Change8

v0.17.0

📦 comfyuiView on GitHub →
7 features🐛 7 fixes🔧 4 symbols

Summary

This release introduces significant refactoring to the asset loading architecture, adds support for the Flux 2 Klein KV cache model, and includes numerous bug fixes related to model loading and batch processing. The CacheProvider API feature was added and immediately reverted.

Migration Steps

  1. If using Flux 2 Klein KV cache, use the new FluxKVCache node.
  2. If encountering issues with older PyTorch versions (< 2.8.0), the guard against torch.AcceleratorError should improve compatibility.

✨ New Features

  • Added support for Flux 2 Klein KV cache model via the new FluxKVCache node.
  • Added the painter node.
  • Added Reve Image nodes via the api-nodes extension.
  • Added functionality for nodes to add pre attention patches to flux models.
  • Allowed model patches to include a cleanup function.
  • Added pre attention and post input patches support for qwen image models.
  • Implemented a CacheProvider API for external distributed caching (though this was immediately reverted).

🐛 Bug Fixes

  • Fixed text encoder lora loading for wrapped models.
  • Improved installation guidance when comfyui-manager is not installed.
  • Fixed an issue with batch_size > 1 on some models.
  • Guarded torch.AcceleratorError for compatibility with torch versions older than 2.8.0.
  • Fixed frontend compatibility issue for Float gradient_stops format.
  • Lowered KV cache memory usage.
  • Fixed audio extraction and truncation bugs.

Affected Symbols