diffusion-pipe

diffusion-pipe trains image and video diffusion models on your own GPU hardware. It's for people fine-tuning models that are too large to fit on a single GPU or who want to distribute training across several GPUs.

• Multi-GPU training: Divide a model across GPUs and distribute training data across model instances using DeepSpeed.
• Image and video models: Train SDXL, Flux, Stable Diffusion 3, Qwen-Image and Qwen-Image-Edit, alongside video models such as Wan and HunyuanVideo, using captioned image or video datasets.
• Lower VRAM use: Reduce memory needs with block swapping on supported models, activation checkpointing and an 8-bit optimizer; LoRA training also supports quantized models.
• Training evaluation: Track TensorBoard metrics and evaluate on held-out data to measure performance beyond the training dataset.
• Checkpoints and exports: Save training state to resume interrupted runs, and export models and LoRAs as safetensors weights.

The Python project is open source under GPL-3.0 and requires DeepSpeed. Windows users can run it through WSL 2; native Windows training is difficult or impossible because DeepSpeed has only partial support. MiniMax H3 supports text-to-image and text-to-video-with-audio training. Qwen-Image-2.1 supports text-to-image training only, and video datasets can't use animated WebP files.