Black Forest Labs (FLUX)

Black Forest Labs is, in many ways, the most important image generation lab in the world right now. Not because it has the biggest model or the most users, but because it sits at the center of the open-source ecosystem. If you use any open-source AI image tool, ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org., Krea's image generation, fal.ai, the entire community of LoRAs and fine-tunes, you are almost certainly using a model that descends from Black Forest Labs in some way.

The company was founded in mid-2024 in Freiburg, Germany, the small city in the Black Forest region that gives the lab its name, by Robin Rombach, Andreas Blattmann, and Patrick Esser. These are the same people we met in the deep-history chapter of the foundations. They were the core authors of the original latent diffusionLatent diffusionDiffusion that operates in a compressed 'latent' space produced by a VAE, rather than directly in pixel space. Dramatically reduces compute cost. Used by essentially every modern image and video model. model paper, the technical leads behind Stable Diffusion 1.x, and the primary architects of what became Stable Diffusion 3 during their time at Stability AI. In early 2024, frustrated with the chaotic management situation at Stability under founder Emad Mostaque (who would resign as Chief Executive Officer (CEO) in March 2024 amid significant company turmoil), they left and founded Black Forest Labs.

They were not alone. The seed funding came from Andreessen Horowitz, with a $31 million round that closed in August 2024, exceptional for a seed round, even in the AI boom of 2024. Additional participation came from Brendan Iribe (the Oculus co-founder), Garry Tan (the Y Combinator president), Michael Ovitz (the legendary Hollywood agent), and Timo Aila from NVIDIA. The story of how Black Forest Labs got its early customers is also instructive: their first major partnership was with Elon Musk's xAI, providing the image generation backend for Grok's image features. This deal was announced almost simultaneously with the seed round, which is part of why a16z was willing to lead at such an aggressive valuation.

In December 2025, Black Forest Labs raised a $300 million Series B at a $3.25 billion post-money valuation, co-led by Salesforce Ventures and AMP (a new investment vehicle from Andreessen Horowitz general partner Anjney Midha). The round also disclosed a previously unannounced Series A. Total funding raised is over $450 million. The company is now treated, by both investors and users, as the European answer to the American closed-model labs, a serious player committed to keeping at least some of its models open while building a paid Application Programming Interface (APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface.) business on top.

FLUX is a 12-billion-parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. rectified-flow MM-DiTMM-DiT (Multimodal Diffusion Transformer)A variant of DiT that processes text and image (or audio) tokens in separate parallel streams joined by attention. Used by Stable Diffusion 3, FLUX, and most newer models. Improves prompt following significantly., meaning it uses the multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. diffusion transformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. architecture (separate parallel streams for text and image with joint attention) and is trained with a flow-matching objective (straight-line trajectories from noise to data). The text conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image. uses a dual encoder setup combining Contrastive Language-Image Pretraining (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.) for visual semantics and Text-to-Text Transfer Transformer (T5)-XXL for linguistic understanding. The model runs in latent spaceLatent spaceThe compressed numeric representation an image is turned into before diffusion. Working here instead of on raw pixels is what makes modern image models fast enough to run. via a Variational Autoencoder (VAEVAE (Variational Autoencoder)A neural network that learns to compress images into a small numerical representation (a latent) and decompress them back. Used in latent diffusion models to make the generation process tractable.), and the team has invested heavily in distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.). to produce faster variants of the same base model.

Architecturally, FLUX is closely related to Stable Diffusion 3, which makes sense, because the same people built both. But FLUX is meaningfully larger (12 billion parameters versus SD3's 8 billion in its largest variant), and it benefits from a year of additional training research and a larger compute budget. The result is a model that, when it launched in August 2024, immediately took the top of public image quality leaderboards and stayed there for months. Adobe, Canva, Meta, and Microsoft all built features on top of FLUX. By any reasonable measure, it is the dominant open-weight image model of the current generation.

FLUX.1 [pro]

Released August 2024, this is the closed-source flagship. It is only accessible through Black Forest Labs' APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. or through partner platforms (Krea, fal.ai, Replicate, etc.). The weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. have never been published. This is the highest-quality FLUX variant, and it is what you get when you select 'FLUX Pro' in any consumer tool. The pricing is in the range of a few cents per image at the API level, with platforms typically marking this up modestly.

FLUX.1 [dev]

Released the same day as [pro], this is a 12-billion-parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. open-weights model that has been distilled from [pro] using a technique called guidance distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.)., which bakes the classifier-free guidance behavior directly into the model weights. The result is a model that generates in roughly 20 to 50 sampling steps (versus the much slower undistilled version) at quality very close to [pro]. The license is non-commercial, meaning you can use it for personal and research projects but you need to buy a separate commercial license for business use.

FLUX.1 [dev] is the model that the open-source community has built on. It runs locally on a consumer Graphics Processing Unit (GPUGPU (Graphics Processing Unit)The parallel processor that runs image and video models. Its speed and memory largely determine what you can generate and how fast.) with about 16 gigabytes of VRAMVRAM (Video RAM)The memory on a GPU. It sets a hard ceiling on the model size and image resolution you can run locally. (or 8 gigabytes with quantizationQuantizationShrinking a model by storing its weights at lower numeric precision, for example 8 or 4 bits instead of 16, so it fits in less memory and runs faster, usually with a small quality cost.), it has been fine-tuned in thousands of variants, and it is the de facto base for almost all the LoRAs and custom checkpoints you will find on platforms like CivitaiCivitaiA community website for sharing Stable Diffusion and FLUX models, LoRAs, and embeddings.. When somebody says 'I trained a FLUX Low-RankRank (LoRA rank)The inner dimension of a LoRA's two small matrices, which sets how much capacity it has to learn. Higher rank captures more complex behavior but risks overfitting and produces larger files. Adaptation (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.),' they almost certainly mean a FLUX.1 [dev] LoRA.

FLUX.1 [schnell]

'Schnell' is German for 'fast,' which is appropriate for a Munich-Freiburg lab. This is a heavily distilled four-step variant of FLUX, generating images in roughly two seconds on consumer hardware. The license is Apache 2.0, the most permissive open-source license, making this the FLUX variant you can use commercially without paying. The trade-off is meaningful quality loss compared to dev or pro. For prototyping, exploration, real-time interactive applications, and any use case where you need to generate at scale, schnell is the workhorse. For finished work, you will probably want dev or pro.

FLUX 1.1 [pro]

Released October 2024, two months after the original launch. This is a meaningful upgrade, six times faster than the original pro (4.5 seconds per generation instead of around 25), with improved prompt adherencePrompt adherenceHow faithfully a model follows the specifics of your prompt, meaning the objects, counts, colors, spatial relationships, and text you asked for. High adherence means you get what you described, not just something in the right neighborhood. and image quality. This is the version that, through late 2024, traded blows with the very best on quality. The scoreboard position mattered less than what it signaled: a lab built on open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models. was matching the closed frontier, which is a large part of why the open community rallied to FLUX in the first place. The codename used in the leaderboard before launch was 'Blueberry,' which is one of those internal-codename-leaks-to-the-public stories that the AI community loves.

FLUX.1 Kontext

Released May 2025, Kontext represents a significant architectural extension. Where the earlier FLUX models do text-to-image generation, Kontext is a unified model that handles both generation and instruction-based image editing in the same network. You can give it an existing image plus a text instruction ('change the background to a beach,' 'put the character in a red jacket,' 'remove the person on the left'), and it will modify the image in-place while preserving identity, style, and unchanged regions.

The trick that makes Kontext work is in-context conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image., the input image is passed to the model as additional context, the way text would be, rather than as a starting state for diffusion. The model has been trained on enormous volumes of (original image, instruction, edited image) triplets, so it has learned what kinds of changes correspond to what kinds of instructions. Kontext comes in three variants: [pro] for fast production work, [max] for the highest quality experimental results, and [dev] as an open-weights 12-billion-parameter model that the community can download and use. Black Forest Labs reports that Kontext is 8 to 10 times faster than equivalent image editing in OpenAI's GPT-4o, and it is one of the headline editing models accessible through Krea.

FLUX.1 Krea [dev]

Released July 2025 as a research collaboration between Krea and Black Forest Labs. The story here is interesting and worth telling in detail because it illustrates how the platform layer interacts with the model layer.

Krea, the creative platform, had been observing for some time that even the best image models had a characteristic 'AI look', slightly oversaturated colors, a certain plasticky quality to skin and surfaces, an aesthetic homogeneity that made it easy to tell at a glance that an image had been generated rather than photographed. They believed this look was not inherent to the architecture but rather a side effect of how the models were trained, the data and the loss functions had inadvertently optimized for a particular aesthetic that, once you noticed it, you could not unsee.

Krea worked with Black Forest Labs to do extensive post-training on top of FLUX [dev]'s base weights, with a deliberately curated dataset and training objective designed to break the model out of the AI lookThe AI lookThe polished, oversmooth, conventionally pretty aesthetic that marks an image as machine-made, a result of models regressing toward the average of their training data. and produce more naturalistic, visually diverse output. The result was Krea-1, Krea's proprietary closed-model. The open-source version of this work was released as FLUX.1 Krea [dev], which is architecturally compatible with the rest of the FLUX dev ecosystem but produces meaningfully different aesthetics. Black Forest Labs described the model as 'opinionated,' meaning it has a strong aesthetic point of view that surprises users with how natural the outputs feel. All existing FLUX dev LoRAs and tools work with it, so it is a drop-in upgrade for many workflows.

FLUX.2 (the current flagship)

Released November 25, 2025, FLUX.2 is the current Black Forest Labs flagship and the model the document refers to whenever it says 'FLUX' in a present-tense context. The family extends the FLUX.1 architecture with several production-grade improvements: native 4-megapixel photorealistic output (versus FLUX.1's typical 1-megapixel ceiling), multi-reference conditioning that supports up to ten input reference images for style and 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., significantly improved text rendering for images with legible typography, better handling of complex spatial composition, and JSON-structured prompting support that lets you pass structured scene descriptions instead of flat prose. The base architecture is still MM-DiTMM-DiT (Multimodal Diffusion Transformer)A variant of DiT that processes text and image (or audio) tokens in separate parallel streams joined by attention. Used by Stable Diffusion 3, FLUX, and most newer models. Improves prompt following significantly. with rectified flow, but the backbone has been scaled up, the training data refreshed with heavy use of synthetic captions, and the text encoderText encoderA neural network that converts text prompts into numerical embeddings that the diffusion model can use. Modern models use combinations including CLIP and T5. stack upgraded.

FLUX.2 ships in five variants, each targeting a different point on the quality-speed-license curve. FLUX.2 [pro] is the hosted production tier with the highest quality at standard latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput., accessed through the Black Forest Labs API or partner platforms like fal.ai, Replicate, and Krea. FLUX.2 [max] is the even-higher-quality variant targeted at hero-asset production work where you want the best possible single image and do not mind waiting a bit longer. FLUX.2 [flex] is an interesting variant that exposes a 'steps' parameter directly to the caller, letting you trade off quality for latency at inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. time (6 steps for fast iteration, 20 steps for default quality, 50 steps for the highest detail). FLUX.2 [dev] is the open-weight variant with a commercial license, the spiritual successor to FLUX.1 [dev] and the model the open community is now building on. FLUX.2 [klein] is the fast small variant, released January 15, 2026, optimized for interactive real-time generation.

The most strategically interesting decision in the FLUX.2 release was opening the source on the FLUX.2 VAEVAE (Variational Autoencoder)A neural network that learns to compress images into a small numerical representation (a latent) and decompress them back. Used in latent diffusion models to make the generation process tractable. under the Apache 2.0 license. The VAE, the variational autoencoder that compresses images to and from latent spaceLatent spaceThe compressed numeric representation an image is turned into before diffusion. Working here instead of on raw pixels is what makes modern image models fast enough to run., is typically where labs hide their competitive advantage, because the quality of the VAE caps the quality of every diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image. trained on top of it. Black Forest Labs releasing theirs openly was a major move, it lets the rest of the open community train new models against a frontier-quality VAE without having to develop one from scratch, and it deepens Black Forest Labs' positioning as the center of the open ecosystem. A March 3, 2026 update doubled generation speed across the FLUX.2 family without quality loss, primarily by optimizing the sampling schedule and integrating improvements from the Spectrum training-free speedup technique covered in the research-frontier part.

FLUX.2's strengths are photorealism, prompt adherencePrompt adherenceHow faithfully a model follows the specifics of your prompt, meaning the objects, counts, colors, spatial relationships, and text you asked for. High adherence means you get what you described, not just something in the right neighborhood., and the quality of its multi-reference conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image.. For product photography, hero ad assets, brand imagery, and any use case where realistic lighting, correct proportions, and professional camera optics matter, FLUX.2 [pro] or [max] is typically the best choice. The multi-reference conditioning is also the best in the field for maintaining style or 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. across a batch of related images, which matters enormously for brand work. Text rendering is much improved over FLUX.1 (which could reliably render about 8-12 word signs) but still not at the level of GPT-Image or Ideogram V3 for dense typography.

FLUX.2's weaknesses are artistic flexibility and stylized aesthetics. The model is optimized for photorealistic and semi-photorealistic work, and it is noticeably weaker at producing the kinds of dreamlike, painterly, or surreal outputs that Midjourney 7 excels at. If your brand visual language is photographic and polished, FLUX.2 is the clear leader. If it is illustrative, artistic, or stylized, Midjourney V8 or Imagen 4 Ultra is probably a better fit. FLUX.2 is also, at mid-2026 pricing, one of the more expensive hosted models to run at scale, partly because of the 4-megapixel output default. Most operators running it at volume use the [dev] variant on their own infrastructure or the [klein] variant for high-volume use cases where perfection is not required.

Black Forest Labs sits in an unusual position. It is the only major image lab that is simultaneously (a) shipping frontier-quality closed-source hosted models that compete directly with OpenAI's GPT-Image and Midjourney, (b) shipping open-weight models that power most of the open-source image ecosystem, and (c) remaining a small, focused, Germany-based lab with a fraction of the headcount of its American competitors. This three-way positioning is what venture investors mean when they describe the company as the European answer to the American closed-model labs, and it is also what lets BFL extract value from both the enterprise APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. customers (who pay for hosted access to [pro] and [max]) and the open ecosystem (which depends on [dev] and [klein] and sends BFL an enormous amount of engineering signal and reputation in exchange).

The open question for 2026 is whether BFL extends into video generation. The lab has signaled interest, the Series B funding round in December 2025 explicitly mentioned video as a future direction, and a text-to-videoText-to-videoGenerating video from text alone, with no starting image. model called SOTA has been referenced in public Wikipedia entries as under development as of February 2026. If BFL ships a frontier-quality open-weight video model, it becomes one of the most strategically important labs in the entire field, because no one else is in a position to both challenge the Chinese video labs (Kling, Wan, Seedance) at quality and also give that capability to the open community. If BFL stays focused on images, it will remain the dominant image model labModel labA company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google. but cede the video frontier to the larger and better-funded Chinese and American players.

The FLUX story stopped being only about images in late 2025. FLUX.2, in November 2025, paired the rectified-flow transformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. with a Mistral-3 vision-language component and multi-reference conditioningConditioningAny extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image. of up to ten images, and shipped an open 32B dev checkpointCheckpointA saved set of model weights. The .safetensors or .ckpt files you download to run a model. alongside the proprietary tiers. Then on July 23, 2026, Black Forest Labs unveiled FLUX 3, a single flow model trained jointly across image, video, audio, and action, generating video with native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward. in clips reported up to about twenty seconds. Shipped the same day was FLUX mimic, a robotics model that trains a small action decoder on the intermediate features of the video-prediction path and was tested at Audi for soft-body manipulation. This is the same lab that anchored the open image era declaring, in one release, that it wants to be the backbone of visual and physical intelligence.

The money backs the ambition. In December 2025 the company raised a 300-million-dollar Series B at a 3.25-billion-dollar valuation, co-led by Salesforce Ventures with a16z, Nvidia, and even Canva and Figma Ventures participating, which is a telling investor list for a company whose models already power features inside Adobe, Picsart, and ElevenLabs. The open question is execution: as of the launch there is no FLUX 3 technical report, no confirmed parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. count, and the cross-model benchmarks are labeled preliminary, so a roughly seventy-person startup is betting it can compete on image, video, and robotics at once against far larger labs.

FLUX comes in an open tier and a pro tier, and the licensing is the catch. The fully permissive option is FLUX.1 schnell, Apache 2.0 and step-distilled for fast local generation, while the stronger dev checkpoints (FLUX.1 dev, Kontext dev, FLUX.2 dev) are open-weight but non-commercial by default, so any paid use needs a separate commercial license from Black Forest Labs. The best raw quality stays behind the proprietary pro, flex, and max endpoints, reached through the company's APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. or through fal and Replicate, and embedded in consumer apps like Adobe and Krea. So the honest shape is: schnell for free and fast, dev for the best you can run yourself once you sort the license, pro when you want the ceiling and will pay per image.

On the craft, FLUX is strongest at photorealism, typography, and, since Kontext, in-context editing, so reach for Kontext or FLUX.2 multi-reference when you need to change one thing about an existing image or hold a character and product consistent across a set. It is weakest where you need a permissive commercial open license out of the box, which is where a truly open model like Wan or an Apache-licensed schnell build wins, and, for now, where you need mature large-scale video, since FLUX 3 is early access with staggered availability.