OpenFLUX.1
huggingface.co
huggingface.co
Context: FLUX is state of the art text-to-image generator. It has very strong prompt adherence, and is able to render complex scenes, though it somewhat lacks in terms of ability to produce classical art and a few other areas.
When FLUX launched, they released 3 models:
FLUX Pro: Weights not available.
FLUX Dev: Distilled (guidance distilled) version of FLUX Pro. Requires >20 steps for good image. Weights available. Non-commercial license.
FLUX Schnell: Distilled (timestep distilled) version of FLUX Pro. Requires ~1-4 steps for good image. Weights available. Apache 2.0.
---
The thing is that Schnell is very hard to fine-tune due to the distillation applied, and is kneecapped as it's not using true CFG based sampling and being limited to ~4 steps. This is likely why BFL (the company behind FLUX) felt comfortable releasing it under Apache 2.0.
The OpenFLUX project fixes both issues and since it's derived from Schnell, it's also Apache 2.0.
note: depending on whether you're using the official unet models from black forest labs, or a quantized checkpoint, the workflows are different. see: https://comfyui-wiki.com/tutorial/advanced/flux1-comfyui-gui...