Key Takeaways
- Dustin Podell built FLUX Kontext at Black Forest Labs to eliminate the "slot machine" workflow where users repeatedly regenerate images to tweak minor details.
- Foundation models already contain latent spatial understanding, meaning developers can unlock editing features through data curation rather than altering network architectures.
- Podell unlocked general in-context image editing using a four-to-six-hour training run on a small photography dataset annotated with simple spatial relationships.
- Andreas Blattmann points to controllability as the primary hurdle for image generation, framing character consistency as a data design challenge rather than a compute bottleneck.
The Method
Dustin Podell wanted to eliminate generative image slot machines. For years, creators using diffusion models suffered through the same loop: submit a prompt, review four images, find one decent composition with a mangled hand, tweak two words, and roll the dice again. The original character vanished on the next spin. Podell suspected that foundation models did not need structural rewrites to preserve identity. They already contained spatial knowledge from pre-training. They simply needed a dataset that taught them how to edit.
Instead of retraining foundation weights or adding complex attention control layers, Podell assembled a small, clean photography dataset. The data focused entirely on relative spatial changes: camera perspectives, positions, and light directions. He trained the model for four to six hours.
The experiment worked immediately. Feeding a reference image alongside an edit instruction allowed the model to modify attributes while keeping the subject stable. Andreas Blattmann noted that this shift unlocked practical consistency: “It was the first kind of model to really do this character consistent editing right in high speed that made it again accessible to so many people.” (play from 1:00:28) Black Forest Labs turned that side project into FLUX Kontext, using multi-reference context expansions in FLUX 2 to push editing precision further.
Where This Breaks Down
This method relies on pre-trained representations that already exist. If a foundation model has not already learned how light wraps around a human face or how fabric drapes over an arm, six hours of fine-tuning cannot teach it. The small dataset functions as a steer, not an educator. Run this workflow on an under-trained checkpoint and the model outputs smeared artifacts instead of coherent edits.
The approach also runs into walls when users demand complex composition changes that conflict with the reference frame. If you ask the model to preserve a character while placing them in an extreme lighting scenario outside the base model's known distributions, identity preservation deteriorates. Multi-character scenes present an identical hurdle: when two reference figures share a prompt context, their visual features tend to bleed together.
What to Do With This
Audit your model fine-tuning roadmap. If your team is debating new adapter modules or custom loss functions to force consistent outputs, pause the engineering work. Assemble 200 high-quality image pairs that isolate the exact spatial or stylistic modification you want, write plain text captions explaining the difference, and run an overnight fine-tuning job on your existing base checkpoint. Test whether the capability already exists in your weights before you write custom code.