Update changelog with comfyui v0.21.0

changelog/index.mdx

@@ -4,6 +4,79 @@
 icon: "clock-rotate-left"
 ---
 
+<Update label="v0.21.1" description="May 13, 2026">
+
+**New Partner Nodes**
+* **Flux2ImageNode and GrokImageEditNodeV2:** Added new nodes for enhanced image generation and editing workflows
+* **ByteDanceSeedreamNodeV2:** New version with DynamicCombo and Autogrow features for improved parameter handling
+* **OpenAI Image node:** New integration with DynamicCombo and Autogrow capabilities
+* **Claude LLM node:** Added Claude language model support for text generation workflows
+
+**Model Support Enhancements**
+* **Anima TE LoRA:** Added support for Anima TE LoRA in Kohya format
+* **HiDream-O1-Image:** Full support for HiDream-O1-Image model with dtype fixes and improved memory usage for non-dynamic VRAM configurations
+
+**Technical Improvements**
+* Fixed safetensors saving of fp8 format in model patcher
+* Corrected LTXV mid-video multi-frame guide alignment for better video processing
+* Improved VOID model compatibility by fixing RuntimeError issues
+* Enhanced Save3D node to support vertex colors and textures for comprehensive 3D workflow support
+
+**UI and Documentation Updates**
+* Added "Create Video" to the essentials tab for easier access to video generation workflows
+* Updated embedded documentation to v0.5.0 and workflow templates to v0.9.75
+* Fixed input directory creation in LoadAudio define_schema for better audio workflow handling
+
+**Bug Fixes**
+* Reverted breaking changes to maintain workflow compatibility
+* Fixed Quiver nodes functionality
+* Corrected deprecated cloud-runtime endpoints marking in API specification
+
+</Update>
+
+<Update label="v0.21.0" description="May 11, 2026">
+
+**Open-source model support**
+* [**BiRefNet**](https://github.com/Comfy-Org/ComfyUI/pull/12747): Background removal
+* [**Gemma4**](https://github.com/Comfy-Org/ComfyUI/pull/13376): LLM text generation
+* [**Void**](https://github.com/Comfy-Org/ComfyUI/pull/13403): Pass 1 & 2 model support
+
+**Partner model support**
+* [**Tripo 3.1**](https://github.com/Comfy-Org/ComfyUI/pull/13788): Production-ready 3D generation
+* **NanoBanana2** node v2: DynamicCombo and Autogrow UX refresh
+
+**Node improvements**
+* **Math Expression:** Added Boolean support for more complex calculations
+* **Flux2 latent previews:** High-quality preview support for Flux2 model latents
+* **Core Blueprints:** New built-in blueprint templates
+
+**Bug fixes & stability**
+* Corrected VAEDecodeAudio compatibility with LTX-2.x generated audio latents
+* Resolved SolidMask and MaskComposite device mismatch issues
+* Other fixes and improvements (CPU TE meta-tensor handling, port-in-use errors, save filenames without trailing underscores, and more)
+
+**Additional model & media updates**
+* **CogVideoX:** Video generation models
+* **Wan-Dancer:** Model support
+* **Major media loading:** Unified audio/video in the video loader; PyAV replaces Pillow for images (JPEG memory, tRNS PNG); automatic metadata rotation
+
+**Performance & video generation**
+* Block prefetch and LoRA async loading (notably faster for LTX)
+* Dynamic VRAM tuning with `--cache-ram 2` and lower peak usage for video
+* Triton ComfyUI kitchen CLI option for advanced optimizations
+* Auto-regressive video; image-to-video for causal models; `causal_window_fix` for long sequences; improved frame interpolation memory and overhead
+
+**More partner integrations**
+* **Luma UNI-1**, **Topaz Astra 2**, **GPT-5.5** / **GPT-5.5-pro**
+
+**Further node fixes**
+* Image blending fixes with `--gpu-only`; batch broadcasting on JoinImageWithAlpha; tile stride fixes on SplitImageToTileList and ImageMergeTileList
+
+**Developer & API**
+* `--feature-flag` registry; OpenAPI cloud-runtime and experiment endpoints; blueprint subgraph descriptions; deterministic ControlNet load order
+
+</Update>
+
 <Update label="v0.20.1" description="April 27, 2026">
 
 **New Model Support**
@@ -13,8 +86,8 @@
 * **LTXV Audio VAE:** Support for standalone LTXV audio VAE models with improved native integration
 
 **Partner Node Enhancements**
 * **Video Generation:** Added 4K resolution support for ByteDance 2, Veo models, and Kling nodes
 * **Veo 3 Lite:** New lightweight version of Veo model for faster video generation
 * **SD2 Real Human:** Added support for Stable Diffusion 2 real human generation
 * **GPT-Image 2:** New version option with additional resolution support and fixed price badges
 * **HappyHorse:** New model integration for enhanced image generation capabilities
@@ -31,10 +104,10 @@
 * Auto-downscaling functionality for ByteDance 2 nodes to optimize processing
 
 **Technical Updates**
 * Fixed SQLAlchemy version pinning (>=2.0) for improved database compatibility
 * Enhanced logging capabilities in ComfyUI app files for better debugging
 * Added OpenAPI 3.1 specification for comprehensive API documentation
 * Improved shader handling for blur/sharpen operations using textureSize
 
 **UI & Frontend**
 * Updated ComfyUI frontend to version 1.42.15 with various improvements
@@ -46,10 +119,10 @@
 <Update label="v0.20.0" description="April 27, 2026">
 
 **New Model Support**
 * **SUPIR model support:** Added enhanced super-resolution capabilities for high-quality upscaling workflows
 * **RIFE and FILM frame interpolation:** Introduced advanced frame interpolation models for smooth video transitions
 * **SAM 3.1 support:** Updated Segment Anything Model with improved segmentation accuracy
 * **LTX audio VAE improvements:** Enhanced native audio processing with support for standalone LTXV audio VAEs
 
 **Performance Optimizations**
 * Optimized Ernie inference for faster generation times
@@ -62,20 +135,20 @@
 * Auto-downscaling capabilities for ByteDance 2 nodes to handle large video files
 
 **Partner Node Updates**
 * **Veo models:** Added 4K resolution support and Veo 3 Lite model
 * **Kling nodes:** Enhanced with 4K resolution capabilities
 * **GPTImage:** Added gpt-image-2 version option and new resolution support
 * **HappyHorse model:** New partner model integration
 * **SD2 real human support:** Enhanced realistic human generation capabilities
 
 **UI and API Improvements**
 * Added OpenAPI 3.1 specification for better API documentation
 * Fixed texture size handling in blur/sharpen shaders for more accurate effects
 * Enhanced Preview Any node to display more tensor values for debugging
 
 **Bug Fixes**
 * Fixed SQLAlchemy compatibility by pinning version ≥2.0
 * Resolved Parameter assignment for Stable_Zero123 cc_projection weights
 * Corrected LTXV Reference Audio node functionality
 * Fixed 4K resolution restriction for veo-3.0 models
 
@@ -92,7 +165,6 @@
 * Corrected StabilityAI price badges for accurate API node pricing information
 
 </Update>
-
 <Update label="v0.19.2" description="April 17, 2026">
 
 **Performance & memory**
@@ -114,7 +186,7 @@
 
 **New Model Support**
 * [Erine Image  Text to Image](https://blog.comfy.org/p/ernie-image-day-0-support)
 * LLM: Ministral model
 
 **Partner Node**
 * [Sonilo: Create music for your video](https://blog.comfy.org/p/comfyui-now-supports-sonilo-via-partner)
@@ -129,7 +201,7 @@
 - Added support for LTX2 reference audio (ID-LoRA) functionality
 - Implemented Qwen3.5 text generation models with enhanced 8B model compatibility for TextGenerate node
 - Added RT-DETRv4 detection model support for improved object detection capabilities
 - Introduced Ace Step 1.5 XL model support with intermediate dtype compatibility
 - Added support for small flux.2 decoder and enabled Flux conditioning without pooled output
 - Implemented Ernie Image model for expanded generation options
 
@@ -138,8 +210,8 @@
 - Introduced Number Convert node with improved precision handling for large numbers
 - Added Image Histogram node for image analysis capabilities
 - Implemented Color Adjustment node with zero default values
 - Enhanced GLSL shader node with curve inputs and raised uniform limits
 - Added has_intermediate_output flag for nodes with interactive UI elements
 
 **Memory & Performance**
 - Integrated RAM cache with model RAM management for improved efficiency
@@ -155,7 +227,7 @@
 - Updated Tencent3D nodes with improved features
 
 **Bug Fixes & Stability**
 - Fixed Train LoRA crash with training_dtype "none" and bfloat16 LoRA weights
 - Resolved text generation regression with LTXAV model
 - Fixed FP8 scaled checkpoints compatibility issues
 - Improved browser security handling for insecure browsers
@@ -192,35 +264,35 @@
 
 **Memory & Performance Optimizations**
 - Added `--fp16-intermediates` flag to use FP16 for intermediate values between operations, reducing VRAM usage
 - Major VRAM reductions for LTX and WAN VAE models with inplace output processing and chunked encoder implementation
 - Improved tiled decode peak memory usage and fixed VRAM leaks in video VAE tiler fallback
 - Enhanced RAM pressure release strategies for Windows with comfy-aimdo updates (0.2.11-0.2.12)
 - Added `--enable-dynamic-vram` option to force enable dynamic VRAM management
 
 **Model & Hardware Support**
 - Added support for mxfp8 precision format for enhanced model compatibility
 - Enabled PyTorch Attention for AMD gfx1150 (Strix Point) GPUs
 - Enhanced training capabilities with FP4, 8, 16 native dtype support and quantized linear autograd functions
 - Fixed text encoder compatibility issues and disabled SageAttention for Hunyuan3D v2.1 DiT models
 
 **Node & API Enhancements**
 - Fixed Tencent TextToModel and ImageToModel API nodes
 - Added Nano Banana 2 "thought_image" support and Quiver SVG nodes
 - Marked seedream-3-0-t2i and seedance-1-0-lite models as deprecated
 - Added slice_cond and per-model context window conditioning resizing capabilities
 - Enhanced EmptyLatentImage and EmptyImage nodes to follow intermediate dtype settings
 
 **Frontend & Infrastructure**
 - Updated frontend package to version 1.41.21 with improved caching headers to prevent stale chunks
 - Added CacheProvider API for external distributed caching support
 - Implemented atomic writes for userdata to prevent data loss on crashes
 - Enhanced essentials category support for nodes and blueprints
 - Updated workflow templates to version 0.9.26
 
 **Bug Fixes & Stability**
 - Fixed pixel space VAE compatibility issues and model finalizer execution at exit
 - Improved weight dtype handling in Load Diffusion Model node (marked as advanced input)
 - Enhanced deferred weight initialization for subclassed models
 - Fixed various memory leaks and corruption issues in rare edge cases
 
 </Update>
@@ -245,13 +317,13 @@
 **Model Support & Enhancements**
 - Added support for Flux 2 Klein KV cache model through the FluxKVCache node
 - Enhanced model patching system with cleanup functions and pre-attention patches for Flux models
 - Added pre-attention and post-input patches to Qwen image model
 - Fixed text encoder LoRA loading for wrapped models
 - Improved batch processing with fixes for batch_size > 1 on various models
 
 **New Nodes & Features**
 - Added Painter node for enhanced image editing capabilities
 - Introduced Reve Image API nodes for expanded functionality
 - Updated workflow templates to version 0.9.21
 - Enhanced gradient stops formatting for better frontend compatibility
 
@@ -269,7 +341,7 @@
 <Update label="v0.16.4" description="March 7, 2026">
 
 **New Nodes & Features**
 - Added Math Expression node with simpleeval evaluation for mathematical operations in workflows
 - Introduced TencentSmartTopology API node for enhanced topology processing capabilities
 - Added Gemini 3.1 Flash-Lite model support to LLM node, expanding AI language model options
 
@@ -287,9 +359,9 @@
 <Update label="v0.16.3" description="March 5, 2026">
 
 **LTX Audio Model Improvements**
 - Fixed LTX2 vocoder conv_transpose1d with manual casting for improved audio processing stability
 - Resolved LTX audio VAE novram compatibility issues for better memory management during audio generation
 - Added causal_fix parameter to add_keyframe_index and append_keyframe functions, enhancing keyframe handling precision in video workflows
 
 </Update>
 <Update label="v0.16.2" description="March 5, 2026">
@@ -308,7 +380,7 @@
 
 **API Nodes Updates**
 - Updated xAI models and pricing for improved cost management and model selection
 - Enabled Kling 3.0 Motion Control for enhanced video generation capabilities
 
 **Workflow Templates**
 - Updated workflow templates to version 0.9.8 with latest improvements and optimizations
@@ -324,7 +396,7 @@
 - Native LongCat-Image implementation for enhanced image processing capabilities
 
 **Model & Format Support**
 - Added ACE-Step 1.5 lycoris key alias mapping for LoKR compatibility
 - Support for SDPose-OOD models
 - Added SCAIL WanVideo model support
 - Support for zeta chroma weights loading
@@ -333,21 +405,21 @@
 
 **Memory & Performance**
 - Dynamic VRAM mode is now the default, improving memory efficiency
 - Implemented LoRA requantization for non-QuantizedTensor fp8
 - Enhanced dynamic offload heuristics for better performance
 - Updated comfy-aimdo to version 0.2.6 with improved allocator handling
 - Fixed dynamic VRAM conflicts with weight hooks and WSL compatibility
 
 **API & UI Improvements**
 - Added "IMAGE+TEXT" support in NanoBanana2 API node
 - Renamed Mahiro CFG to Similarity-Adaptive Guidance for clarity
 - Fixed GLSL node input dimension requirements
 - Enhanced VAEDecodeAudioTiled to properly use tile_size input
 
 **Bug Fixes**
 - Fixed WanVAE encoder/decoder confusion in feat_map processing
 - Resolved ACE-1.5 memory estimation issues with missing lm_metadata
 - Fixed context window step handling for substep sigmas
 - Prevented VideoFromComponents crashes when writing to BytesIO
 - Limited image tile overlap to prevent processing issues
 
@@ -359,20 +431,20 @@
 <Update label="v0.15.1" description="February 26, 2026">
 
 **Bug Fixes & Stability**
 - Fixed KeyError when prompt entries lack class_type key, preventing workflow crashes
 - Resolved GLSL node compatibility issues with pyopengl accelerate numpy textures
 - Fixed LTXAV text encoder minimum length and memory estimation problems
 - Disabled dynamic VRAM when using torch compiler to prevent conflicts
 
 **Model & API Updates**
 - Updated comfy aimdo to version 0.2.2 with improved fallback handling
 - Added WanVideo-based segmentation model FlowRVS support for advanced video processing
 - Added NanoBanana2 API node for expanded functionality
 - Fixed Aimdo fallback to prevent zero-copy sft issues
 
 **UI & Frontend**
 - Patched frontend to version 1.39.19, fixing progress text serialization in API nodes
 - Moved essentials_category to correct replacement nodes for better organization
 
 **New Features**
 - Added per-guide attention strength control in self-attention for fine-tuned workflow control
@@ -382,10 +454,10 @@
 <Update label="v0.15.0" description="February 24, 2026">
 
 **New Nodes & Features**
 - Added BBox widget for precise bounding box selection in image workflows
 - Introduced 3-band equalizer node for audio processing workflows
 - Added basic text generation support with native models, initially supporting Gemma3 and Qwen 3
 - New GLSL shader node using PyOpenGL for advanced visual effects
 - Added ElevenLabs API nodes for text-to-speech functionality
 - New gradient-slider display mode for FLOAT inputs with improved UI
 - Added SplitImageToTileList and ImageMergeTileList nodes for tiled processing workflows
@@ -398,19 +470,19 @@
 - Forced Gemini to return uncompressed images for better quality
 
 **Performance & Bug Fixes**
 - Fixed FP8 dynamic VRAM workflow performance issues by limiting return of requants
 - Resolved non-contiguous audio waveform crashes in video save operations
 - Fixed LTXAV model loading issues including fp8 checkpoint support and embedding connector dtype problems
 - Improved PyOpenGL compatibility for versions < 3.1.4
 
 **UI Improvements**
 - Marked 429 widgets as advanced for collapsible UI
 - Added essentials_category for better node organization
 - Enhanced essential subgraph blueprints with improved categorization
 - Updated frontend to version 1.39.16
 
 **Model Support**
 - Enhanced text generation with tokenization improvements and minimum length enforcement
 - Updated LTXAV av embedding connectors integration
 - Temporary compatibility fixes for LTX 2.0 workflows during migration period
 
@@ -428,7 +500,7 @@
 
 **New Model Support**
 - Added "viduq3-turbo" model support in API nodes
 - Added Recraft V4 nodes for enhanced creative workflows
 
 **Updates**
 - Updated workflow templates to v0.8.43
@@ -449,7 +521,7 @@
 
 **Bug Fixes**
 - Fixed stable release workflow to properly pull latest comfy kitchen updates
 - Removed confusing load statistics from model patcher
 - Updated workflow templates to v0.7.69
 
 </Update>
@@ -461,21 +533,21 @@
 - Added Sage Attention 3 support via CLI argument
 
 **API Improvements**
 - V3 API: Exposed DynamicCombo and Autogrow functionality
 - API nodes: Kling Omni 720p resolution support, WAN2.6 ReferenceToVideo
 - Tripo3D: Optimized face_limit parameter handling
 - Fixed URL encoding preservation for Vidu API signed URLs
 
 **Fixes**
 - Fixed upscale model CPU offloading
 - Fixed LTXV2 text encoder lowvram compatibility
 - Fixed fp8 processing issues
 - Fixed MPO format images to use first frame only
 - Conditional comfy-kitchen CUDA support based on PyTorch version
 
 **Other**
 - Added OOM memory summary for debugging
 - Mahiro CFG display name update
 - Workflow templates updated to v0.7.67
 - Code refactoring and cleanup (CLIP preprocessing, comfy-kitchen integration)
 </Update>
@@ -484,7 +556,7 @@
 - PyTorch 2.4+ minimum version requirement
 
 **Fixes & Improvements**
 - Enabled async memory offload by default for AMD GPUs
 - Fixed ancestral sampler noise on CPU inference
 - Improved memory management error handling (pin/unpin operations)
 - Added VACE context handling in context windows
@@ -496,13 +568,13 @@
 
 **New Nodes**
 - Added ManualSigmas node for sampling control
 - Added Kling Motion Control node
 - ResizeByLongerSide now supports video processing
 
 **API Nodes**
 - Gemini API: Force enable prompt enhancement
 - Billing changed from dollars to credits
 - Deprecated "seededit", updated Seedream node display names
 
 **Technical**
 - Converted image processing nodes to V3 schema
@@ -511,7 +583,7 @@
 <Update label="v0.6.0" description="December 24, 2025">
 
 **New Model Support**
 - Qwen Image Layered
 - NewBie Image Exp0.1
 
 **New Features**
@@ -521,8 +593,8 @@
 - Disabled watermark generation by default in API nodes
 - Refactored trainer with resolution bucketing for better training efficiency
 - Enhanced AMD GPU support
 - Extended Kling O1 StartEndFrame node duration range
 - Updated Topaz 4k video upscaling functionality
 
 **Bug Fixes**
 - Fixed mask concatenation issues in ZImageFunControlNet in --gpu-only mode
@@ -540,11 +612,11 @@
 
 **New Models & Features**
 - WanMove model for advanced motion processing
 - Qwen Edit 2511 with automatic detection
 - Z-Image Fun Control Union 2.0 with inpainting
 
 **API Nodes**
 - Kling Omni Image, TextToVideoWithAudio, ImageToVideoWithAudio
 - Wan2.6 model integration
 - Tripo3.0 for 3D model generation
 
@@ -557,21 +629,21 @@
 <Update label="v0.4.0" description="December 10, 2025">
 
 **Model Support**
 - Ovis Image
 - Kandinsky 5.0 (T2V/I2V/T2I variants)
 - Z-Image Alibaba PAI-Fun ControlNet support
 - Improved Z-Image FP16 compatibility 
 - Added PATCH method to CORS headers and chroma-radiance-x0 mode
 
 **Workflow &  Enhancements**
 - Context windows fixes and enhanced temporal processing
 - Kling API @image reference format
 - ComfyUI-Manager pip installation support
 
 **Frontend Improvements**
 - Improve frontend load/startup time significantly
 - Restored settings button in sidebar
 - GPU accelerated maskeditor rendering
 - Improve asset cards and queue progress panel designs
 - Add missing nodes warning UI to queue button and breadcrumb
 - Mobile fixes
@@ -580,18 +652,18 @@
 **Major VRAM Optimizations** 
 - Temporal rolling VAE delivering massive VRAM reductions for video models (Hunyuan/Kandinsky)
 - Reduced LoRA memory reservations, especially beneficial for Flux2 workflows
 - Fixed Flux2 OOM errors with improved dequantization offload accounting
 - Enhanced LoRA computation speed with FP16 processing
 - Optimized text encoder GPU usage to prevent unnecessary CPU fallbacks
 
 **V3 Schema Migration**
 - Converted 3D, audio, freelunch, and mask nodes to V3 schema
 - Added MatchType, DynamicCombo, and Autogrow support for improved workflow flexibility
 
 **Bug Fixes & Compatibility**
 - Fixed text encoder regression when loaded directly on GPU
 - Resolved transformer_options clearing issue in Lumina models
 - Fixed Qwen Image LoRA training, HunyuanVideo 1.5 meanflow distil issues
 - Corrected EmptyAudio node input types for proper audio workflow integration
 - Enhanced VRAM calculation accuracy to prevent unwanted full offloading
 - Support for "transformer." LoRA prefix in Z-Image workflows
@@ -607,7 +679,7 @@
 **Frontend UI/UX**
 - New UI: **Nodes 2.0** public beta
 - Linear mode beta version (You can enable it by setting a hot-key in keybinding.)
 - Subgraphs improvement and bug fixes
 - Improve missing node UX
 - New workflow progress panel
 - New Assets sidebar
@@ -621,14 +693,14 @@
 - Support Z-Image LoRA training
 
 **Performance & model support**
 - Support video tiny VAEs
 - Support z-image LoRA formats
 - Enable async offloading by default on Nvidia
 
 **Partner Nodes**
 - Veo3 First-Last-Frame Node
 - Add Kling v2.5 turbo to Kling FirstLastFrame node
 - Kling O1 model support
 
 </Update>
 
@@ -684,8 +756,8 @@
 
 **API Node Expansions**
 - **New Topaz API Nodes**: Added support for Topaz video enhancement workflows directly within ComfyUI
 - **Nano Banana Pro Integration**: Expanded API node collection with Nano Banana Pro for enhanced processing capabilities
 - **Kling Lip Sync Improvements**: Fixed audio format conversion issues in KlingLipSyncAudioToVideoNode, ensuring proper MP3 format handling
 
 **Image Processing & Workflow Improvements**
 - **Enhanced Image Batching**: Fixed ImageBatch node to handle images with different channel counts and automatically add alpha channels when needed
@@ -710,11 +782,11 @@
 - Enhanced block swap functionality by removing potentially harmful native custom node implementations
 
 **API Node Enhancements**
 - **New Gemini Models Added**: Expanded AI model options for text and multimodal generation workflows
 - Improved API node development infrastructure with updated PR templates and Python 3.10 minimum version requirements
 
 **Development & Infrastructure**
 - Enhanced pylint configuration for better code quality in custom node development
 - Improved release automation and distribution processes for more reliable updates
 
 </Update>
@@ -722,19 +794,19 @@
 <Update label="v0.3.69" description="November 18, 2025">
 
 **Memory & Performance Optimizations**
 - **Pinned Memory Enabled by Default** for NVIDIA and AMD GPUs
 - **Reduced VRAM Usage** for Flux, Qwen, and LTX-Video models
 - Smart model unloading that automatically frees memory when VRAM usage increases
 - Improved weight casting performance on offload streams
 
 **New Features**
 - **ScaleROPE Node Now Works with Flux Models**
 - Added left padding support to tokenizers
 - Added `create_time` field to `/history` and `/queue` endpoints
 
 **Bug Fixes**
 - Fixed custom node import errors for SingleStreamBlock/DoubleStreamBlock (temporary fix)
 - Fixed Qwen ControlNet regression
 - Enhanced quantized operations with offload support and stability improvements
 - Unified RoPE function implementation across models
 
@@ -750,11 +822,11 @@
 
 **New Nodes & Execution Features**
 - **ScaleROPE Node**: Rope scaling support for WAN and Lumina models
 - Enhanced subgraph execution allowing multiple runs within single workflows
 - Improved caching system with proper handling of bytes data and None outputs
 
 **API Node Enhancements**
 - Migrated API nodes to V3 client architecture: Luma, Minimax, Pixverse, Ideogram, StabilityAI, Pika, Recraft, Hypernetwork, OpenAI
 - Extended LTXV API nodes with 12s-20s duration options
 - Fixed img2img operations in DALL-E 2 node
 - Enhanced Rodin3D nodes to return proper relative paths
@@ -768,18 +840,18 @@
 <Update label="v0.3.67" description="October 28, 2025">
 
 **API Nodes**
 - **LTXV API Integration**: Added new LTXV API nodes for Lightricks LTX video generation
 - Network Client V2 upgrade with async operations and cancellation support
 - Converted Tripo and Gemini API nodes to V3 schema
 
 **Performance & Compatibility**
 - Improved AMD GPU support by only disabling cudnn on newer AMD GPUs
 - Fixed Windows-specific network issues in API nodes for better retry handling
 
 **Core Improvements**
 - Enhanced dependency-aware caching system that fixes --cache-none behavior with loops
 - Added support for multi-dimensional latents
 - Added custom node published subgraphs endpoint
 
 **Updates**
 - Frontend bumped to version 1.28.8
@@ -789,26 +861,26 @@
 <Update label="v0.3.66" description="October 21, 2025">
 
 **Frontend Updates**
 - **Subgraph Widget Editing**: Edit subgraph parameters directly from new Parameters panel without entering the subgraph
 - **Template Modal Redesign**: New template browser with advanced filtering by model tags and categories
 
 **Performance Optimizations**
 - Iimproved workflow cancellation speed
 - Fixed VAE memory issue consuming 3x more on PyTorch 2.9 with NVIDIA GPUs  
 - Enhanced chroma radiance processing speed and fixed batch size issues above 1
 
 **API Nodes**
 - Added Veo 3.1 model support
 - Added TemporalScoreRescaling node for advanced temporal control in video workflows
 
 **Hardware & Compatibility**
 - Disabled FP8 operations for AMD gfx942 GPUs
 - Improved CUDA memory management in --fast autotune mode
 
 **Execution & Schema**
 - Converted ControlNet nodes to V3 schema
 - Enhanced EasyCache with proper batch_slice handling
 - Improved merge_nested_dicts functionality with proper input ordering
 - Added deprecation warnings for unused files
 </Update>
 <Update label="v0.3.65" description="October 14, 2025">
@@ -817,18 +889,18 @@
 - Migrated core node categories to V3 schema including model downscale, LoRA extraction, compositing, latent ops, SD3/SLG, Flux, upscale models, and HunyuanVideo nodes
 
 **Audio & Model Improvements**
 - Added MMaudio 16K VAE support for high-fidelity audio workflows
 - Fixed mono audio incorrectly saving as stereo
 - Refactored model sampling sigmas code and fixed FP8 scaled LoRA issues
 - Fixed loading of older Stable Diffusion checkpoints on newer NumPy versions
 
 **AMD GPU Optimizations**
 - Better memory estimation for SD/Flux VAE operations
 - Enabled RDNA4 PyTorch attention on ROCm 7.0+
 
 **API Node Updates**
 - Added price extractor and improvements to Kling/Pika API nodes
 - Enhanced Gemini Image API with aspect_ratio support
 
 **Updates**
 - Template v0.1.95, node docs v0.3.0
@@ -852,25 +924,25 @@
 - **AMD Support**: Enabled TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL by default
 
 **API Node Updates**
 - **Kling 2.5 Turbo**: Added support for kling-2-5-turbo in both txt2video and img2video nodes, with proper mode configuration fixes
 - **API Node Fixes**: Improved Gemini node base64 handling and fixed indentation issues in Recraft API node functions
 
 **Node Schema Migration (V3)**
 - **Extensive V3 Conversion**: Migrated numerous node categories to V3 schema including audio encoder, GITS, differential diffusion, optimal steps, PAG, LT, IP2P, morphology, torch compile, EPS, Pixverse, TomeSD, edit model, Rodin, and Stable3D nodes for improved workflow compatibility
 
 **Developer Experience Improvements**
 - **Code Quality**: Added pylint support for comfy_api_nodes folder and updated example_node.py to use V3 schema, making custom node development more consistent
 - **Documentation Updates**: Enhanced AMD installation instructions with nightly PyTorch commands for Windows users
 
 **Frontend Updates**
 - **Subgraph Publish**: Allows publishing subgraphs to the node library
 - **Node Selection Toolbox Redesign**: Redesigned the node selection toolbox
 </Update>
 <Update label="v0.3.62" description="September 30, 2025">
 
 **API Node**
 - **Rodin3D-Gen2 Parameter Fix**
 - **Seedance Pro Model Support**
 </Update>
 
 <Update label="v0.3.61" description="September 30, 2025">
@@ -880,32 +952,32 @@
 - **WAN Image-to-Image**: Wan2.5 Image-to-Image API node, support image editing.
 
 **Enhanced Audio Capabilities**
 - **New Audio Nodes**: Added new audio nodes for enhanced audio-driven workflows and multimodal content creation
 
 **Model Compatibility Fixes**
 - **Qwen2.5VL Template Handling**: Improved template management for Qwen2.5VL models when templates are already present in prompts
 - **HuMo View Operation**: Fixed .view() operation issues in HuMo models for more stable video generation
 
 **Memory and Performance Optimizations**
 - **Memory Leak Fix**: Resolved memory leaks by explicitly detaching model finalizers, improving long-running workflow stability
 - **Sampler CFG Enhancement**: Added 'input_cond' and 'input_uncond' parameters to sampler CFG function arguments for more flexible conditioning control
 </Update>
 
 <Update label="v0.3.60" description="September 23, 2025">
 
 **New Model Support**
 - **Wan2.2 Animate**: Support for Wan2.2 Animate video generation model with character replacement and motion transfer capabilities
 - **Qwen Image Edit 2509 Support**: Support for the updated Qwen Image Edit 2509 with multi-image editing, higher consistency, and native ControlNet support
 - **HuMo Models**: Added support for both 1.7B and 17B HuMo models that use audio to drive video generation while maintaining lip sync
 - **Chroma Radiance**: A model that performs image generation in pixel space, reducing losses during the image generation process
 - **Omnigen2 UMO LoRA**: Added support for Omnigen2 UMO LoRA models
 
 **API Node Additions**
 - **Kling v2.1 Support**: Added kling-v2-1 model to KlingStartEndFrame node
 - **Seedream4 Fixes**: Fixed the flag that ignores errors on partial success, making workflows more robust
 
 **Node Schema Migration (V3)**
 - **Core Node Updates**: Migrated multiple node categories to V3 schema including Minimax API, Cosmos, conditioning, CFG, and Canny nodes
 
 **Performance & Technical Improvements**
 - **FP8 Operations**: Enabled FP8 operations by default on gfx1200 hardware for faster processing
@@ -917,19 +989,19 @@
 </Update>
 <Update label="v0.3.59" description="September 10, 2025">
 
 **ByteDance Seedream 4.0 Integration**
 - **New Seedream Node**: Added ByteDanceSeedream (4.0) node.
 
 </Update>
 <Update label="v0.3.58" description="September 6, 2025">
 
 **New Model Support**
 - Hunyuan Image 2.1 regular model
 - Hunyuan 3D 2.1
 
 **New API Nodes**
 - Stable Audio 2.5 API
 - Seedance Video API
 </Update>
 <Update label="v0.3.57" description="September 4, 2025">
 
@@ -940,7 +1012,7 @@
 **Workflow Utilities**
 - **ImageScaleToMaxDimension Node**: New utility for intelligent image scaling
 - **SEEDS Noise System**: Updated noise decomposition with improved algorithms
 - **Enhanced Prompt Control**: Interrupt handler now accepts prompt_id parameters
 
 **Performance & Architecture**
 - **V3 Schema Migration**: Converted some core nodes to V3 schema
@@ -968,11 +1040,11 @@
 - **Automatic Audio Trimming**: Video saves now automatically trim audio to match video length, eliminating audio-video sync issues in final output files.
 
 **Advanced Latent Processing**
 - **LatentCut Node**: New node for cutting latents at precise points, enabling more granular control over latent space manipulation in complex generation workflows. This is particularly useful for batch processing and temporal video workflows, such as removing specific frames from videos.
 
 **Wan2.2 5B Model Integration**
 - **Fun Control Model Support**: Added support for Wan2.2 5B fun control model.
 - **Fun Inpaint Model Support**: Integrated Wan2.2 5B fun inpaint model.
 
 </Update>
 <Update label="v0.3.54" description="August 28, 2025">
@@ -995,7 +1067,7 @@
 This release adds ComfyUI audio processing capabilities and includes performance improvements and model compatibility updates:
 
 **Audio Processing Updates**
 - **Wav2vec2 Audio Encoder**: Added native wav2vec2 implementation as an audio encoder model, enabling audio-to-embedding workflows for multimodal applications
 - **Audio Encoders Directory**: Added models/audio_encoders directory, which is the audio encoder directory for Wan2.2 S2V
 - **AudioEncoderOutput V3 Support**: Made AudioEncoderOutput compatible with V3 node schema, ensuring seamless integration with modern workflow architectures
 
@@ -1009,46 +1081,46 @@
 - **Negative Audio Handling**: Fixed negative audio input handling in S2V workflows to use proper zero values
 
 **Sampling & Node Enhancements**
 - **DPM++ 2M SDE Heun (RES) Sampler**: New advanced sampler by @Balladie provides additional sampling options for fine-tuned generation control
 - **LatentConcat Node**: New node for concatenating latent tensors, enabling advanced latent space manipulation workflows
 - **EasyCache/LazyCache Stability**: Fixed critical crashes when tensor properties (shape/dtype/device) change during sampling, ensuring workflow reliability
 
 **Model Compatibility Improvements**
 - **ControlNet Type Models**: Enhanced compatibility fixes for ControlNet-type models working with Qwen Edit and Kontext workflows
 - **Flux Memory Optimization**: Adjusted Flux model memory usage factors for better resource utilization
 
 **Infrastructure & Reliability**
 - **Template Updates**: Updated to versions 0.1.66 and 0.1.68
 - **Documentation Cleanup**: Removed incompletely implemented models from readme to avoid user confusion
 
 
 </Update>
 <Update label="v0.3.52" description="August 23, 2025">
 
 **Enhanced Model Support & Qwen Image ControlNet Integration**
 
 This release significantly expands ControlNet capabilities and improves model compatibility, making ComfyUI workflows more versatile and reliable:
 
 **Qwen ControlNet Ecosystem**
 - **Diffsynth ControlNet Support**: Added support for Qwen Diffsynth ControlNets with Canny and depth conditioning, enabling precise edge and depth-based image control
 - **InstantX Qwen ControlNet**: Integrated InstantX Qwen ControlNet for expanded creative control options
 - **Inpaint ControlNet/Model Patches**: Enhanced inpainting capabilities with dedicated Diffsynth inpaint ControlNet support
 
 **Node Architecture & API Evolution**
 - **V3 Architecture Migration**: String nodes, Google Veo API, and Ideogram API nodes upgraded to V3 architecture for better performance and consistency
 - **Enhanced API Nodes**: OpenAI Chat node renamed to "OpenAI ChatGPT" for clarity, Gemini Chat node now includes copy button functionality
 - **Improved Usability**: API nodes now provide better user experience with clearer labeling and enhanced interaction features
 
 **Workflow Reliability & Performance**
 - **LTXV Noise Mask Fix**: Resolved key frame noise mask dimension issues when real noise masks exist, ensuring stable video workflow execution
 - **3D Latent Conditioning Control**: Fixed conditioning masks on 3D latents, enabling proper depth-aware conditioning control in advanced workflows
 - **Invalid Filename Handling**: Improved workflow save functionality with proper handling of invalid filenames, preventing save failures
 - **EasyCache & LazyCache**: Implemented advanced caching systems for improved workflow execution performance
 
 **Platform & Development Improvements**
 - **Python 3.13 Support**: Full compatibility with Python 3.13, keeping ComfyUI current with latest Python developments
 - **Frontend Update**: Updated to v1.25.10 with improved navigation and user interface enhancements
 - **Elementwise Fusions**: Added performance optimizations through elementwise operations fusion
 - **Navigation Mode Rollback**: Rolled back navigation default to traditional legacy mode, avoiding user experience issues caused by default enabled standard navigation mode. Users can still enable standard navigation mode in settings
 
 </Update>
@@ -1059,7 +1131,7 @@
 - **Qwen-Image-Edit Model**: Native support for Qwen-Image-Edit
 - **FluxKontextMultiReferenceLatentMethod Node**: Multi-reference input node for Flux workflows
 - **WAN 2.2 Fun Camera Model Support**: Support for video generation through camera control
 - **Template Updates**: Upgraded to version 0.1.62, added Wan2.2 Fun Camera and Qwen Image Edit templates
 
 **Core Function Improvements**
 - **Context Windows Support**: Enhanced sampling code to support longer sequence generation tasks
@@ -1071,9 +1143,9 @@
 
 **API Node Support Updates**
 - **GPT-5 Series Models**: Support for the latest GPT-5 models
 - **Kling V2-1 and V2-1-Master**: Updated video generation model functionality
 - **Minimax Hailuo Video Node**: New video generation node
 - **Vidu Video Node**: Vidu API node support
 - **Google Model Updates**: Added new Google Gemini models
 - **OpenAI API Fix**: Fixed MIME type errors in OpenAI API node input images
 
@@ -1084,7 +1156,7 @@
 - **Memory Management**: Optimized installation size and memory efficiency
 
 **Frontend Changes**
 - **Subgraph Support**: Subgraph functionality support
 - **Shortcut Panel**: Added bottom shortcut panel
 - **UI Layout Modifications**: Modified terminal entry layout, added template, log panel and other entries
 - **Standard Canvas Mode**: Added standard canvas mode, can be switched in `Lite Graph` > `Canvas` > `Canvas Navigation Mode`
@@ -1098,31 +1170,31 @@
 
 **Model Integration & Performance Enhancements**
 
 This release expands ComfyUI's model ecosystem with enhanced Qwen support, async API capabilities, and stability improvements for complex workflows:
 
 **Qwen Model Ecosystem**
 - **Qwen Image Model Support**: Improved integration including proper LoRA loading and model merging capabilities for sophisticated vision workflows
 - **Qwen Model Merging Node**: New dedicated node for merging Qwen image models, allowing creators to combine different model strengths
 - **SimpleTuner Lycoris LoRA Support**: Extended compatibility with SimpleTuner-trained Lycoris LoRAs for Qwen-Image models
 
 **API & Performance Infrastructure**
 - **Async API Nodes**: Introduction of asynchronous API nodes, enabling non-blocking workflow execution for better performance
 - **Memory Handling**: Enhanced RepeatLatentBatch node now properly handles multi-dimensional latents, fixing workflow interruptions
 - **WAN 2.2 Fun Control Support**: Added support for WAN 2.2 fun control features, expanding creative control for video workflows
 
 **Hardware Optimization & Compatibility**
 - **AMD GPU Improvements**: Enhanced AMD Radeon support with improved FP16 accuracy handling and performance optimization
 - **RDNA3 Architecture Fixes**: Resolved issues with gfx1201 GPUs when using Flux models with PyTorch attention
 - **Updated PyTorch Support**: Bumped CUDA and ROCM PyTorch versions with testing on Python 3.13 and CUDA 12.9
 
 **Developer Experience Enhancements**
 - **Cleaner Logging**: Feature flags now only display in verbose mode, reducing console noise
 - **Audio Processing Safety**: Enhanced torchaudio import safety checks prevent crashes when audio dependencies are unavailable
 - **Kling API Improvements**: Fixed image type parameter handling in Kling Image API nodes
 
 **Workflow Benefits**
 - **Async Workflow Execution**: New async API capabilities enable more responsive workflows when integrating external services
 - **Model Flexibility**: Expanded Qwen support allows for more diverse vision-language workflows with improved LoRA compatibility
 - **Hardware Utilization**: AMD GPU optimizations and updated PyTorch support improve performance across hardware configurations
 - **Batch Processing**: Fixed RepeatLatentBatch ensures reliable operation with complex multi-dimensional data structures
 - **Video Control**: WAN 2.2 fun control features provide advanced creative control for video generation workflows
@@ -1139,8 +1211,8 @@
 - **Workflow Navigation**: Enhanced user experience with better organization of commonly accessed elements
 
 **Model Integration**
 - **Qwen Vision Model Support**: Initial support for Qwen image models with configuration options
 - **Image Processing**: Enhanced Qwen model integration allows for more versatile image analysis and generation workflows
 
 **Video Generation**
 - **Veo3 Video Generation**: Added Veo3 video generation node with integrated audio support
@@ -1157,7 +1229,7 @@
 
 **Workflow Benefits**
 - **Faster Iteration**: Recently used items API enables quicker workflow assembly and modification
 - **Enhanced Creativity**: Qwen vision models open new possibilities for image understanding and manipulation workflows
 - **Video Production**: Veo3 integration transforms ComfyUI into a comprehensive multimedia creation platform
 - **Reliability**: Memory optimizations and device management fixes ensure stable operation with complex workflows
 - **Performance**: Optimized VRAM usage allows for more ambitious projects on systems with limited resources
@@ -1224,26 +1296,26 @@
 - **Audio Integration**: Enhanced audio handling for multimedia generation workflows
 
 **Hardware Support**
 - **Iluvatar CoreX Support**: Added native support for Iluvatar CoreX accelerators
 - **Intel XPU Optimization**: XPU support improvements including async offload capabilities
 - **AMD ROCm Enhancements**: Enabled PyTorch attention by default for gfx1201 on Torch 2.8
 - **CUDA Memory Management**: Fixed CUDA malloc to only activate on CUDA-enabled PyTorch installations
 
 **Sampling Algorithm Improvements**
 - **Euler CFG++ Enhancement**: Separated denoised and noise estimation processes in Euler CFG++ sampler
 - **WAN Model Support**: Added support for WAN (Wavelet-based Attention Network) models
 
 **Training Features**
 - **Training Nodes**: Added algorithm support, gradient accumulation, and optional gradient checkpointing
 - **Training Flexibility**: Better memory management and performance optimization for custom model training
 
 **Node & Workflow Enhancements**
 - **Moonvalley V2V Node**: Added Moonvalley Marey V2V node with enhanced input validation
 - **Negative Prompt Updates**: Improved negative prompt handling for Moonvalley nodes
 - **History API Enhancement**: Added map_function parameter to get_history API
 
 **API & System Improvements**
 - **Frontend Version Tracking**: Added required_frontend_version parameter in /system_stats API response
 - **Device Information**: Enhanced XPU device name printing for better hardware identification
 - **Template Updates**: Multiple template updates (0.1.40, 0.1.41) ensuring compatibility
 
@@ -1262,7 +1334,7 @@
 
 **Sampling & Generation Features**
 - **SA-Solver Sampler**: New reconstructed SA-Solver sampling algorithm providing enhanced numerical stability
 - **Experimental CFGNorm Node**: Classifier-free guidance normalization for improved control over generation quality
 - **Nested Dual CFG Support**: Added nested style configuration to DualCFGGuider node
 - **SamplingPercentToSigma Node**: New utility node for precise sigma calculation from sampling percentages
 
@@ -1273,19 +1345,19 @@
 
 **Platform & Performance Improvements**
 - **Async Node Support**: Full support for asynchronous node functions with earlier execution optimization
 - **Chroma Flexibility**: Un-hardcoded patch_size parameter in Chroma
 - **LTXV VAE Decoder**: Switched to improved default padding mode for better image quality
 - **Safetensors Memory Management**: Added workaround for mmap issues
 
 **API & Integration Enhancements**
 - **Custom Prompt IDs**: API now allows specifying prompt IDs for better workflow tracking
 - **Kling API Optimization**: Increased polling timeout to prevent user timeouts
 - **History Token Cleanup**: Removed sensitive tokens from history items
 - **Python 3.9 Compatibility**: Fixed compatibility issues ensuring broader platform support
 
 **Bug Fixes & Stability**
 - **MaskComposite Fixes**: Resolved errors when destination masks have 2 dimensions
 - **Fresca Input/Output**: Corrected input and output handling for Fresca model workflows
 - **Reference Bug Fixes**: Resolved incorrect reference bugs in Gemini node implementations
 - **Line Ending Standardization**: Automated detection and removal of Windows line endings
 
@@ -1315,12 +1387,12 @@
 **Image Processing Improvements**
 - **Transform Nodes**: Added ImageRotate and ImageFlip nodes for enhanced image manipulation
 - **ImageColorToMask Fix**: Corrected mask value returns for more accurate color-based masking
 - **3D Model Support**: Upload 3D models to custom subfolders for better organization
 
 **Guidance & Conditioning Enhancements**
 - **PerpNeg Guider**: Updated with improved pre and post-CFG handling
 - **Latent Conditioning Fix**: Resolved issues with conditioning at index > 0 for multi-step workflows
 - **Denoising Steps**: Added denoising step support to several samplers
 
 **Platform Stability**
 - **PyTorch Compatibility**: Fixed contiguous memory issues with PyTorch nightly builds
@@ -1328,7 +1400,7 @@
 - **Audio Processing**: Removed deprecated torchaudio.save function dependencies
 
 **Model Integration**
 - **Moonvalley Nodes**: Added native support for Moonvalley model workflows
 - **Scheduler Reordering**: Simple scheduler now defaults first
 - **Template Updates**: Multiple template version updates (0.1.31-0.1.35)
 
@@ -1344,9 +1416,9 @@
 
 This release brings improvements to model compatibility and workflow stability:
 
 **Expanded Model Documentation**: Added support documentation for Flux Kontext and Omnigen 2 models
 **VAE Encoding Improvements**: Removed unnecessary random noise injection during VAE encoding
 **Memory Management Fix**: Resolved a memory estimation bug affecting Kontext model usage
 
 </Update>
 
@@ -1358,23 +1430,23 @@
 - **LoRA Training Integration**: New native LoRA training node using weight adapter scheme
 
 **Performance & Hardware Optimizations**
 - **AMD GPU Enhancements**: Enabled FP8 operations and PyTorch attention on AMD GPUs
 - **Apple Silicon Fixes**: Addressed FP16 attention issues on Apple devices
 - **Flux Model Stability**: Resolved black image generation issues with certain Flux models
 
 **Sampling Improvements**
 - **Rectified Flow Samplers**: Added SEEDS and multistep DPM++ SDE samplers with RF support
 - **ModelSamplingContinuousEDM**: New cosmos_rflow option for enhanced sampling control
 - **Memory Optimization**: Improved memory estimation for Cosmos models
 
 **Developer & Integration Features**
 - **SQLite Database Support**: Enhanced data management capabilities for custom nodes
 - **PyProject.toml Integration**: Automatic web folder registration from pyproject files
 - **Frontend Flexibility**: Support for semver suffixes and prerelease frontend versions
 - **Tokenizer Enhancements**: Configurable min_length settings with tokenizer_data
 
 **Quality of Life Improvements**
 - **Kontext Aspect Ratio Fix**: Resolved widget-only limitation
 - **SaveLora Consistency**: Standardized filename format across all save nodes
 - **Python Version Warnings**: Added alerts for outdated Python installations
 - **WebcamCapture Fixes**: Corrected IS_CHANGED signature
@@ -1394,7 +1466,7 @@
 
 **Model Compatibility**
 - **Tensor Handling**: Streamlined list processing makes multi-model workflows more reliable
 - **BFL API Optimization**: Refined support for Kontext models with cleaner node interfaces
 - **Performance Boost**: Fused multiply-add operations in chroma processing for faster generation
 
 **Developer Experience**
@@ -1404,12 +1476,12 @@
 
 **Frontend & UI Enhancements**
 - **Frontend Updated to v1.21.7**: Stability fixes and performance improvements
 - **Custom API Base Support**: Better subpath handling for custom deployment configurations
 - **Security Hardening**: XSS vulnerability fixes
 
 **Bug Fixes & Stability**
 - **Pillow Compatibility**: Updated deprecated API calls
 - **ROCm Support**: Improved version detection for AMD GPU users
 - **Template Updates**: Enhanced project templates for custom node development
 
 </Update>

ja/changelog/index.mdx

@@ -6,6 +6,79 @@
 translationFrom: changelog/index.mdx, zh/changelog/index.mdx
 ---
 
+<Update label="v0.21.1" description="2026年5月13日">
+
+**新しいパートナーノード**
+* **Flux2ImageNode / GrokImageEditNodeV2：** 画像生成・編集ワークフローを強化するノードを追加
+* **ByteDanceSeedreamNodeV2：** DynamicCombo と Autogrow によりパラメータを扱いやすくした新版
+* **OpenAI Image ノード：** DynamicCombo／Autogrow に対応した新しい統合
+* **Claude LLM ノード：** テキスト生成ワークフロー向けに Claude 言語モデルを追加
+
+**モデル対応の強化**
+* **Anima TE LoRA：** Kohya 形式の Anima TE LoRA に対応
+* **HiDream-O1-Image：** dtype 修正と非ダイナミック VRAM 構成でのメモリ使用改善を含め、HiDream-O1-Image をフル対応
+
+**技術面の改善**
+* モデルパッチャーで fp8 形式の Safetensors 保存が正しく行われるよう修正
+* LTXV の動画中盤でのマルチフレームガイドの整合を修正し、動画処理を改善
+* RuntimeError を解消し VOID モデルの互換性を向上
+* Save3D ノードで頂点カラーおよびテクスチャに対応し、3D ワークフローを拡充
+
+**UI・ドキュメント**
+* 「動画を作成」をエッセンシャルタブに追加し、動画生成ワークフローへのアクセスを改善
+* 埋め込みドキュメントを v0.5.0、ワークフローテンプレートを v0.9.75 に更新
+* LoadAudio の `define_schema` で入力ディレクトリ作成を修正し、音声ワークフローを改善
+
+**バグ修正**
+* ワークフロー互換のための破壊的変更をロールバック
+* Quiver ノードの動作を修正
+* OpenAPI で非推奨のクラウドランタイムエンドポイントの記述を是正
+
+</Update>
+
+<Update label="v0.21.0" description="2026年5月11日">
+
+**オープンソースモデル対応**
+* [**BiRefNet**](https://github.com/Comfy-Org/ComfyUI/pull/12747)：背景除去
+* [**Gemma4**](https://github.com/Comfy-Org/ComfyUI/pull/13376)：LLM テキスト生成
+* [**Void**](https://github.com/Comfy-Org/ComfyUI/pull/13403)：パス 1／2 モデル対応
+
+**パートナーモデル対応**
+* [**Tripo 3.1**](https://github.com/Comfy-Org/ComfyUI/pull/13788)：本番運用を想定した 3D 生成
+* **NanoBanana2** ノード v2：DynamicCombo と Autogrow の UX 更新
+
+**ノード改善**
+* **Math Expression：** Boolean 対応でより複雑な計算が可能に
+* **Flux2 潜在プレビュー：** Flux2 潜在の高品質プレビュー
+* **コア Blueprint：** 新しい組み込み Blueprint テンプレート
+
+**バグ修正・安定性**
+* LTX-2.x が出力した音声潜在との VAEDecodeAudio 互換性を修正
+* SolidMask／MaskComposite のデバイス不一致を解消
+* そのほかの修正（CPU TE メタテンソル処理、ポート使用中エラー、末尾アンダースコアなしの保存ファイル名など）
+
+**その他のモデル・メディア**
+* **CogVideoX：** 動画生成モデル対応
+* **Wan-Dancer：** モデル対応
+* **メディア読み込み：** 動画ローダーで音声／動画を統合読み込み；画像は Pillow から PyAV へ（JPEG メモリ、tRNS PNG）；メタデータに基づく向きの自動補正
+
+**パフォーマンス・動画生成**
+* ブロック先読みと LoRA 非同期読み込み（LTX などで効果大）
+* 動的 VRAM 調整（`--cache-ram 2`、動画処理のピーク VRAM 低減）
+* Triton ComfyUI kitchen 用 CLI オプション
+* 自己回帰動画、因果モデル向け I2V、`causal_window_fix`、フレーム補間のメモリ／オーバーヘッド改善
+
+**追加のパートナー統合**
+* **Luma UNI-1**、**Topaz Astra 2**、**GPT-5.5**／**GPT-5.5-pro**
+
+**ノードまわりの追加修正**
+* `--gpu-only` 時の画像ブレンド、JoinImageWithAlpha のバッチブロードキャスト、SplitImageToTileList／ImageMergeTileList のタイルストライド
+
+**開発者・API**
+* `--feature-flag` レジストリ、OpenAPI（クラウドランタイムと実験エンドポイント）、Blueprint サブグラフの説明、ControlNet の決定的な読み込み順序
+
+</Update>
+
 <Update label="v0.20.1" description="2026年4月27日">
 
 **新しいモデル対応**
@@ -15,8 +88,8 @@
 * **LTXV Audio VAE：** 単体の LTXV Audio VAE モデルに対応し、ネイティブ統合を改善
 
 **パートナーノードの強化**
 * **動画生成：** ByteDance 2 / Veo / Kling ノードに 4K 解像度対応を追加
 * **Veo 3 Lite：** より高速な動画生成向けの軽量版 Veo モデルを追加
 * **SD2 Real Human：** Stable Diffusion 2 のリアル人物生成に対応
 * **GPT-Image 2：** 新しいバージョンオプション、追加解像度対応、価格バッジ修正
 * **HappyHorse：** 新しいモデル統合で画像生成機能を強化
@@ -33,10 +106,10 @@
 * ByteDance 2 ノードに自動スケーリング機能を追加し処理を最適化
 
 **技術更新**
 * SQLAlchemy のバージョン固定（>= 2.0）を修正し、DB 互換性を改善
 * ComfyUI アプリファイルのログ機能を強化し、デバッグを改善
 * OpenAPI 3.1 仕様を追加し、API ドキュメントを拡充
 * textureSize を用いて、ぼかし／シャープ化のシェーダー処理を改善
 
 **UI / フロントエンド**
 * ComfyUI フロントエンドを 1.42.15 に更新（各種改善）
@@ -64,20 +137,20 @@
 * 大容量動画を扱うため、ByteDance 2 ノードに自動ダウンスケール機能を追加
 
 **パートナーノードの更新**
 * **Veo モデル：** 4K 解像度対応と Veo 3 Lite モデルを追加
 * **Kling ノード：** 4K 解像度対応を強化
 * **GPTImage：** `gpt-image-2` のバージョン選択と新しい解像度に対応
 * **HappyHorse モデル：** 新しいパートナーモデルを統合
 * **SD2 Real Human 対応：** リアルな人物生成能力を強化
 
 **UI / API の改善**
 * OpenAPI 3.1 仕様を追加し、API ドキュメントを改善
 * ぼかし／シャープ化シェーダーの textureSize 処理を修正し、効果をより正確に
 * Preview Any ノードを強化し、デバッグ用により多くのテンソル値を表示
 
 **バグ修正**
 * SQLAlchemy 互換性のため、バージョンを >= 2.0 に固定
 * Stable_Zero123 の cc_projection 重みのパラメータ割り当てを修正
 * LTXV Reference Audio ノードの動作を修正
 * veo-3.0 モデルの 4K 解像度制限を修正
 
@@ -132,7 +205,7 @@
 - LTX2 の参照音声（ID-LoRA）機能に対応
 - Qwen3.5 テキスト生成モデルを実装し、TextGenerate ノードの 8B モデル互換性を向上
 - 物体検出機能を強化する RT-DETRv4 検出モデルに対応
 - 中間 dtype 互換性を備えた Ace Step 1.5 XL モデルに対応
 - 小型 flux.2 デコーダに対応し、pooled output なしでの Flux conditioning を有効化
 - Ernie Image モデルを実装し、生成オプションを拡張
 
@@ -142,7 +215,7 @@
 - 画像解析のため、Image Histogram ノードを追加
 - デフォルト値を 0 とする Color Adjustment ノードを実装
 - 曲線入力に対応し uniform 上限を引き上げるなど、GLSL シェーダーノードを強化
 - インタラクティブ UI を持つノード向けに has_intermediate_output フラグを追加
 
 **メモリおよびパフォーマンス**
 - モデル RAM 管理と統合された RAM キャッシュを導入し、効率を向上
@@ -158,7 +231,7 @@
 - Tencent3D ノードを更新し、機能を改善
 
 **バグ修正および安定性**
 - training_dtype が "none" の場合と bfloat16 LoRA 重みで発生する Train LoRA のクラッシュを修正
 - LTXAV モデルにおけるテキスト生成のリグレッションを解消
 - FP8 scaled checkpoints の互換性問題を修正
 - 安全でないブラウザ向けのブラウザセキュリティ処理を改善
@@ -204,16 +277,16 @@
 
 **モデルおよびハードウェア対応**
 - mxfp8 精度フォーマットへの対応を追加し、モデル互換性を向上
 - AMD gfx1150 (Strix Point) GPU 向けに PyTorch Attention を有効化
 - FP4、FP8、FP16 のネイティブ dtype 対応および量子化線形自動微分関数により、学習機能を強化
 - テキストエンコーダの互換性問題を修正し、Hunyuan3D v2.1 DiT モデル向け SageAttention を無効化
 
 **ノードおよび API の強化**
 - Tencent TextToModel および ImageToModel API ノードを修正
 - Nano Banana 2 の `"thought_image"` 対応および Quiver SVG ノードを追加
 - `seedream-3-0-t2i` および `seedance-1-0-lite` モデルを非推奨としてマーク
 - `slice_cond` およびモデルごとのコンテキストウィンドウ条件付けサイズ調整機能を追加
 - `EmptyLatentImage` および `EmptyImage` ノードを強化し、中間 dtype 設定に従うように変更
 
 **フロントエンドおよび基盤構造**
 - フロントエンドパッケージをバージョン 1.41.21 へ更新し、古くなったチャンクを防止するためのキャッシュヘッダーを改善
@@ -224,7 +297,7 @@
 
 **バグ修正および安定性向上**
 - ピクセル空間 VAE の互換性問題および終了時のモデルファイナライザ実行を修正
 - `Load Diffusion Model` ノードにおける重み dtype 処理を改善（高度入力としてマーク）
 - サブクラス化されたモデル向けの遅延重み初期化を強化
 - 稀なエッジケースにおけるさまざまなメモリリークおよび破損問題を修正
 
@@ -251,7 +324,7 @@
 **モデル対応および強化**
 - `FluxKVCache` ノード経由で Flux 2 Klein KV キャッシュモデルをサポート
 - Flux モデル向けにクリーンアップ関数および事前アテンションパッチを備えたモデルパッチングシステムを強化
 - Qwen 画像モデル向けに事前アテンションおよび後入力パッチを追加
 - ラップされたモデルにおけるテキストエンコーダ LoRA の読み込みを修正
 - さまざまなモデルにおいて `batch_size > 1` のバッチ処理を修正し、改善
 
@@ -293,7 +366,7 @@
 <Update label="v0.16.3" description="2026年3月5日">
 
 **LTX 音声モデルの改善**
 - LTX2 vocoder `conv_transpose1d` の手動キャストを修正し、音声処理の安定性を向上
 - LTX 音声 VAE の `novram` 互換性問題を解決し、音声生成中のメモリ管理を改善
 - 動画ワークフローにおけるキーフレーム処理の精度を高めるため、`add_keyframe_index` および `append_keyframe` 関数に `causal_fix` パラメーターを追加
 
@@ -314,7 +387,7 @@
 
 **API ノードの更新**
 - コスト管理およびモデル選択の向上のため、xAI モデルおよび料金体系を更新
 - 動画生成能力を強化する Kling 3.0 Motion Control を有効化
 
 **ワークフローテンプレート**
 - 最新の改善および最適化を含むワークフローテンプレートをバージョン 0.9.8 へ更新
@@ -330,7 +403,7 @@
 - 強化された画像処理機能を提供する LongCat-Image のネイティブ実装
 
 **モデルおよびフォーマット対応**
 - LoKR 互換性のため、ACE-Step 1.5 lycoris キー別名マッピングを追加
 - SDPose-OOD モデルをサポート
 - SCAIL WanVideo モデルをサポート
 - zeta chroma 重みの読み込みをサポート
@@ -366,7 +439,7 @@
 
 **バグ修正および安定性向上**
 - `prompt` エントリに `class_type` キーが欠落している場合に発生する `KeyError` を修正し、ワークフローのクラッシュを防止
 - pyopengl accelerate numpy テクスチャとの GLSL ノード互換性問題を解決
 - LTXAV テキストエンコーダの最小長およびメモリ推定問題を修正
 - torch コンパイラー使用時に動的 VRAM を無効化し、衝突を防止
 
@@ -374,7 +447,7 @@
 - 後方互換性処理の向上のため、`comfy aimdo` をバージョン 0.2.2 へ更新
 - 高度な動画処理のため、WanVideo ベースのセグメンテーションモデル FlowRVS を追加
 - 機能拡張のため、NanoBanana2 API ノードを追加
 - ゼロコピー sft 問題を防止するための Aimdo 後方互換性処理を修正
 
 **UI およびフロントエンド**
 - API ノードにおける進行状況テキストのシリアル化を修正するため、フロントエンドをバージョン 1.39.19 へパッチ適用
@@ -390,7 +463,7 @@
 **新規ノードおよび機能**
 - 画像ワークフローにおける正確な境界ボックス選択のための `BBox` ウィジェットを追加
 - 音声処理ワークフロー向けの 3 バンドイコライザー・ノードを導入
 - Gemma3 および Qwen 3 を初期対応モデルとして、ネイティブモデルによる基本的なテキスト生成をサポート
 - 高度な視覚効果のための PyOpenGL を用いた GLSL シェーダーノードを新規追加
 - テキスト・ツー・スピーチ機能のための ElevenLabs API ノードを追加
 - 改善された UI を実現するため、FLOAT 入力向けの新しいグラデーションスライダーディスプレイモードを追加
@@ -406,7 +479,7 @@
 **パフォーマンスおよびバグ修正**
 - 再量子化の返却を制限することで、FP8 動的 VRAM ワークフローパフォーマンス問題を修正
 - 動画保存操作における非連続音声波形のクラッシュを解決
 - fp8 チェックポイント対応および埋め込みコネクタ dtype 問題を含む LTXAV モデル読み込み問題を修正
 - PyOpenGL バージョン < 3.1.4 への互換性を向上
 
 **UI 改善**
@@ -435,7 +508,7 @@
 
 **新規モデル対応**
 - API ノードに `"viduq3-turbo"` モデルを追加
 - 創造的なワークフローを強化する Recraft V4 ノードを追加
 
 **更新**
 - ワークフローテンプレートを v0.8.43 へ更新
@@ -468,21 +541,21 @@
 - CLI 引数経由で Sage Attention 3 をサポート
 
 **API 改善**
 - V3 API：DynamicCombo および Autogrow 機能を公開
 - API ノード：Kling Omni 720p 解像度対応、WAN2.6 ReferenceToVideo
 - Tripo3D：`face_limit` パラメーター処理を最適化
 - Vidu API 署名付き URL の URL エンコーディング保持を修正
 
 **修正**
 - アップスケールモデルの CPU オフロードを修正
 - LTXV2 テキストエンコーダの lowvram 対応を修正
 - fp8 処理の問題を修正
 - MPO 形式画像を最初のフレームのみ使用するよう修正
 - PyTorch バージョンに基づく `comfy-kitchen` CUDA 対応の条件付き有効化
 
 **その他**
 - デバッグ用に OOM メモリ概要を追加
 - Mahiro CFG の表示名を更新
 - ワークフローテンプレートを v0.7.67 へ更新
 - コードのリファクタリングおよびクリーンアップ（CLIP 前処理、`comfy-kitchen` 統合）
 
@@ -504,7 +577,7 @@
 
 **新規ノード**
 - サンプリング制御のための `ManualSigmas` ノードを追加
 - Kling Motion Control ノードを追加
 - `ResizeByLongerSide` を動画処理に対応
 
 **API ノード**
@@ -520,7 +593,7 @@
 <Update label="v0.6.0" description="2025年12月24日">
 
 **新規モデル対応**
 - Qwen Image Layered
 - NewBie Image Exp0.1
 
 **新規機能**
@@ -530,7 +603,7 @@
 - API ノードにおける透かし生成をデフォルトで無効化
 - 解像度バケット化を備えたトレーナーをリファクタリングし、トレーニング効率を向上
 - AMD GPU 対応を強化
 - Kling O1 StartEndFrame ノードの持続時間範囲を拡張
 - Topaz 4k 動画アップスケーリング機能を更新
 
 **バグ修正**
@@ -549,11 +622,11 @@
 
 **新規モデルおよび機能**
 - 高度なモーション処理のための WanMove モデル
 - 自動検出対応の Qwen Edit 2511
 - ペイントインを含む Z-Image Fun Control Union 2.0
 
 **API ノード**
 - Kling Omni Image、TextToVideoWithAudio、ImageToVideoWithAudio
 - Wan2.6 モデル統合
 - 3D モデル生成のための Tripo3.0
 
@@ -566,15 +639,15 @@
 <Update label="v0.4.0" description="2025年12月10日">
 
 **モデル対応**
 - Ovis Image
 - Kandinsky 5.0（T2V/I2V/T2I 変種）
 - Z-Image Alibaba PAI-Fun ControlNet 対応
 - Z-Image FP16 対応の改善
 - CORS ヘッダーへの PATCH メソッド追加および chroma-radiance-x0 モード
 
 **ワークフローおよび強化**
 - コンテキストウィンドウの修正および時系列処理の強化
 - Kling API `@image` 参照形式
 - ComfyUI-Manager の pip インストール対応
 
 **フロントエンド改善**
@@ -599,7 +672,7 @@
 **バグ修正および互換性**
 - GPU 直接読み込み時のテキストエンコーダリグレッションを修正
 - Lumina モデルにおける `transformer_options` クリアリング問題を解決
 - Qwen Image LoRA トレーニング、HunyuanVideo 1.5 meanflow distil 問題を修正
 - 音声ワークフロー統合のための `EmptyAudio` ノード入力タイプを修正
 - 不要な完全オフロードを防ぐための VRAM 計算精度の向上
 - Z-Image ワークフローにおける `"transformer."` LoRA プレフィックス対応
@@ -636,8 +709,8 @@
 
 **パートナー・ノード**
 - Veo3 First-Last-Frame ノード
 - Kling FirstLastFrame ノードに Kling v2.5 turbo を追加
 - Kling O1 モデル対応
 
 </Update>
 
@@ -693,8 +766,8 @@
 
 **API ノードの拡張**
 - **新規 Topaz API ノード**：ComfyUI 内で直接 Topaz 動画強化ワークフローを実行可能に
 - **Nano Banana Pro 統合**：処理能力を拡張するための Nano Banana Pro API ノード群を追加
 - **Kling リップシンクの改善**：`KlingLipSyncAudioToVideoNode` の音声フォーマット変換の問題を修正し、適切な MP3 フォーマット処理を確保
 
 **画像処理およびワークフローの改善**
 - **画像バッチ処理の強化**：異なるチャンネル数の画像を処理し、必要に応じてアルファチャンネルを自動追加するよう ImageBatch ノードを修正

zh/changelog/index.mdx

@@ -4,6 +4,79 @@
 icon: "clock-rotate-left"
 ---
 
+<Update label="v0.21.1" description="2026年5月13日">
+
+**新的合作伙伴节点**
+* **Flux2ImageNode 和 GrokImageEditNodeV2：** 新增节点，用于增强图像生成和编辑工作流
+* **ByteDanceSeedreamNodeV2：** 新版本，带有 DynamicCombo 和 Autogrow 功能，改善参数处理
+* **OpenAI Image 节点：** 新的集成，具备 DynamicCombo 和 Autogrow 能力
+* **Claude LLM 节点：** 为文本生成工作流添加了 Claude 语言模型支持
+
+**模型支持增强**
+* **Anima TE LoRA：** 新增对 Kohya 格式的 Anima TE LoRA 支持
+* **HiDream-O1-Image：** 完全支持 HiDream-O1-Image 模型，修复了数据类型问题并改善了非动态 VRAM 配置下的内存使用
+
+**技术改进**
+* 修复了模型补丁器中 fp8 格式的 safetensors 保存问题
+* 修正了 LTXV 视频中段多帧引导对齐，优化视频处理效果
+* 通过修复 RuntimeError 问题，提高了 VOID 模型兼容性
+* 增强了 Save3D 节点，支持顶点颜色和纹理，为全面的 3D 工作流提供支持
+
+**界面和文档更新**
+* 在基础功能选项卡中添加了"创建视频"，便于访问视频生成工作流
+* 更新嵌入式文档到 v0.5.0，工作流模板到 v0.9.75
+* 修复了 LoadAudio define_schema 中的输入目录创建，改善音频工作流处理
+
+**错误修复**
+* 撤销了破坏性更改以维持工作流兼容性
+* 修复了 Quiver 节点功能
+* 修正了 API 规范中已弃用的云运行时端点标记
+
+</Update>
+
+<Update label="v0.21.0" description="2026年5月11日">
+
+**开源模型支持**
+* [**BiRefNet**](https://github.com/Comfy-Org/ComfyUI/pull/12747)：背景移除
+* [**Gemma4**](https://github.com/Comfy-Org/ComfyUI/pull/13376)：LLM 文本生成
+* [**Void**](https://github.com/Comfy-Org/ComfyUI/pull/13403)：第 1、2 阶段模型支持
+
+**合作伙伴模型支持**
+* [**Tripo 3.1**](https://github.com/Comfy-Org/ComfyUI/pull/13788)：面向生产的 3D 生成能力
+* **NanoBanana2** 节点 v2：DynamicCombo 与 Autogrow 交互更新
+
+**节点改进**
+* **数学表达式：** 增加布尔运算，支持更复杂计算
+* **Flux2 潜在预览：** Flux2 模型潜在变量的高质量预览
+* **核心蓝图：** 新增内置蓝图模板
+
+**错误修复与稳定性**
+* 修正 VAEDecodeAudio 与 LTX-2.x 生成音频潜在变量的兼容性
+* 解决 SolidMask 与 MaskComposite 的设备不匹配问题
+* 其他修复与改进（CPU TE 元张量处理、端口占用报错、不带尾随下划线的保存文件名等）
+
+**其他模型与媒体**
+* **CogVideoX：** 视频生成模型支持
+* **Wan-Dancer：** 模型支持
+* **媒体加载：** 视频加载器统一读取音视频；以 PyAV 替代 Pillow（JPEG 内存、tRNS PNG）；自动方向元数据处理
+
+**性能与视频生成**
+* 块预取与 LoRA 异步加载（对 LTX 等更明显）
+* 动态 VRAM 调优：`--cache-ram 2`、降低视频处理峰值显存占用
+* Triton ComfyUI kitchen CLI，用于进阶优化选项
+* 自回归视频、因果模型的图生视频、`causal_window_fix`；帧插值内存与开销优化
+
+**更多合作伙伴集成**
+* **Luma UNI-1**、**Topaz Astra 2**、**GPT-5.5** / **GPT-5.5-pro**
+
+**其他节点修复**
+* `--gpu-only` 下的图像混合；JoinImageWithAlpha 批处理广播；SplitImageToTileList / ImageMergeTileList 瓦片步幅
+
+**开发者与 API**
+* `--feature-flag` 注册表；OpenAPI 云运行时与实验端点；蓝图子图描述；ControlNet 加载顺序确定性
+
+</Update>
+
 <Update label="v0.20.1" description="2026年4月27日">
 
 **新增模型支持**
@@ -14,7 +87,7 @@
 
 **合作伙伴节点增强**
 * **视频生成：** 为ByteDance 2、Veo模型和Kling节点新增4K分辨率支持
 * **Veo 3 Lite：** 新的轻量级Veo模型版本，用于更快的视频生成
 * **SD2 Real Human：** 新增Stable Diffusion 2真人生成支持
 * **GPT-Image 2：** 新版本选项，支持更多分辨率并修复价格标签
 * **HappyHorse：** 新的模型集成，增强图像生成能力
@@ -140,7 +213,7 @@
 - 新增 Image Histogram 节点，提供图像分析功能
 - 实现了 Color Adjustment 节点，使用零默认值
 - 增强了 GLSL 着色器节点，支持曲线输入并提高了统一变量限制
 - 为具有交互式 UI 元素的节点添加了 has_intermediate_output 标志
 
 **内存与性能**
 - 集成了 RAM 缓存与模型 RAM 管理，提升了效率
@@ -156,7 +229,7 @@
 - 更新了 Tencent3D 节点，改进了功能特性
 
 **错误修复与稳定性**
 - 修复了 Train LoRA 在 training_dtype 为 "none" 和 bfloat16 LoRA 权重时的崩溃问题
 - 解决了 LTXAV 模型的文本生成回归问题
 - 修复了 FP8 缩放检查点兼容性问题
 - 改进了不安全浏览器的浏览器安全处理
@@ -200,7 +273,7 @@
 
 **模型与硬件支持**
 - 添加了对mxfp8精度格式的支持，增强模型兼容性
 - 为AMD gfx1150 (Strix Point) GPU启用了PyTorch注意力机制
 - 通过FP4、8、16原生数据类型支持和量化线性自动求导函数增强训练能力
 - 修复了文本编码器兼容性问题，禁用了Hunyuan3D v2.1 DiT模型的SageAttention
 
@@ -248,13 +321,13 @@
 **模型支持与增强**
 - 通过 FluxKVCache 节点添加了对 Flux 2 Klein KV 缓存模型的支持
 - 增强了模型补丁系统，为 Flux 模型提供清理函数和预注意力补丁
 - 为 Qwen 图像模型添加了预注意力和后输入补丁
 - 修复了包装模型的文本编码器 LoRA 加载问题
 - 改进了批处理，修复了各种模型在 batch_size > 1 时的问题
 
 **新节点与功能**
 - 添加了 Painter 节点，增强图像编辑功能
 - 引入了 Reve Image API 节点，扩展功能范围
 - 将工作流模板更新至版本 0.9.21
 - 增强了渐变停靠点格式化，提升前端兼容性
 
@@ -291,9 +364,9 @@
 <Update label="v0.16.3" description="2026年3月5日">
 
 **LTX Audio 模型改进**
 - 修复了 LTX2 vocoder conv_transpose1d 的手动类型转换问题，提高了音频处理稳定性
 - 解决了 LTX audio VAE novram 兼容性问题，改善了音频生成过程中的内存管理
 - 为 add_keyframe_index 和 append_keyframe 函数添加了 causal_fix 参数，增强了视频工作流中关键帧处理的精度
 
 </Update>
 
@@ -313,7 +386,7 @@
 
 **API 节点更新**
 - 更新了 xAI 模型和定价，以改进成本管理和模型选择
 - 启用了 Kling 3.0 Motion Control，增强视频生成功能
 
 **工作流模板**
 - 工作流模板更新至 0.9.8 版本，包含最新改进和优化
@@ -329,7 +402,7 @@
 - 原生 LongCat-Image 实现，增强图像处理能力
 
 **模型和格式支持**
 - 为 LoKR 兼容性添加了 ACE-Step 1.5 lycoris 键别名映射
 - 支持 SDPose-OOD 模型
 - 添加了 SCAIL WanVideo 模型支持
 - 支持 zeta chroma 权重加载
@@ -345,13 +418,13 @@
 
 **API 和 UI 改进**
 - 在 NanoBanana2 API 节点中添加了"IMAGE+TEXT"支持
 - 将 Mahiro CFG 重命名为 Similarity-Adaptive Guidance 以提高清晰度
 - 修复了 GLSL 节点输入维度要求
 - 增强了 VAEDecodeAudioTiled，正确使用 tile_size 输入
 
 **错误修复**
 - 修复了 WanVAE 编码器/解码器在 feat_map 处理中的混淆问题
 - 解决了缺少 lm_metadata 时 ACE-1.5 内存估算问题
 - 修复了子步长 sigma 的上下文窗口步长处理
 - 防止 VideoFromComponents 写入 BytesIO 时崩溃
 - 限制图像瓦片重叠以防止处理问题
@@ -363,20 +436,20 @@
 <Update label="v0.15.1" description="2026年2月26日">
 
 **错误修复与稳定性**
 - 修复了当提示条目缺少 class_type 键时的 KeyError，防止工作流崩溃
 - 解决了 GLSL 节点与 pyopengl accelerate numpy 纹理的兼容性问题
 - 修复了 LTXAV 文本编码器最小长度和内存估算问题
 - 在使用 torch 编译器时禁用动态 VRAM 以防止冲突
 
 **模型与 API 更新**
 - 将 comfy aimdo 更新至 0.2.2 版本，改进了后备处理机制
 - 添加了基于 WanVideo 的分割模型 FlowRVS 支持，用于高级视频处理
 - 添加了 NanoBanana2 API 节点以扩展功能
 - 修复了 Aimdo 后备机制以防止零拷贝 sft 问题
 
 **UI 与前端**
 - 修补前端至 1.39.19 版本，修复了 API 节点中进度文本序列化问题
 - 将 essentials_category 移至正确的替换节点以更好地组织
 
 **新功能**
 - 在自注意力机制中添加了每个引导的注意力强度控制，实现精细的工作流控制
@@ -386,9 +459,9 @@
 <Update label="v0.15.0" description="2026年2月24日">
 
 **新节点和功能**
 - 添加了 BBox 小部件，用于在图像工作流中精确选择边界框
 - 引入了 3 段均衡器节点，用于音频处理工作流
 - 添加了基础文本生成支持，支持原生模型，最初支持 Gemma3 和 Qwen 3
 - 新增使用 PyOpenGL 的 GLSL 着色器节点，用于高级视觉效果
 - 添加了 ElevenLabs API 节点，实现文本转语音功能
 - 为 FLOAT 输入新增渐变滑块显示模式，改进了用户界面
@@ -464,21 +537,21 @@
 - 新增 Sage Attention 3 支持（通过 CLI 参数）
 
 **API 改进**
 - V3 API：公开 DynamicCombo 和 Autogrow 功能
 - API 节点：Kling Omni 支持 720p 分辨率、WAN2.6 ReferenceToVideo
 - Tripo3D：优化 face_limit 参数处理
 - 修复 Vidu API 签名 URL 编码保留问题
 
 **修复**
 - 修复放大模型 CPU 卸载问题
 - 修复 LTXV2 文本编码器 lowvram 兼容性
 - 修复 fp8 处理问题
 - 修复 MPO 格式图像只使用第一帧
 - 根据 PyTorch 版本条件性支持 comfy-kitchen CUDA
 
 **其他**
 - 新增 OOM 内存摘要用于调试
 - Mahiro CFG 显示名称更新
 - 工作流模板更新至 v0.7.67
 - 代码重构与清理（CLIP 预处理、comfy-kitchen 集成）
 </Update>
@@ -487,7 +560,7 @@
 
 **新增节点**
 - 新增 ManualSigmas 节点用于采样控制
 - 新增 Kling Motion Control 节点
 - ResizeByLongerSide 支持视频处理
 
 **系统要求**
@@ -507,7 +580,7 @@
 **API 节点**
 - Gemini API：强制启用提示词增强
 - 计费系统从美元改为积分制
 - 弃用 "seededit"，更新 Seedream 节点显示名称
 
 **技术**
 - 图像处理节点转换为 V3 架构
@@ -516,7 +589,7 @@
 <Update label="v0.6.0" description="2025年12月24日">
 
 **新模型支持**
 - Qwen Image Layered
 - NewBie Image Exp0.1
 
 **新功能**
@@ -526,7 +599,7 @@
 - 设置 API 节点默认禁用水印生成
 - 重构训练器，添加分辨率分桶功能提高训练效率
 - 增强 AMD GPU 支持
 - 扩展 Kling O1 StartEndFrame 节点持续时间范围
 - 更新 Topaz 4k 视频超分辨率功能
 
 **错误修复**
@@ -545,11 +618,11 @@
 
 **新模型和功能**
 - WanMove 高级运动处理模型
 - Qwen Edit 2511 自动检测支持
 - Z-Image Fun Control Union 2.0 修复功能
 
 **API 节点**
 - Kling Omni Image, TextToVideoWithAudio, ImageToVideoWithAudio
 - Wan2.6 模型集成
 - Tripo3.0 模型生成
 
@@ -563,15 +636,15 @@
 <Update label="v0.4.0" description="2025年12月10日">
 
 **模型支持**
 - Ovis Image 多模态图像模型
 - Kandinsky 5.0 (文本转视频/图像转视频/文本转图像变体)
 - Z-Image Alibaba PAI-Fun ControlNet 支持
 - 改进的 Z-Image FP16 兼容性
 - 为CORS标头添加PATCH方法和chroma-radiance-x0模式
 
 **工作流与API增强**
 - 上下文窗口修复和增强的时序处理
 - Kling API @image引用格式
 - ComfyUI-Manager pip安装支持
 
 **前端改进**
@@ -596,7 +669,7 @@
 **错误修复与兼容性**
 - 修复直接在GPU上加载时的文本编码器回归问题
 - 解决Lumina模型中transformer_options清除问题
 - 修复Qwen Image LoRA训练、HunyuanVideo 1.5 meanflow distil问题
 - 纠正EmptyAudio节点输入类型，确保音频工作流正确集成
 - 增强显存计算准确性，防止不必要的完全卸载
 - 支持Z-Image工作流的"transformer." LoRA前缀
@@ -629,12 +702,12 @@
 **性能与模型支持**
 - 支持视频微型 VAE
 - 支持 z-image LoRA 格式
 - 默认在 Nvidia 上启用异步卸载
 
 **合作伙伴节点**
 - Veo3 First-Last-Frame 节点
 - 在 Kling FirstLastFrame 节点中添加 Kling v2.5 turbo
 - Kling O1 模型支持
 
 </Update>
 <Update label="v0.3.75" description="2025年11月26日">
@@ -692,8 +765,8 @@
 
 **API 节点扩展**
 - **新增 Topaz API 节点**：在 ComfyUI 中直接添加了对 Topaz 视频增强工作流的支持
 - **Nano Banana Pro 集成**：通过 Nano Banana Pro 扩展 API 节点集合，增强处理能力
 - **Kling 唇形同步改进**：修复了 KlingLipSyncAudioToVideoNode 中的音频格式转换问题，确保正确处理 MP3 格式
 
 **图像处理与工作流改进**
 - **增强图像批处理**：修复了 ImageBatch 节点以处理不同通道数的图像，并在需要时自动添加 alpha 通道
@@ -776,17 +849,17 @@
 <Update label="v0.3.67" description="2025年10月28日">
 
 **API 节点**
 - **LTXV API 集成**：添加了新的 LTXV API 节点，支持 Lightricks LTX 视频生成
 - Network Client V2 升级，支持异步操作和取消功能
 - 将 Tripo 和 Gemini API 节点转换为 V3 架构
 
 **性能与兼容性**
 - 通过仅在较新的 AMD GPU 上禁用 cudnn 来改进 AMD GPU 支持
 - 修复了 API 节点中 Windows 特定的网络问题，改进重试处理
 
 **核心改进**
 - 增强了依赖感知缓存系统，修复了 --cache-none 在循环中的行为
 - 添加了对多维 latents 的支持
 - 添加了自定义节点发布子图端点
 
 **其他更新**
@@ -806,17 +879,17 @@
 - 加快色度辐射处理速度并修正批量大于 1 时的问题
 
 **API 节点**
 - 新增 Veo 3.1 模型支持
 - 新增 TemporalScoreRescaling 节点，实现视频工作流高级时序控制
 
 **硬件与兼容性**
 - 禁用 AMD gfx942 GPU 的 FP8 运算
 - 优化 --fast autotune 模式下的 CUDA 内存管理
 
 **执行与架构**
 - ControlNet 节点迁移至 V3 架构
 - EasyCache 优化，支持正确的 batch_slice 处理
 - merge_nested_dicts 功能改进，保证输入顺序正确
 - 为未使用文件新增废弃提示
 </Update>
 
@@ -826,18 +899,18 @@
 - 将核心节点迁移到 V3 架构,包括模型缩放、LoRA 提取、合成、潜在操作、SD3/SLG、Flux、放大模型和 HunyuanVideo 节点
 
 **音频和模型改进**
 - 新增 MMaudio 16K VAE 支持高保真音频工作流
 - 修复单声道音频错误保存为立体声问题
 - 重构模型采样代码并修复 FP8 缩放 LoRA 问题
 - 修复旧版 Stable Diffusion 检查点在新版 NumPy 上的加载问题
 
 **AMD GPU 优化**
 - 改进 SD/Flux VAE 操作的内存估算
 - 在 ROCm 7.0+ 上启用 RDNA4 PyTorch 注意力
 
 **API 节点更新**
 - 新增价格提取器并改进 Kling/Pika API 节点
 - 增强 Gemini Image API 的 aspect_ratio 支持
 
 **其他更新**
 - 模板 v0.1.95,节点文档 v0.3.0
@@ -863,17 +936,17 @@
 - **AMD 相关支持**：默认启用 TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL 
 
 **API 节点更新**
 - **Kling 2.5 Turbo**：在 txt2video 和 img2video 节点中新增对 kling-2-5-turbo 的支持，并修复了模式配置问题
 - **API 节点修复**：改进了 Gemini 节点的 base64 处理，并修复了 Recraft API 节点函数中的缩进问题
 
 **节点架构迁移（V3）**
 - **大规模 V3 转换**：将众多节点类别迁移到 V3 架构，包括音频编码器、GITS、差分扩散、最优步数、PAG、LT、IP2P、形态学、torch compile、EPS、Pixverse、TomeSD、编辑模型、Rodin 和 Stable3D 节点，提高了工作流兼容性
 
 **音频系统更新**
 - **依赖清理**：移除了 soundfile 依赖并取消了 torchaudio 加载/保存操作，简化了音频处理管道
 
 **开发者体验改进**
 - **代码质量**：为 comfy_api_nodes 文件夹添加了 pylint 支持，并更新 example_node.py 使用 V3 架构，使自定义节点开发更加一致
 - **文档更新**：为 Windows 用户增加了 AMD 安装说明，添加了 nightly PyTorch 命令
 
 **前端更新**
@@ -901,24 +974,24 @@
 
 **内存和性能优化**
 - **内存泄漏修复**：通过显式分离模型终结器解决了内存泄漏问题，提高长时间运行工作流的稳定性
 - **采样器 CFG 增强**：在采样器 CFG 函数参数中添加了 'input_cond' 和 'input_uncond' 参数，提供更灵活的条件控制
 
 </Update>
 <Update label="v0.3.60" description="2025年9月23日">
 
 **新模型支持**
 - **Wan2.2 Animate**：支持具有角色替换和动作迁移功能的 Wan2.2 Animate 视频生成模型
 - **Qwen Image Edit 2509 支持**：支持更新的 Qwen Image Edit 2509，具有多图像编辑、更高一致性和原生 ControlNet 支持
 - **HuMo 模型**：添加了对 1.7B 和 17B HuMo 模型的支持，使用音频驱动视频生成并保持唇形同步
 - **Chroma Radiance**：一个在像素空间执行图像生成的模型，减少图像生成过程中的损失
 - **Omnigen2 UMO LoRA**：添加了对 Omnigen2 UMO LoRA 模型的支持
 
 **API 节点新增**
 - **Kling v2.1 支持**：在 KlingStartEndFrame 节点中添加了 kling-v2-1 模型
 - **Seedream4 修复**：修复了在部分成功时忽略错误的标志，使工作流更加稳健
 
 **节点架构迁移（V3）**
 - **核心节点更新**：将多个节点类别迁移到 V3 架构，包括 Minimax API、Cosmos、条件、CFG 和 Canny 节点
 
 **性能和技术改进**
 - **FP8 操作**：在 gfx1200 硬件上默认启用 FP8 操作，实现更快的处理速度
@@ -930,20 +1003,20 @@
 </Update>
 <Update label="v0.3.59" description="2025年9月10日">
 
 **字节跳动 Seedream 4.0 集成**
 - **新增 Seedream 节点**：添加了 ByteDanceSeedream (4.0) 节点。
 
 </Update>
 
 <Update label="v0.3.58" description="2025年9月6日">
 
 **新模型支持**
 - Hunyuan Image 2.1 常规模型
 - Hunyuan 3D 2.1
 
 **新增 API 节点**
 - Stable Audio 2.5 API
 - Seedance Video API
 </Update>
 
 <Update label="v0.3.57" description="2025年9月4日">
@@ -955,7 +1028,7 @@
 **工作流实用工具**
 - **ImageScaleToMaxDimension 节点**：用于智能图像缩放的新实用工具
 - **SEEDS 噪声系统**：使用改进算法更新噪声分解
 - **增强提示控制**：中断处理器现在接受 prompt_id 参数
 
 **性能和架构**
 - **V3 架构迁移**：将一些核心节点转换为 V3 架构
@@ -978,7 +1051,7 @@
 **模型支持**
 - Wan2.2 S2V
 - Wan2.2 5B Fun control
 - Wan2.2 5B Fun inpaint
 
 **高级潜在空间处理**
 - **LatentCut节点**：用于在精确点位切割潜在变量的新节点，在复杂生成工作流中实现对潜在空间操作的更精细控制。这对于批处理和时序视频工作流特别有用，比如删除视频中的某些帧。
@@ -1020,12 +1093,12 @@
 - **负音频处理**：修复 S2V 工作流中负音频输入处理，使用适当的零值
 
 **采样与节点增强**
 - **DPM++ 2M SDE Heun (RES) 采样器**：@Balladie 提供的全新高级采样器，为精细调节生成控制提供额外的采样选项
 - **LatentConcat 节点**：用于连接潜在张量的新节点，启用高级潜在空间操作工作流
 - **EasyCache/LazyCache 稳定性**：修复当张量属性（形状/数据类型/设备）在采样期间发生变化时的关键崩溃问题，确保工作流可靠性
 
 **模型兼容性改进**
 - **ControlNet 类型模型**：增强与 Qwen Edit 和 Kontext 工作流配合使用的 ControlNet 类型模型的兼容性修复
 - **Flux 内存优化**：调整 Flux 模型内存使用因子以获得更好的资源利用
 
 **基础设施与可靠性**
@@ -1035,17 +1108,17 @@
 </Update>
 <Update label="v0.3.52" description="2025年8月23日">
 
 **增强模型支持和 Qwen Image ControlNet 集成等**
 
 此版本显著扩展了 ControlNet 功能并改进了模型兼容性，让 ComfyUI 工作流更加多样化和可靠：
 
 **Qwen ControlNet 生态系统**
 - **Diffsynth ControlNet 支持**：新增支持 Qwen Diffsynth ControlNet 的 Canny 和深度条件控制，实现精确的边缘和基于深度的图像控制
 - **InstantX Qwen ControlNet**：集成 InstantX Qwen ControlNet，扩展创意控制选项
 - **修复补丁 ControlNet/模型**：通过专用的 Diffsynth 修复 ControlNet 支持增强修复功能
 
 **节点架构和 API 演进**
 - **V3 架构迁移**：字符串节点、Google Veo API 和 Ideogram API 节点升级到 V3 架构，提供更好的性能和一致性
 - **增强的 API 节点**：OpenAI Chat 节点重命名为"OpenAI ChatGPT" 提高辨识度，Gemini Chat 节点现在包含复制按钮功能
 - **改进的可用性**：API 节点现在通过更清晰的标签和增强的交互功能提供更好的用户体验
 
@@ -1068,7 +1141,7 @@
 - **Qwen-Image-Edit模型**：原生支持 Qwen-Image-Edit
 - **FluxKontextMultiReferenceLatentMethod 节点**：用于 Flux 工作流的多图参考输入节点
 - **WAN 2.2 Fun Camera模型支持**：支持通过相机控制生成视频
 - **模板更新**：升级到版本 0.1.62，新增了 Wan2.2 Fun Camera 和 Qwen Image Edit 等模板
 
 **核心功能改进**
 - **Context Windows上下文节点支持**：增强采样代码，支持更长序列的生成任务
@@ -1080,9 +1153,9 @@
 
 **API节点支持更新**
 - **GPT-5 系列模型**：支持最新的 GPT-5 模型
 - **Kling V2-1 和 V2-1-Master**：更新视频生成模型功能
 - **Minimax Hailuo 视频节点**：新增视频生成节点
 - **Vidu 视频节点**：Vidu API 节点支持
 - **Google 模型更新**：新增 Google Gemini 的新模型
 - **OpenAI API 修复**：修复 OpenAI API 节点输入图像的 MIME 类型导致的错误
 
@@ -1112,7 +1185,7 @@
 **Qwen模型生态**
 - **Qwen图像模型支持**：改进了LoRA加载和模型合并功能，支持更复杂的视觉工作流程
 - **Qwen模型合并节点**：新增节点，支持合并Qwen图像模型
 - **SimpleTuner Lycoris LoRA支持**：扩展了对Qwen-Image模型的Lycoris LoRA兼容性
 
 **API与性能**
 - **异步API节点**：引入异步API节点，支持非阻塞工作流程执行
@@ -1127,7 +1200,7 @@
 **开发体验**
 - **日志记录优化**：功能标志仅在详细模式下显示
 - **音频处理安全**：增强了torchaudio导入安全检查
 - **Kling API改进**：修复了Kling Image API节点的图像类型参数
 
 </Update>
 <Update label="v0.3.49" description="2025年8月5日">
@@ -1207,7 +1280,7 @@
 - **音频集成优化**：增强了多媒体生成工作流的音频处理
 
 **硬件支持**
 - **Iluvatar CoreX支持**：添加了对Iluvatar CoreX加速器的支持
 - **Intel XPU优化**：改进了XPU支持
 - **AMD ROCm增强**：在Torch 2.8上为gfx1201启用了PyTorch attention
 - **CUDA内存管理**：修复了CUDA malloc激活问题
@@ -1221,7 +1294,7 @@
 - **训练灵活性**：提供了更好的内存管理和性能优化
 
 **节点和工作流**
 - **Moonvalley V2V节点**：添加了Moonvalley Marey V2V节点
 - **负面提示词更新**：改进了Moonvalley节点的负面提示词处理
 - **历史API增强**：向get_history API添加了map_function参数
 
@@ -1261,7 +1334,7 @@
 
 **API和集成**
 - **自定义提示ID**：API允许指定提示ID
 - **Kling API优化**：增加了轮询超时时间
 - **历史令牌清理**：从历史项目中删除了敏感令牌
 - **Python 3.9兼容性**：修复了兼容性问题