●Change8

v0.40.0

Breaking Changes
📦 diffusersView on GitHub →
⚠ 1 breaking✨ 12 features🐛 3 fixes🔧 25 symbols

Summary

This release introduces several new pipelines including LTX2.5, MiniMax H3, and Wan Animate 2, and graduates Modular Diffusers to stable support. It also includes minimal support for tensor-parallel.

⚠️ Breaking Changes

  • The `guidance_scale` argument is no longer available in the Wan-Animate-2 pipeline because guidance is now handled by the pipeline's guider component.

Migration Steps

  1. When using the Wan-Animate-2 pipeline, remove the `guidance_scale` argument and manage guidance through the pipeline's guider component.
  2. For `ConditionalPipelineBlocks` with sibling blocks declaring different defaults for the same input, be aware that `combine_inputs` now merges defaults to `None` and stores per-block defaults in `InputParam.defaults_by_block`. `get_block_state` will fall back to the block's own declared default.

✨ New Features

  • Introduced LTX2.5 pipeline with Gemma 4 text encoder and new video decoder.
  • Added MiniMax H3 pipeline for joint video and soundtrack generation.
  • Released MiniMax Music 3 pipeline for generating full songs from lyrics and music descriptions.
  • Integrated Stable Audio 3, a text-to-audio model from Stability AI, with three associated pipelines.
  • Launched Wan-Animate-2 pipeline for animating reference character images with driving video motion.
  • Introduced JoyAI-Image-Edit-Plus for multi-image instruction-guided editing.
  • Enhanced Cosmos 3 coverage with a modular pipeline, transfer support for precomputed control videos, edge support, distilled pipelines, and model parallelism.
  • Modular Diffusers is now stable and out of experimental phase.
  • Added modular pipelines for Cosmos 3 and Krea 2.
  • Anima pipeline now supports img2img blocks.
  • Implemented branch-specific input defaults for `ConditionalPipelineBlocks` to handle conflicting defaults.
  • Added minimal support for tensor-parallel.

🐛 Bug Fixes

  • Fixed a group-offload device mismatch issue.
  • Cleaned up unused `intermediate_inputs` in modular pipelines.
  • Ensured stale auto-docstrings are detected in CI.

Affected Symbols

MiniMaxH3BlocksMiniMaxH3ModularPipelineQwen3VLForConditionalGenerationMiniMax Music 3StableAudio3PipelineStableAudio3AudioToAudioPipelineStableAudio3InpaintPipelineStableAudio3DurationEmbedderAutoencoderSAMELTX2PipelineLTX2VideoTransformer3DModelAutoencoderKLLTX2Videogemma4_unifiedLTX2VideoDiffusionDecoderModelLTX2VideoDiffusionDecodePipelineLTX25AutoBlocksWan-Animate-2Cosmos 3Cosmos3OmniModularPipelineCosmos3DistilledModularPipelineKrea 2AnimaConditionalPipelineBlocksInputParam.defaults_by_blockget_block_state