FLUX.1 Kontext (Cog inference)

Compact Cog wrapper around Black Forest Labs' FLUX.1 Kontext dev model. It loads the Transformer, auto-encoder, CLIP/T5 text encoders, and optional NSFW safety checker, then exposes a single predict endpoint that performs image-to-image editing or style transfer conditioned on a text prompt.

# basic usage
cog predict -i prompt="make the hair blue" -i [email protected]

Everything required (weights download, Torch 2 compilation, etc.) happens automatically on first run.

Licensed under Apache-2.0 for the wrapper code; see model card for FLUX.1 license.

Performance Optimizations

• torch.compile is used in dynamic mode
• the two linear layers in the single stream block are quantized to run in FP8, using a modified version of aredden's fp8 linear layer
• taylor seer style activation caching, enabled by the go_fast option in the cog predictor. May cause quality degradation for more complex editing tasks.
• enable pytorch's cudnn attention backend