●Change8

v0.36.0

📦 diffusersView on GitHub →
✨ 13 features🐛 6 fixes🔧 10 symbols

Summary

This release introduces several major image and video pipelines (Flux2, Hunyuan 1.5, Sana-Video), high-performance attention backends via the 'kernels' library, and the TaylorSeer caching method for significant speed improvements.

Migration Steps

  1. Install the 'kernels' library via 'pip install kernels' to use new attention backends.
  2. Update attention backends using 'pipe.transformer.set_attention_backend("_flash_3_hub")' or similar.
  3. Review the new modality-organized documentation to locate updated pipeline paths.

✨ New Features

  • Added Flux2 image generation and editing pipeline supporting multiple input images.
  • Added Z-Image 6B parameter image generation model.
  • Added QwenImage Edit Plus with multi-image reference capabilities.
  • Added Bria FIBO for precise control using structured JSON captions.
  • Added Kandinsky 5.0 Image Lite (6B) and Video Lite (2B) models.
  • Added ChronoEdit for image editing via temporal video reasoning.
  • Added Sana-Video with linear attention for long video sequences.
  • Added HunyuanVideo-1.5 (8.3B) for high-quality motion coherence.
  • Added Wan-Animate for character animation and replacement.
  • Introduced 'kernels'-powered attention backends for Flash Attention 2/3 and SAGE.
  • Integrated TaylorSeer cache for up to 3x speedups.
  • New LoRA fine-tuning script for Flux.2 with consumer GPU optimizations.
  • Introduced AutoencoderMixin and AttentionMixin to streamline model codebases.

🐛 Bug Fixes

  • Fixed clapconfig for text backbone in audioldm2 tests.
  • Removed redundant RoPE Cache in Qwen Image.
  • Changed missing imports for custom code from errors to warnings.
  • Fixed incorrect temporary variable key when replacing adapter names.
  • Fixed Kandinsky5 No CFG issue.
  • Added _skip_keys for AutoencoderKLWan.

🔧 Affected Symbols

AttentionMixinAutoencoderMixinAutoencoderKLWanFlux2PipelineSanaVideoPipelineKandinsky5PipelineHunyuanVideoPipelineChronoEditPipelineVAETesterMixinset_attention_backend