Image Generation and Visual Intelligence with Black Forest Labs

Dustin Podell, co-founder and researcher at Black Forest Labs, describes image generation quality as having advanced from vague color blobs to near-photorealistic scenes over roughly three to four years, with whole short films now being made entirely with AI generation. He emphasizes that the core technology has not changed dramatically and that steady forward progress rather than a sudden paradigm shift explains the improvement.

Diffusion models work by progressively adding noise to images during training until the original signal is destroyed, then learning to reverse that process at inference starting from a fully noisy image. The same fundamental noise-removal process used in early image models underlies modern video models producing sequences up to ten seconds, and the same technique applied to different mediums drives audio and music generation models such as Suno. Flow matching is a refinement that trains a model to learn an underlying velocity or flow map, but all current leading image and video generation models still perform the same noise-addition and removal process, with architectural and training optimizations rather than process changes driving recent gains.

Defining the best image or video model is difficult because expected outputs vary significantly across user types, and leaderboards rely on general single-preference voting rather than granular task-specific metrics, unlike LLM benchmarks covering programming and math. Seed Dance is described as the current state of the art for text-to-video generation, with Seed Dance 2 supporting 4K generations up to fifteen seconds, while Sora 2 is considered strong despite ranking lower on community leaderboards.

Black Forest Labs released an in-context editing model called Flux Context approximately one year ago, capable of cleaning up photos, adding objects, and handling interrelational edits. Podell notes that for a model to show what happens when a water glass is knocked over, it must model real-world physical relationships and not just visual patterns, reflecting a broader field transition from one-way prompt-to-image generation toward editing and world-modeling capabilities. The Flux 2 family and a smaller variant called Klein introduced the ability to take many image references simultaneously, described as a fundamental change compared to Flux 1, which accepted only one image reference. Current commercial use cases include clothing try-on, home decoration visualization, product photography generation, and safety planning simulations such as modeling what fire exits would look like during a crowd emergency. Podell also sees the internal representations developed by large-scale image generation models as a strong foundation for robotics training.

The Flux One family was released approximately two years ago after a four to five month sprint following the company's formation, and included Flux One Pro on the Black Forest Labs API, Flux One Dev with commercially licensable open weights, and Flux One Schnell, a step-distilled high-speed model with a permissive open license. Flux Two was described as a very large upgrade advancing text-to-image generation, editing, and multi-image omni-edit handling many people and items with complex relationships, built on a 32 billion parameter architecture requiring more powerful hardware to run locally. Black Forest Labs has followed a consistent pattern of releasing a large high-performance model followed by a smaller, faster distilled version, doing this first with Flux One and Schnell and again with Flux Two and Klein.

The Klein series was designed to pack as much performance as possible into a small model for both API use and open weight release, covering text-to-image and editing, and ran faster than the Flux One series. The Klein KV model introduced KV caching to the image editing domain, an optimization technique common in language models that produced a very large speed-up for local editing. Podell estimated approximately 99 percent certainty that Klein can run on a modern M-series MacBook Pro but had not personally tested it and could not provide speed numbers. Black Forest Labs also published a research blog post called Self Flow and has a new release planned for later in the summer without a specific date confirmed.

Podell identified long context truly multimodal models and real-time interaction as the two areas he is most personally excited about for the future of visual intelligence, while noting these comments reflect personal enthusiasm and not a signal of what Black Forest Labs is building. He predicted future models will be able to think visually, generate audio, and maintain full context of user activity over weeks without requiring explicit prompts. He observed that real-time video and audio duplex interaction is very early but will become significant, connecting to robotics use cases where models must take in and reason about the real world in real time, with interactive game-like experiences such as Genie representing early examples of that direction. He expressed hope that long context multimodal capability and real-time interaction will eventually converge into a single unified capability area.