchronization are not yet at the level needed for dialogue-heavy scenes. By late 2026, expect the audio generation quality to improve to the point where AI-generated dialogue scenes are usable for production, which will move the field meaningfully closer to generating full scripted content rather than just visual shots with audio added separately. The lipsync specialists (Sync Labs in particular) will probably still be better than the native video models for the specific task of retroactively syncing existing dialogue to generated video, but the gap will narrow and the joint-generation approach will start to dominate for new scenes.

The third change, and the most speculative one, is the arrival of world models as production tools. If Marble from World Labs and Genie 3 from DeepMind continue to improve at their current rate, by late 2026 or early 2027 they will be good enough that a director can generate a 3D environment from a text description, walk a virtual camera through it to block out shots, and export either rendered footage or set extensionSet extensionDigitally enlarging a physical set beyond what was actually built, so a small stage reads as a full location. assets for traditional production. This would represent a fundamental shift in how pre-visualization and virtual production work, and the film and TV operators who figure out how to integrate world models into their pipelines early will have a meaningful creative advantage over the ones who wait. The risk is that the scaling story behind world models does not work and this entire paragraph turns out to be wishful thinking, but the bet is worth tracking.

Beyond 2027, the interesting question for film and TV is not whether AI can generate a feature film, the technical answer will certainly be 'yes.' The interesting question is what the market for AI-generated film actually looks like: a replacement for traditional production, a new category alongside it, or simply another tool in the filmmaker's kit the way CGI became one. Nobody knows yet, and the operators building AI-assisted workflows today are the ones who will define the answer.