Black Forest Labs: FLUX, Open Research, and the Future of Visual AI

Do you know Black Forest Labs—or do you know FLUX? Stephen Batifol opens with those two questions before introducing the connection. Speaking as a developer relations engineer at BFL, he describes a team whose research lineage includes Stable Diffusion and Latent Diffusion, alongside FLUX. Batifol reports more than 200,000 academic citations across the team and names Microsoft, Adobe, Canva, and Mistral as customers. The company combines model research with work on enterprise applications.

FLUX.1, released in August 2024, began with text-to-image generation. Batifol describes an openly released model that could run on a laptop, with anatomy that compared favorably with much larger models. Its early reception included a shout-out from Hugging Face’s Clem and, for a time, the position of most-liked model on Hugging Face. That ranking belongs to the launch history; Batifol explicitly says it no longer holds.

FLUX.1 Kontext extends that starting point from generation to editing. Batifol presents it as a first open-source combination of the two capabilities. The launch announcement establishes the combined workflow, but described Kontext dev as a private beta alongside hosted pro and max models; it does not establish worldwide priority or public availability of every variant at launch. The useful technical distinction is that an existing image can become the context for the next instruction.

The demonstration follows one person through successive changes:

1. Remove the snowflake from her face.
2. Move her to Freiburg, where BFL is headquartered, taking a selfie in the street.
3. Make the background snowy, with snow appearing on her face too.

Character consistency makes these edits a sequence rather than three unrelated generations. The location and weather change while the subject remains recognizable. Batifol recalls Kontext generation or editing taking about 7–8 seconds, compared with roughly 40–50 seconds for early GPT image generation or editing; he does not specify the test conditions.

That continuity also supports storyboards. A seagull wearing a VR headset drinks beer in a bar; a friend joins; the characters acquire hats; then they go outside. Each image advances the story without requiring the characters to be reinvented. Partners and customers used sequences like this as input frames or end frames for video and animation models, giving a motion generator concrete visual destinations.

FLUX.2, released in November 2025, broadens both image quality and editing. Batifol points to hands, veins, and bracelets as details that make the human samples convincing to him, followed by turtles and an animal in a bath. The next examples move beyond people and animals: a waffle presented as product photography and a moped rider carrying balloons. These are his qualitative examples of realism and range.

The editing examples introduce a more consequential capability: multiple references can jointly constrain one output. Six source images accompany an instruction to create an outfit from them. The result puts the jacket and tie in plausible positions instead of merely assembling a collage. A second example places a sofa into a room, letting a furniture maker or e-commerce customer visualize the product in a flat.

FLUX.2 advertises support for up to ten simultaneous reference images at the model-family level. Individual variants and hosting services need not expose that same limit. Batifol identifies character, product, and style consistency as the important properties: the references supply things that should survive the transformation, while the instruction specifies how they should be composed or changed.

BFL’s first operating principle, in Batifol’s account, is to improve model quality with each release. He positions FLUX.2 as its best image model at that point, its first multi-reference release, and a leader among openly available generation and editing models. January’s FLUX.2 Klein then shifts attention toward interactive use. Batifol tentatively recalls fastest-case Klein latency of 500 milliseconds for editing and 300 milliseconds for generation. The release family also makes the licensing distinction concrete: its 4B models use Apache 2.0, while its 9B models use the FLUX Non-Commercial License. Open availability is not a single license shared by every model.

Faster products are only one part of that trajectory. Batifol turns next to BFL’s research-first approach and public papers: improving generation also means changing what the model learns during training.

A glass should rest on a table rather than pass through it. A person sitting in a chair should not intersect the chair. Batifol uses these ordinary relationships to explain the limitation of a training objective built around corrupting images with noise and learning to remove it. Denoising supplies a generation task, but does not explicitly state those physical relationships.

Representation alignment adds another learning signal. An external image encoder supplies features that describe image content, and the generator learns to align its internal features with them. Batifol describes the encoder as teaching the generator about objects and their relationships, contrasting its visual analysis role with the generator’s synthesis objective. In the displayed external-alignment comparison, Batifol reports 70-times-faster convergence; the baseline, target metric, and compute accounting are not specified in his explanation.

That extra supervision introduces three constraints:

• Scaling ceiling: the external encoder is a frozen checkpoint. Enlarging the generator does not enlarge the source of its representation targets.
• Modality dependence: an image encoder such as DINOv2 does not supply the corresponding targets for audio and video. Adding separate encoders for each modality complicates the training system.
• Objective mismatch: features useful for visual analysis are not automatically the best features for generation. The two models were trained to do different things.

The encoder comparison makes the last limitation concrete. DINOv3 may be a stronger encoder, yet the displayed generation experiments perform worse with its representations than with DINOv2’s. The chart plots FID against training steps; it concerns the quality of generators trained with those encoders, not a general ranking of DINO models. A better external representation model does not guarantee a better generation target, and Batifol says there is no clear selection rule that resolves the mismatch.

Can a generator learn representations without an external encoder? Self-Flow, published roughly a month and a half before the talk, combines representation learning and generation in one self-supervised training approach. The intended scope includes images, video, and audio. Instead of importing a fixed teacher from another task, it develops the teacher alongside the generator.

The training procedure uses two differently corrupted views of the same asset:

1. Create the noisy views. Batifol describes a heavily corrupted view and a lightly corrupted view. More precisely, the paper gives the student a mixture of token noise levels and the teacher the lower-noise view.
2. Run the student and teacher. The student handles the harder generation task. The teacher is a more stable version of the student, maintained through an exponential moving average of its weights.
3. Train for generation and representation together. The student learns to generate while also matching the teacher’s representation targets.

The cleaner view gives the teacher a less ambiguous input; the student must learn useful structure despite stronger corruption.

The central change is architectural: the representation teacher grows with the generator. Scaling the model scales both student and teacher, rather than leaving a fixed external encoder as the source of supervision. The approach also avoids having to import a separate representation model for every modality. Batifol says BFL is already using it in models under training.

The next comparisons use research models, not production releases. Batifol reports better results for Self-Flow than the flow-matching baseline across audio, image, and video generation in the displayed experiments. Dashed curves mark the baseline and solid curves mark Self-Flow. He also points to faster convergence: the baseline begins to plateau while Self-Flow’s loss continues falling. His suggestion that this advantage would persist toward two million training steps is an extrapolation, not a measured endpoint.

The image examples translate that training signal into recognizable errors. A baseline attempt at “the future is FLUX” has missing or duplicated letters; another word, “worlds,” contains a doubled L. In the Self-Flow examples, Batifol points to more coherent lettering, including text on a mirror and a tree. A further side-by-side text comparison repeats the pattern, followed by a face whose anatomy looks less distorted with Self-Flow. These samples illustrate what improved representations can change without implying production-perfect faces or typography.

The video demonstrations extend the comparison to motion. In a push-up clip, Batifol contrasts awkward baseline movement with a Self-Flow result whose arms, hair, and exercise form appear more coherent. Next, a walking-bird example contrasts baseline flickering and odd behavior with what he describes as a clean, flicker-free Self-Flow sequence. These are temporal artifacts: individual frames are not enough if the subject changes implausibly from one moment to the next.

Joint training also permits joint generation. The audiovisual example asks a person to say “Hello from the Black Forest.” Batifol plays the baseline first and points out unwanted sounds at the end. He then plays the Self-Flow version, which he describes as stopping after the requested phrase. The distinction is not only whether recognizable speech appears, but whether the generated sequence ends where the instruction ends.

Batifol connects this audiovisual result to the preceding image and video work, then extends the discussion to action prediction. The robot’s task is concrete: pick up a can and bring it closer. At the same number of steps, he describes the baseline arm as flickering and behaving erratically, while the Self-Flow result picks up the can and moves it directly. The paper’s robotics experiment uses an RT-1 fine-tuned model evaluated in SIMPLER simulation. This is transfer from multimodal training into an action task, rather than evidence that one unchanged checkpoint controls a physical robot. It introduces BFL’s interest in physical AI beyond media generation.

Returning to available image tools, Batifol shows Klein running through Krea. The display changes continuously, but it is not a video model: it is a succession of image edits. That distinction matters because a sufficiently fast image model can support an interactive workflow without generating a temporally modeled video sequence.

Batifol compares Klein’s 4B and 9B variants with other openly available models, describing quality as at least comparable while latency is much lower. His approximate readings of the comparison are:

The vendor’s benchmark caption specifies GB200 hardware and bf16 precision. The talk does not spell out every model version, resolution, or timing boundary, so these figures describe the presented comparison rather than expected latency on a laptop or arbitrary hosted service. The plots place Elo against latency to show the quality–speed tradeoff across the three tasks.

The application goal is to remove the pause between an idea and its visual feedback. Mock-ups could be rendered and guided as the user works, rather than requiring repeated waits. Batifol extends that ambition to interactive visual engines for games and films, including the future possibility of rendering a movie while prompting it. These are proposed directions for visual intelligence, beyond the editing demonstration.

World models add another requirement: learning geometry, relationships, and interactions well enough to simulate how a world behaves. Batifol’s motivation is robotics and automation. Generative worlds could become training environments for agents, with longer-term applications in self-driving and manufacturing. The target is no longer just an image that looks right; it is an environment in which actions have useful, learnable consequences.

The audience first asks what data underlies this world-model work. Batifol declines to disclose it, citing trade secrets and data sensitivity, while saying BFL is working with multiple partners. He does not identify those partners or datasets.

A second question asks where an action-prediction model stores the state of the world. Batifol initially describes learned representations and internal memory. Pressed on whether that means external storage or the context window, he identifies context-window tokens: the sequence records movement and informs where the model should be next.

That answer leaves a practical limit. Asked whether the model can run indefinitely or compact its history, Batifol says he is unsure about indefinite operation, expects a limit, and suggests a possible sliding window. He does not describe a confirmed compaction mechanism. The closing engineering question is therefore how to preserve enough state for useful action prediction once the history no longer fits in context.