DiffSynth-Studio
An open-source Python diffusion engine generates images, video and audio locally, with low-VRAM inference and model training under Apache 2.0.
DiffSynth-Studio is a diffusion model engine for developers and researchers generating media or training models on their own hardware. It combines inference and training, with memory-saving features for consumer GPUs.
- Media generation: Work with image, video and audio models in one framework, with support for Qwen-Image, FLUX, Wan and MiniMax-H3.
- Low-VRAM operation: Memory offloading helps large models fit on GPUs with limited memory, while quantization reduces memory demands for inference and LoRA training.
- Model training: Train base models, LoRAs and adapters with additional inputs where the selected model supports them; training availability differs by model.
- Workflow interfaces: DiffSynth-ComfyUI brings inference into ComfyUI workflows, while the separate DiffSynth-WebUI provides a LoRA training interface for consumer GPUs.
The Python engine runs locally and is open source under Apache 2.0. It supports NVIDIA GPUs and provides installation guidance for non-NVIDIA GPUs. EntroPack is an available weight-compression backend for inference and LoRA training, with lossless and controlled-loss compression.
Some older features are no longer maintained and require historical releases. The small development team limits the pace of feature development and issue fixes.
Logotyper
