v0.8.1

What's Changed

• Fix stable release workflow not pulling latest comfy kitchen. by @comfyanonymous in #11695
• Support gemma 12B with quant weights. by @comfyanonymous in #11696
• model_patcher: Remove confusing load stat by @rattus128 in #11710
• Add device selection for LTXAVTextEncoderLoader by @kijai in #11700
• Lower ltxv text encoder vram use. by @comfyanonymous in #11713
• Bump required comfy-kitchen version. by @comfyanonymous in #11714
• Increase ltxav mem estimation by a bit. by @comfyanonymous in #11715
• Add memory estimation function to ltxav text encoder. by @comfyanonymous in #11716
• ops: Fix offloading with FP8MM performance by @rattus128 in #11697
• Add warning for old pytorch and comfy kitchen. Update your pytorch to cu130 to get a performance boost. by @comfyanonymous in #11718
• Update template to 0.7.69 by @comfyui-wiki in #11719

Full Changelog: v0.8.0...v0.8.1

v0.8.0

What's Changed

• Add support for sage attention 3 in comfyui, enable via new cli arg by @mengqin in #11026
• V3 Improvements + DynamicCombo + Autogrow exposed in public API by @Kosinkadink in #11345
• fix(api-nodes-vidu): preserve percent-encoding for signed URLs by @bigcat88 in #11564
• Update workflow templates to v0.7.65 by @comfyui-wiki in #11579
• Refactor: move clip_preprocess to comfy.clip_model by @comfyanonymous in #11586
• Remove duplicate import of model_management by @comfyanonymous in #11587
• New Year ruff cleanup. by @comfyanonymous in #11595
• Ignore all frames except the first one for MPO format. by @bigcat88 in #11569
• Give Mahiro CFG a more appropriate display name by @throttlekitty in #11580
• [API Nodes] Tripo3D: pass face_limit parameter only when it differs from default by @bigcat88 in #11601
• Remove leftover scaled_fp8 key. by @comfyanonymous in #11603
• Print memory summary on OOM to help with debugging. by @comfyanonymous in #11613
• [API Nodes] add support for 720p resolution for Kling Omni nodes by @bigcat88 in #11604
• Fix case where upscale model wouldn't be moved to cpu. by @comfyanonymous in #11633
• Support the LTXV 2 model. by @comfyanonymous in #11632
• Add LTXAVTextEncoderLoader node. by @comfyanonymous in #11634
• Refactor module_size function. by @comfyanonymous in #11637
• Fix name. by @comfyanonymous in #11638
• Initial ops changes to use comfy_kitchen: Initial nvfp4 checkpoint support. by @comfyanonymous in #11635
• Use rope functions from comfy kitchen. by @comfyanonymous in #11647
• Revert "Use rope functions from comfy kitchen." by @comfyanonymous in #11648
• Update workflow templates to v0.7.66 by @comfyui-wiki in #11652
• Add helpful message to portable. by @comfyanonymous in #11671
• Update comfy-kitchen version to 0.2.1 by @comfyanonymous in #11672
• Use rope functions from comfy kitchen. by @comfyanonymous in #11674
• Update workflow templates to v0.7.67 by @comfyui-wiki in #11667
• Fix lowvram issue with ltxv2 text encoder. by @comfyanonymous in #11675
• Disable ltxav previews. by @comfyanonymous in #11676
• Skip fp4 matrix mult on devices that don't support it. by @comfyanonymous in #11677
• Disable comfy kitchen cuda if pytorch cuda less than 13 by @comfyanonymous in #11681
• Update comfy-kitchen version to 0.2.3 by @comfyanonymous in #11685
• Force sequential execution in CI test jobs by @yoland68 in #11687
• [API Nodes] add WAN2.6 ReferenceToVideo by @bigcat88 in #11644
• Fix fp8 fast issue. by @comfyanonymous in #11688

New Contributors

• @mengqin made their first contribution in #11026
• @throttlekitty made their first contribution in #11580

Full Changelog: v0.7.0...v0.8.0

v0.7.0

What's Changed

• Bump comfyui-frontend-package to 1.35.9 by @comfy-pr-bot in #11470
• Update workflow templates to v0.7.64 by @comfyui-wiki in #11496
• Add a ManualSigmas node. by @comfyanonymous in #11499
• Specify in readme that we only support pytorch 2.4 and up. by @comfyanonymous in #11512
• bump comfyui_manager version to the 4.0.4 by @ltdrdata in #11521
• Fix noise with ancestral samplers when inferencing on cpu. by @comfyanonymous in #11528
• feat(api-nodes): add Kling Motion Control node by @bigcat88 in #11493
• [V3] converted nodes_images.py to V3 schema by @bigcat88 in #11206
• fix(api-nodes-gemini): always force enhance_prompt to be True by @bigcat88 in #11503
• chore(api-nodes): move to credits from $ by @bigcat88 in #11489
• Enable async offload by default for AMD. by @comfyanonymous in #11534
• Comment out unused norm_final in lumina/z image model. by @comfyanonymous in #11545
• mm: discard async errors from pinning failures by @rattus128 in #10738
• Add some warnings for pin and unpin errors. by @comfyanonymous in #11561
• ResizeByLongerSide: support video by @tavihalperin in #11555
• Mark "seededit" as deprecated, adjust display name of Seedream node by @bigcat88 in #11490
• Add handling for vace_context in context windows by @drozbay in #11386

New Contributors

• @tavihalperin made their first contribution in #11555

Full Changelog: v0.6.0...v0.7.0

v0.6.0

What's Changed

• Resolution bucketing and Trainer implementation refactoring by @KohakuBlueleaf in #11117
• skip_load_model -> force_full_load by @comfyanonymous in #11390
• Add unified jobs API with /api/jobs endpoints by @ric-yu in #11054
• Update workflow templates to v0.7.60 by @comfyui-wiki in #11403
• Support loading Wan/Qwen VAEs with different in/out channels. by @comfyanonymous in #11405
• Trim/pad channels in VAE code. by @comfyanonymous in #11406
• Diffusion model part of Qwen Image Layered. by @comfyanonymous in #11408
• Add LatentCutToBatch node. by @comfyanonymous in #11411
• [API Nodes] add Flux2MaxImage API Node by @bigcat88 in #11420
• bump comfyui_manager version to the 4.0.3b7 by @ltdrdata in #11422
• Allow enabling use of MIOpen by setting an env var by @BradPepersAMD in #11366
• Support nested tensor denoise masks. by @comfyanonymous in #11431
• Fix error from logging line in Context Windows by @drozbay in #11423
• Implement sliding attention in Gemma3 by @woct0rdho in #11409
• Disable prompt weights on newbie te. by @comfyanonymous in #11434
• ZImageFunControlNet: Fix mask concatenation in --gpu-only by @rattus128 in #11421
• Fix issue with batches and newbie. by @comfyanonymous in #11435
• Implement Jina CLIP v2 and NewBie dual CLIP by @woct0rdho in #11415
• Only apply gemma quant config to gemma model for newbie. by @comfyanonymous in #11436
• chore(api-nodes): by default set Watermark generation to False by @bigcat88 in #11437
• fix(api-nodes): Topaz 4k video upscaling by @bigcat88 in #11438
• Core release process. by @comfyanonymous in #11447
• Add node to create empty latents for qwen image layered model. by @comfyanonymous in #11460
• [API Nodes] extend possible duration range for Kling O1 StartEndFrame node by @bigcat88 in #11451
• Update workflow templates to v0.7.62 by @comfyui-wiki in #11467
• Make denoised output on custom sampler nodes work with nested tensors. by @comfyanonymous in #11471
• api-nodes: use new custom endpoint for Nano Banana by @bigcat88 in #11311
• Update workflow templates to v0.7.63 by @comfyui-wiki in #11482

New Contributors

• @BradPepersAMD made their first contribution in #11366
• @woct0rdho made their first contribution in #11409

Full Changelog: v0.5.1...v0.6.0

v0.5.1

What's Changed

• [API Nodes] GPT-Image-1.5 by @bigcat88 in #11368
• fix regression in V3 nodes processing by @bigcat88 in #11375
• SA-Solver: return the result of the last predictor step correctly by @chaObserv in #11380
• Better handle torch being imported by prestartup nodes. by @comfyanonymous in #11383

Full Changelog: v0.5.0...v0.5.1

v0.5.0

What's Changed

• Bump frontend to 1.34.8 by @benceruleanlu in #11220
• fix: correctly process the Node V1 dict results on V3 node by @bigcat88 in #11237
• Tweak Z Image memory estimation. by @comfyanonymous in #11254
• Fix: filter hidden files from /internal/files endpoint by @Myestery in #11191
• Lower VAE loading requirements：Create a new branch for GPU memory calculations in qwen-image vae by @jiangchengchengark in #11199
• feat(api-nodes): enable Kling Omni Image node by @bigcat88 in #11229
• Adjust memory usage factor for sd3. by @comfyanonymous in #11257
• Fix regular chroma radiance by @comfyanonymous in #11276
• This only works on radiance. by @comfyanonymous in #11277
• Better chroma radiance vram estimation. by @comfyanonymous in #11278
• Make portable updater work with repos in unmerged state. by @comfyanonymous in #11281
• WanMove support by @kijai in #11247
• Respect the dtype the op was initialized in for non quant mixed op. by @comfyanonymous in #11282
• feat(api-nodes): Kling TextToVideoWithAudio and ImageToVideoWithAudio nodes by @bigcat88 in #11267
• Fix bias dtype issue in mixed ops. by @comfyanonymous in #11293
• Basic implementation of z image fun control union 2.0 by @comfyanonymous in #11304
• Fix pytorch warnings. by @comfyanonymous in #11314
• seeds_2: add phi_2 variant and sampler node by @chaObserv in #11309
• Update warning for old pytorch version. by @comfyanonymous in #11319
• bump manager to the 4.0.3b5 by @ltdrdata in #11324
• Update workflow templates to v0.7.59 by @comfyui-wiki in #11337
• Disable guards on transformer_options when torch.compile by @comfyanonymous in #11317
• [API Nodes] drop very old Kling models by @bigcat88 in #11307
• drop Pika API nodes by @bigcat88 in #11306
• feat(preview): add per-queue live preview method override by @ltdrdata in #11261
• comfy_api: remove usage of "Type","List" and "Dict" types by @bigcat88 in #11238
• Add context windows callback for custom cond handling by @drozbay in #11208
• Support the new qwen edit 2511 reference method. by @comfyanonymous in #11340
• [add] tripo3.0 by @seed93 in #10663
• Add code to detect if a z image fun controlnet is broken or not. by @comfyanonymous in #11341
• [BlockInfo] Lumina by @Haoming02 in #11227
• [BlockInfo] Wan by @Haoming02 in #10845
• Only enable fp16 on ZImage on newer pytorch. by @comfyanonymous in #11344
• Patch frontend to 1.34.9 (from 1.34.8) by @christian-byrne in #11342
• Inpainting for z image fun control. Use the ZImageFunControlnet node. by @comfyanonymous in #11346
• Add a way to set the default ref method in the qwen image code. by @comfyanonymous in #11349
• Update workflows for new release process by @benceruleanlu in #11064
• [API Nodes] add Wan2.6 model to video nodes by @bigcat88 in #11357
• Check state dict key to auto enable the index_timestep_zero ref method. by @comfyanonymous in #11362
• Add exp_heun_2_x0 sampler series by @chaObserv in #11360

New Contributors

• @jiangchengchengark made their first contribution in #11199
• @drozbay made their first contribution in #11208
• @seed93 made their first contribution in #10663

Full Changelog: v0.4.0...v0.5.0

v0.4.0

What's Changed

• Implement the Ovis image model. by @comfyanonymous in #11030
• Bump frontend to 1.33 by @christian-byrne in #11028
• feat: Support ComfyUI-Manager for pip version by @ltdrdata in #7555
• Add @guill as a code owner by @yoland68 in #11031
• Fix CODEOWNERS formatting to have all on the same line by @Kosinkadink in #11053
• add check for the format arg type in VideoFromComponents.save_to function by @bigcat88 in #11046
• Hack to make zimage work in fp16. by @comfyanonymous in #11057
• attention: use flag based OOM fallback by @rattus128 in #11038
• Support Z Image alipai controlnets. by @comfyanonymous in #11062
• Added PATCH method to CORS headers by @jheising in #11066
• Implement temporal rolling VAE (Major VRAM reductions in Hunyuan and Kandinsky) by @rattus128 in #10995
• Add MatchType, DynamicCombo, and Autogrow support to V3 Schema by @Kosinkadink in #10832
• Fix issue with portable updater. by @comfyanonymous in #11070
• Prs/lora reservations (reduce massive Lora reservations especially on Flux2) by @rattus128 in #11069
• use empty list defaults for Schema.inputs/outputs/hidden to avoid None issues by @bigcat88 in #11083
• [API nodes] add support for "@image" reference format in Kling Omni API nodes by @bigcat88 in #11082
• [V3] convert nodes_load_3d.py to V3 schema by @bigcat88 in #10990
• [V3] convert nodes_audio.py to V3 schema by @bigcat88 in #10798
• Qwen Image Lora training fix from #11090 by @comfyanonymous in #11094
• Fix case where text encoders where running on the CPU instead of GPU. by @comfyanonymous in #11095
• mp: use look-ahead actuals for stream offload VRAM calculation (fixes unwanted TE full offload) by @rattus128 in #11096
• sd: revise hy VAE VRAM by @rattus128 in #11105
• sd: bump HY1.5 VAE estimate by @rattus128 in #11107
• [API Nodes]: fixes and refactor by @bigcat88 in #11104
• Forgot to put this in README. by @comfyanonymous in #11112
• Fix transformer_options getting cleared in Lumina model after z-image PR by @Kosinkadink in #11118
• Make old scaled fp8 format use the new mixed quant ops system. by @comfyanonymous in #11000
• Fix regression when text encoder loaded directly on GPU. by @comfyanonymous in #11129
• Context windows fixes and features by @kijai in #10975
• Fix some custom nodes. by @comfyanonymous in #11134
• docs: add ComfyUI-Manager documentation and update to v4.0.3b4 by @ltdrdata in #11133
• Kandinsky5 model support by @kijai in #10988
• Fix regression. by @comfyanonymous in #11137
• [V3] convert nodes_freelunch.py to the V3 schema by @bigcat88 in #10904
• [V3] convert nodes_mask.py to V3 schema by @bigcat88 in #10669
• Set OCL_SET_SVM_SIZE on AMD. by @comfyanonymous in #11139
• marked all Pika API nodes a deprecated by @bigcat88 in #11146
• Bugfix: Wrong EmptyAudio node input types by @kijai in #11149
• Speed up lora compute and lower memory usage by doing it in fp16. by @comfyanonymous in #11161
• Fix on-load VRAM OOM by @rattus128 in #11144
• Fix qwen scaled fp8 not working with kandinsky. Make basic t2i wf work. by @comfyanonymous in #11162
• Properly load the newbie diffusion model. by @comfyanonymous in #11172
• chore(comfy_api): replace absolute imports with relative by @bigcat88 in #11145
• Update workflow templates to v0.7.51 by @comfyui-wiki in #11150
• chore: replace imports of deprecated V1 classes by @bigcat88 in #11127
• [API Nodes] Add System Instruction support & Visual Enforcement for Banana Nodes by @bigcat88 in #11177
• [API Nodes] add support for seedance-1-0-pro-fast model by @bigcat88 in #10947
• Support "transformer." LoRA prefix for Z-Image by @dxqb in #11135
• retune lowVramPatch VRAM accounting by @rattus128 in #11173
• Update workflow templates to v0.7.54 by @comfyui-wiki in #11192
• Fix regression. by @comfyanonymous in #11194
• Fix potential issue. by @comfyanonymous in #11201
• dequantization offload accounting (fixes Flux2 OOMs - incl TEs) by @rattus128 in #11171
• Patch frontend to 1.33.13 (from 1.33.10) by @christian-byrne in #11200
• add chroma-radiance-x0 mode by @lodestone-rock in #11197
• ops: delete dead code by @rattus128 in #11204
• Fix for HunyuanVideo1.5 meanflow distil by @kijai in #11212
• Fix nan issue when quantizing fp16 tensor. by @comfyanonymous in #11213

New Contributors

• @jheising made their first contribution in #11066
• @dxqb made their first contribution in #11135
• @lodestone-rock made their first contribution in #11197

Full Changelog: v0.3.76...v0.4.0

v0.3.76

What's Changed

• Add cheap latent preview for flux 2. by @comfyanonymous in #10907
• [API Nodes] add Veo3 First-Last-Frame node by @bigcat88 in #10878
• [API Nodes] improve UX for batch uploads in upload_images_to_comfyapi by @bigcat88 in #10913
• [API Nodes] fix(gemini): use first 10 images as fileData (URLs) and remaining images as inline base64 by @bigcat88 in #10918
• Bump frontend to 1.32.9 by @christian-byrne in #10867
• Merge 3d animation node by @jtydhr88 in #10025
• Fix the CSP offline feature on latest frontend. by @comfyanonymous in #10923
• Add Z Image to readme. by @comfyanonymous in #10924
• chore(api-nodes): remove chat widgets from OpenAI/Gemini nodes by @bigcat88 in #10861
• [V3] convert nodes_custom_sampler.py to V3 schema by @bigcat88 in #10206
• Dataset Processing Nodes and Improved LoRA Trainer Nodes with multi resolution supports. by @KohakuBlueleaf in #10708
• Make lora training work on Z Image and remove some redundant nodes. by @comfyanonymous in #10927
• [BlockInfo] Flux by @Haoming02 in #10841
• Account for the VRAM cost of weight offloading by @rattus128 in #10733
• quant ops: Dequantize weight in-place (reduce flux2 VRAM usage) by @rattus128 in #10935
• Update template to 0.7.23 by @comfyui-wiki in #10949
• Enable async offloading by default on Nvidia. by @comfyanonymous in #10953
• feat(Kling-API-Nodes): add v2-5-turbo model to FirstLastFrame node by @bigcat88 in #10938
• fix(user_manager): fix typo in move_userdata dest validation by @ltdrdata in #10967
• Disable offload stream when torch compile. by @comfyanonymous in #10961
• fix QuantizedTensor.is_contiguous (#10956) by @urlesistiana in #10959
• mm: wrap the raw stream in context manager by @rattus128 in #10958
• Update driver link in AMD portable README by @comfyanonymous in #10974
• Support video tiny VAEs by @kijai in #10884
• Support some z image lora formats. by @comfyanonymous in #10978
• feat(security): add System User protection with __ prefix by @ltdrdata in #10966
• Add missing z image lora layers. by @comfyanonymous in #10980
• Make the ScaleRope node work on Z Image and Lumina. by @comfyanonymous in #10994
• update template to 0.7.25 by @comfyui-wiki in #10996
• Next AMD portable will have pytorch with ROCm 7.1.1 by @comfyanonymous in #11002
• Bumps frontend to 1.32.10 (from 1.32.9) by @christian-byrne in #11018
• Update qwen tokenizer to add qwen 3 tokens. by @comfyanonymous in #11029
• [API Nodes] add Kling O1 model support by @bigcat88 in #11025

New Contributors

• @urlesistiana made their first contribution in #10959

Full Changelog: v0.3.75...v0.3.76

v0.3.75

What's Changed

• Z Image model. by @comfyanonymous in #10892
• Adjustments to Z Image. by @comfyanonymous in #10893
• Fix loras not working on mixed fp8. by @comfyanonymous in #10899
• Fix Flux2 reference image mem estimation. by @comfyanonymous in #10905

Full Changelog: v0.3.73...v0.3.75

v0.3.73

What's Changed

• Fix crash. by @comfyanonymous in #10885
• Update workflow templates to v0.7.20 by @comfyui-wiki in #10883
• Lower vram usage for flux 2 text encoder. by @comfyanonymous in #10887

Full Changelog: v0.3.72...v0.3.73