●Change8

v0.28.0

Breaking Changes
📦 comfyuiView on GitHub →
⚠ 3 breaking✨ 19 features🐛 22 fixes🔧 20 symbols

Summary

This release introduces new features like text overlay and advanced 3D nodes, alongside numerous bug fixes and performance optimizations. It also includes updates to partner nodes and documentation, while dropping support for PyTorch 2.4 and removing some older nodes.

⚠️ Breaking Changes

  • Support for PyTorch 2.4 has been dropped. Users on PyTorch 2.4 will need to upgrade to a supported version.
  • IdeogramV1 and IdeogramV2 nodes have been removed. Users should migrate to newer Ideogram nodes if available or find alternative solutions.
  • StabilityAI nodes have been removed. Users should migrate to alternative nodes or solutions.

Migration Steps

  1. Migrate from IdeogramV1 and IdeogramV2 nodes to newer Ideogram nodes or alternative solutions.
  2. Migrate from StabilityAI nodes to alternative nodes or solutions.
  3. Upgrade PyTorch to a version other than 2.4 if you are currently using PyTorch 2.4.

✨ New Features

  • Added AGENTS.md documentation.
  • Added support for Seed Audio 1.0 via Partner Nodes (ByteDance).
  • Implemented a basic text overlay node (CORE-137).
  • Added support for Seedream 5 Pro model via Partner Nodes (ByteDance).
  • Added namespaced model_type tags and aligned tag semantics.
  • Added --comfy-api-base launch argument to target ephemeral testenvs.
  • Exposed effective extension filter on /experiment/models via API.
  • Added support for convrot int4 models.
  • Enabled comfy-kitchen Triton backend by default on ROCm/AMD (with specific GPU matrix-core requirement).
  • Added SeedVR2 support (CORE-6).
  • Added a widget to disable thinking in Seedream Partner Nodes.
  • Added Save Text Node (CORE-176).
  • Added bboxes input to Create Bounding Boxes node.
  • Added Save 3D (Advanced) node family (CORE-329).
  • Added support for PID 1.5 models.
  • Added ComfyUI Core version in request headers for Partner Nodes (client).
  • Added support for single-channel images in Save Image (Advanced).
  • Rerouted Gemini Image preview models to release versions (Partner Nodes - Google).
  • More optimized int8 and int4 on Turing GPUs.

🐛 Bug Fixes

  • Fixed Qwen3-VL tokenizer crash with custom embeddings.
  • Fixed error when embedding is loaded with models using llama_template.
  • Fixed bug when setting transparency in color picker.
  • Masked authorization headers in logs (Partner Nodes - logs-auth).
  • Skipped unloading dynamic model patchers in the current workflow.
  • Fixed crash on videos with undecodable audio streams (Video).
  • Fixed crash when using UNetSelfAttentionMultiply.
  • Fixed AttributeError in VAE.is_dynamic() for VAEs constructed without a patcher.
  • Fixed qwen3vl reference images when used as text encode models.
  • Fixed Load3D failing path validation from double path resolution.
  • Undeprecated the text node.
  • Updated preview nodes.
  • Attempted to fix issues with the seedvr VAE.
  • Attempted to fix flash attention related issues on AMD.
  • Fixed SaveVideo description: it incorrectly stated it saved images when it saves video.
  • Corrected HLG inverse-OETF clamp in hlg_to_linear (image).
  • Fixed hidream o1 regression.
  • Fixed black image on Turing GPUs when using int4 models.
  • Restored Tencent3DPartNode FBX output (Partner Nodes - Tencent).
  • Fixed 3D advanced nodes crashing in API mode (no UI).
  • Fixed FP8 activation quantization for >2D activations in mixed_precision_ops.
  • Fixed int8 performance regression.

Affected Symbols

Qwen3-VL tokenizerIdeogramV1IdeogramV2StabilityAI nodesVAE.is_dynamic()UNetSelfAttentionMultiplyTencent3DPartNodeSave Image (Advanced)SaveVideohlg_to_linearhidream o1comfy-kitchen Triton backendSeedVR2SeedreamSave Text NodeCreate Bounding Boxes nodeSave 3D (Advanced) node familyPID 1.5 modelsGemini Image preview modelstext overlay node