●Change8

v0.3.76

📦 comfyuiView on GitHub →
✨ 13 features🐛 6 fixes⚡ 1 deprecations🔧 5 symbols

Summary

This release introduces significant feature enhancements, including cheap latent preview for flux 2, new API nodes, and improved LoRA training capabilities. It also includes several performance optimizations related to VRAM usage and bug fixes across various components.

Migration Steps

  1. If you rely on chat widgets in OpenAI/Gemini API nodes, note they have been removed.

✨ New Features

  • Added cheap latent preview for flux 2.
  • Added Veo3 First-Last-Frame node in API Nodes.
  • Merged 3d animation node.
  • Added Dataset Processing Nodes and Improved LoRA Trainer Nodes with multi resolution supports.
  • Enabled async offloading by default on Nvidia.
  • Added v2-5-turbo model support to FirstLastFrame node in API Nodes.
  • Support video tiny VAEs.
  • Support some z image lora formats.
  • Added System User protection with '__' prefix for security.
  • Added missing z image lora layers.
  • Made the ScaleRope node work on Z Image and Lumina.
  • Update qwen tokenizer to add qwen 3 tokens.
  • Added Kling O1 model support in API Nodes.

🐛 Bug Fixes

  • Fixed Gemini API node to use first 10 images as fileData (URLs) and remaining images as inline base64.
  • Fixed the CSP offline feature on the latest frontend.
  • Fixed typo in move_userdata dest validation in user_manager.
  • Disabled offload stream when using torch compile.
  • Fixed QuantizedTensor.is_contiguous (#10956).
  • Wrapped the raw stream in a context manager in mm.

🔧 Affected Symbols

nodes_custom_sampler.pyQuantizedTensor.is_contiguousOpenAI nodesGemini nodesScaleRope node

⚡ Deprecations

  • Removed chat widgets from OpenAI/Gemini API nodes.