hugging-apps/flux-cmo-compositional

FLUX.1-dev CMO — Compositional Text-to-Image

This Space demonstrates Correlation-Weighted Multi-Reward Optimization (CMO) applied to FLUX.1-dev. CMO improves compositional text-to-image generation by analyzing the correlation structure among concept-wise rewards (object existence, attributes, numeracy, size, spatial relations) and adaptively weighting conflicting concepts during training.

Model

• —Base: black-forest-labs/FLUX.1-dev
• —LoRA adapter: Bruece/FLUX.1-dev-CMO

Usage

Enter a compositional prompt (e.g. "A red apple is on the left of the yellow banana") and click Run. The model is loaded in bfloat16 with a tiny VAE for fast preview.

License

FLUX.1-dev carries a non-commercial license. See LICENSE.