YangLing0818/RPG-DiffusionMaster

Training-free framework that leverages multimodal LLMs (GPT-4, Gemini-Pro, DeepSeek-R1, o1) as prompt recaptioners and regional planners, combined with regional diffusion to decompose complex text prompts into spatially-aware generation tasks. Supports both proprietary and open-source LLM backbones with multiple diffusion models (SD v1.5, SDXL, IterComp) via the Hugging Face diffusers library, enabling high-resolution compositional image generation without additional training. Flexible architecture that generalizes across arbitrary MLLM and diffusion model combinations for improved text-to-image fidelity on complex multi-object scenes.

1,843 stars. No commits in the last 6 months.