Change8

v0.32.2

📦 diffusersView on GitHub →
4 features🐛 5 fixes🔧 6 symbols

Summary

This patch release fixes regressions in Flux single-file and LoRA loading (including 4-bit support) and resolves several bugs in HunyuanVideo related to batch processing and LoRA compatibility.

✨ New Features

  • Added support for loading LoRAs into 4-bit bitsandbytes quantized Flux models.
  • Added unload_lora_weights support for Flux Control models.
  • Added support for loading HunyuanVideo LoRAs created from the original repository format.
  • Improved LoRA unloading for expanded Flux LoRAs.

🐛 Bug Fixes

  • Fixed regression in loading ComfyUI format single-file checkpoints for Flux Dev finetunes.
  • Fixed bug preventing HunyuanVideo from running with batch sizes greater than 1.
  • Fixed HunyuanVideo producing NaN values on PyTorch versions earlier than 2.5.
  • Fixed attention mask dimension issues in HunyuanVideo.
  • Fixed a bug affecting multiple LoRA loading in Flux models.

🔧 Affected Symbols

FluxControlNetModelHunyuanVideoTransformer3DModelunload_lora_weightsload_lora_into_text_encoderfuse_lorabitsandbytes