modelscope /DiffSynth-Studio

About DiffSynth-Studio

DiffSynth-Studio runs open-source image, video and audio models on your own GPU. The ModelScope community team built it, and it covers both running models and training them. Low-VRAM options and quantization let consumer cards handle big models. A small maintainer team means new features and issue replies arrive slowly.

Best for: Developers and researchers who want to run or fine-tune big generative models on one consumer GPU.

• Runs many open-source image, video and audio diffusion models
• Moves model weights between disk, memory and GPU to fit small VRAM
• Quantizes weights to NF4 or INT8 to cut VRAM use
• Trains base models, LoRAs and adapter models
• Splits training into two stages to save memory
• Pairs with DiffSynth-ComfyUI nodes and DiffSynth-WebUI for LoRA training

1. 1
   Open the README and find the model you want in the supported models list.
2. 2
   Go to that model's folder under examples/ and copy the example script.
3. 3
   Run the script on your own GPU. If you run out of memory, use the low-VRAM or quantized example instead.

• DiffSynth-Studio vs Wan2GP