●Change8

v0.34.0

📦 diffusersView on GitHub →
✨ 10 features🐛 1 fixes🔧 7 symbols

Summary

This release introduces several major video and image generation pipelines including Wan VACE, Cosmos Predict2, and Chroma, while significantly improving torch.compile performance and memory optimization techniques.

Migration Steps

  1. To load FusionX models, use WanTransformer3DModel.from_single_file() with the safetensors URL.
  2. To use LoRAs with Wan2.1, use the pipe.load_lora_weights() method pointing to the specific safetensors file.
  3. For optimized performance on Flux, combine pipe.enable_model_cpu_offload() with pipe.transformer.compile().
  4. When using torch.compile with offloading, it is recommended to increase torch._dynamo.config.cache_size_limit or recompile_limit.

✨ New Features

  • Added Wan VACE video generation pipeline supporting 1.3B and 14B variants with controllable generation (Depth, Pose, Sketch, etc.).
  • Added Cosmos Predict2 Video2World and Text2Image models (2B and 14B variants) for physical AI and world modeling.
  • Added LTX 0.9.7 and its distilled variants for video generation.
  • Added Hunyuan Video Framepack and F1 variants for long video generation.
  • Added support for FusionX models and LoRAs via from_single_file() and load_lora_weights().
  • Added support for AccVideo and CausVid distillation LoRAs to speed up generation.
  • Added Chroma (8.9B parameter model) based on FLUX.1-schnell with Apache 2.0 license.
  • Added VisualCloze framework for universal image generation via visual in-context learning.
  • Enhanced torch.compile support for Flux models to reduce recompilation and graph breaks.
  • Enabled compatibility between torch.compile, model CPU offloading, and quantization (BitsAndBytes).

🐛 Bug Fixes

  • Fixed recompilation and graph break issues in torch.compile for widely used models like Flux.

🔧 Affected Symbols

WanTransformer3DModelWanTransformer3DModel.from_single_fileDiffusionPipeline.load_lora_weightsDiffusionPipeline.enable_model_cpu_offloadFluxPipelineBitsAndBytesConfigtorch.compile