●Change8

v0.23.0

📦 comfyuiView on GitHub →
✨ 14 features🐛 12 fixes🔧 34 symbols

Summary

This release introduces significant performance improvements via multi-threaded model loading and adds support for new models like MediaPipe face detection, NVIDIA PixelDiT, and Microsoft Lens. It also includes extensive cleanup and standardization across OpenAPI definitions and various partner nodes.

Migration Steps

  1. If using OpenAPI endpoints related to assets, note that the cloud-only job_ids query parameter has been removed from GET /api/assets.
  2. If using logic nodes, note their category has moved to the utils category.
  3. If using LTX AV, ensure correct timesteps are passed to cross-attention AdaLN modulation.
  4. If using LTXVCropGuides, be aware of a fix preventing stray latent frames when guides share a start position.
  5. If using nodes_lt_upsampler nodes, they have been converted to V3 schema.

✨ New Features

  • Implemented MediaPipe face detection (CORE-235).
  • Enabled multi-threaded loading of models from disk, resulting in significant load time speedups and support for offloading to disk (CORE-43,CORE-152,CORE-164,CORE-165,CORE-117).
  • Added OAuth 2.1 and RFC 7591 DCR endpoints to OpenAPI specification (feat(openapi)).
  • Added OpenRouter LLM node ([Partner Nodes]).
  • Added widget for auto upscaling for the ByteDance2Reference node ([Partner Nodes]).
  • Added And, Or, and Not logic nodes.
  • Added new Rodin2.5 nodes ([Partner Nodes]).
  • Added Save Image advanced node (CORE-32).
  • Added support for Microsoft Lens (CORE-248).
  • Added support for NVIDIA PixelDiT and PiD (CORE-201).
  • Added Krea2 Image nodes ([Partner Nodes]).
  • Added Beeble SwitchX nodes ([Partner Nodes]).
  • Added camera intrinsics fields to Load3DCamera info.
  • Added overwrite/increment functionality to SaveImageTextDataSetToFolderNode.

🐛 Bug Fixes

  • Corrected POST /api/assets/import to importPublishedAssets in OpenAPI spec (fix(openapi)).
  • Corrected dimension transposition for single-channel tensors in lanczos operations (CORE-228).
  • Fixed reshaping of LoRA application.
  • Fixed SIGPIPE false negative in Backport Release workflow validation.
  • Fixed --use-flash-attention being ignored when xformers is installed.
  • Corrected description of where compiled Frontend (FE) files live.
  • Fixed invalid BillingStatus schema (object + enum hybrid) in OpenAPI spec.
  • Lowered cache-ram thresholds.
  • Fixed LTXVCropGuides leaving stray latent frames when guides share a start position (CORE-206).
  • Fixed runtime check for SeeDance2 image inputs ([Partner Nodes]).
  • Fixed context window for PiD and corrected lq_latent rounding.
  • Disabled sage attention in stable audio dit and VAE.

Affected Symbols

MediaPipe face detectionopenapiPOST /api/assets/importGET /api/assetsMoGe nodeslanczoslora applicationOpenRouter LLM nodeByteDance2Reference nodeBackport Release workflowAnd nodeOr nodeNot nodelogic nodesRodin2.5 nodesQueueManageResponsePOST /api/queueBillingStatus schemaSave Image advanced nodeMicrosoft LensNVIDIA PixelDiTPiDLTX AVLTXVCropGuidesnodes_lt_upsampler nodesBeeble SwitchX nodesLoad3DCamera infoSaveImageTextDataSetToFolderNodeLoad3D model_fileSeeDance 2Krea2 Image nodesstable audio ditVAEpyav package