Black Forest Labs is moving from image generation toward models that handle images, video, audio, and actions in one system.
2
Self-supervised training can remove the need for separate modality-specific encoders and improve generation across images, video, and audio.
3
Real-time visual models and world models are intended to support interactive media, robotics, self-driving, and manufacturing automation.
Summary
Stephen Batifol traces Black Forest Labs' model releases from FLUX.1, a text-to-image model, through Kontext, FLUX.2, and FLUX.2 Klein. The company has focused on open releases, model quality, image editing, reference consistency, and faster generation. He then explains why standard generative training relies on external encoders that teach models visual representations, and why this creates scaling and multimodal problems. BFL's Self-Flow research combines representation learning and generation through a student-teacher setup with different noise levels. Batifol shows research results for images, video, audio, combined video and audio, and robot actions. The longer-term direction is visual intelligence: models that understand relationships, geometry, and interactions in the world. Faster models could support interactive image editing, games, and films. World models could train agents for robotics, self-driving, and manufacturing. Batifol is direct about the limits of the research models and the company's unwillingness to disclose sensitive training data.
FLUX began as an open text-to-image model that could run locally
Batifol introduces Black Forest Labs as the team behind Stable Diffusion, Latent Diffusion, and the FLUX models. FLUX.1 launched in August 2024 as a text-to-image model released in open source. He says it could run on a laptop and gained attention for image quality, especially anatomy, compared with larger models. The first release established BFL's pattern of publishing capable models openly while also working with customers such as Microsoft, Adobe, Canva, and Mistral.
Kontext combined image generation with editing and made visual stories easier to build
BFL's Kontext model combined text-to-image generation with image editing. Batifol shows edits that remove a snowflake from a character's face, move the character to Freiburg, and change the background while preserving character consistency. He also describes storyboard use cases, where a sequence of edited images can provide input frames for video or animation models. At the time, he says Kontext could edit or generate in roughly seven to eight seconds, while other systems could take 40 or 50 seconds.
FLUX.2 added multi-reference editing and stronger visual consistency
FLUX.2 is presented as BFL's best model at that point and its first multi-reference release. Batifol shows generated people, animals, product photography, and scenes, then demonstrates editing with six input images. A prompt can combine those references into a coherent outfit, with the jacket and tie worn correctly. Another example places a sofa into a customer's room. FLUX.2 can take up to 10 images at once and is designed for character, product, and style consistency.
External encoders improve generative training but limit scaling across modalities
Batifol explains that ordinary generative models add noise to images and learn to remove it, without directly learning facts about objects and physical relationships. Researchers therefore use an external image encoder to teach representations such as a glass sitting on a table or a person sitting on a chair. He says this alignment can make training converge 70 times faster, but the encoder creates a scaling ceiling. Separate encoders for images, video, and audio would also produce a complicated system with mismatched objectives.
Self-Flow trains representation and generation together without an external encoder
BFL's Self-Flow paper proposes a self-supervised approach for multimodal generative models. The method combines representation learning and generation in one flow and trains across images, video, and audio. It uses two noise levels: a heavily corrupted input for a student model and a lightly corrupted input for a teacher model. The student learns both generation and representation losses, while the teacher is a more stable version of the student. Scaling the system scales both models instead of leaving an external encoder fixed.
The research model improves text, anatomy, video motion, and audio alignment
Batifol compares Self-Flow with flow matching and says the research model performs better on audio, images, and video while continuing to reduce loss after the baseline plateaus. The examples show more accurate text, improved anatomy, a person doing a push-up with better form, and birds walking without the baseline's flicker. A joint video-and-audio example follows the prompt to say "Hello from the Black Forest" and stops more cleanly than the baseline. He stresses that these are research models, not production releases.
The same multimodal direction can predict robot actions
BFL is also training models on actions. Batifol shows a robot trying to pick up a can and move it closer. The baseline has flickering and awkward arm movement, while the comparison model completes the action more directly with the same number of steps. He describes this as part of BFL's interest in physical AI, extending beyond image, video, and audio generation toward systems that can model and perform actions.
FLUX.2 Klein makes image generation and editing close to real time
FLUX.2 Klein is designed for interactive generation and editing. Batifol gives figures of about 500 milliseconds for editing and 300 milliseconds for generation at its fastest, and later compares its 4B and 9B versions with other open models. In the examples, Klein stays under a second while another model takes around 15 seconds for some tasks and around 20 seconds for multi-reference editing. He argues that this speed allows users to guide visual output as they think instead of waiting for each result.
World models connect visual intelligence to robotics and automation
Batifol describes the longer-term goal as visual intelligence, with models that understand and simulate geometry, relationships, and interactions in the world. Real-time generation could support interactive visual engines for games and films, where scenes are rendered as users prompt them. World models matter to BFL because they could generate environments for training agents. He names robotics and automation, scaling self-driving, and automating manufacturing as applications for that direction.
"This is where we're really interested is like not only image generation or video, but it's also doing actions and doing more things toward physical AI."18:21
Who should watch
You are building image or video products and need editing that preserves characters, products, or styles across references.
You are researching multimodal generative models and want a concrete explanation of the limits of external representation encoders.
You are working on interactive media, robotics, or world models and want to understand how BFL connects generation with actions.
