What's Changed

• Frontend fix to prevent some api nodes from executing again when none of the inputs change.
• Memory estimation code can now take into account conds. by @comfyanonymous in #8307
• Update frontend to v1.20.7 (patch) by @christian-byrne in #8312

Full Changelog: v0.3.37...v0.3.38

What's Changed

• Return proper error if diffusion model not detected properly. by @comfyanonymous in #8272
• Enable pytorch attention by default on AMD gfx1151 by @comfyanonymous in #8282
• Disable initial GPU load when novram is used. by @comfyanonymous in #8294
• [BugFix] Update frontend to 1.20.6 by @webfiltered in #8296
• More API nodes: Gemini/Open AI Chat, Tripo, Rodin, Runway Image by @robinjhuang in #8295

Full Changelog: v0.3.36...v0.3.37

What's Changed

• Make torch.compile LoRA/key-compatible by @Kosinkadink in #8213
• Code Optimization and Issues Fixes in ComfyUI server by @programming-pupil in #8196
• Improve performance with large number of queued prompts by @miabrahams in #8176
• Update templates to 0.1.18 by @comfyui-wiki in #8224
• Update frontend to 1.20 by @huchenlei in #8232
• remove mtl from 3d model file list by @jtydhr88 in #8192
• Add append feature to conditioning_set_values by @comfyanonymous in #8239
• Make VACE conditionings stackable. by @comfyanonymous in #8240
• Remove blocking call from timestep embedding of Chroma by @drhead in #8255
• Add argument to explicitly enable fp8 compute support. by @comfyanonymous in #8257
• Update frontend to v1.20.5 (patch) by @christian-byrne in #8260
• Better error if sageattention is installed but a dependency is missing. by @comfyanonymous in #8264

New Contributors

• @programming-pupil made their first contribution in #8196
• @comfyui-wiki made their first contribution in #8224
• @drhead made their first contribution in #8255

Full Changelog: v0.3.35...v0.3.36

What's Changed

• add opus and mp3 to audio output node by @thot-experiment in #8019
• Add nodes for basic string operations by @jtydhr88 in #7952
• Update frontend to v1.19 by @huchenlei in #8076
• add support to record video as output for 3d node by @jtydhr88 in #7927
• implement APG guidance by @thot-experiment in #8081
• Add comment in example script denoting how to use api keys without needing frontend by @christian-byrne in #8082
• Support official ACE Step loras. by @comfyanonymous in #8094
• Fix issue with old pytorch RMSNorm. by @comfyanonymous in #8095
• fix negative momentum in APG node by @thot-experiment in #8100
• Hunyuan Custom initial untested implementation. by @comfyanonymous in #8101
• rework client.py to be more robust, add logging of api requests by @thot-experiment in #7988
• Display progress and result URL directly on API nodes by @christian-byrne in #8102
• Fix bug: don't retry request if API reports task failure by @christian-byrne in #8111
• Update comment on ROCm pytorch attention in README. by @comfyanonymous in #8123
• Add get_duration method to Comfy VIDEO type by @christian-byrne in #8122
• Remove old hack used to fix windows pytorch 2.4 on the portable. by @comfyanonymous in #8139
• Add Wan-FUN Camera Control models and Add WanCameraImageToVideo node by @George0726 in #8013
• Fix typo in last PR. by @comfyanonymous in #8144
• Remove Desktop versioning claim from README by @webfiltered in #8155
• Make ImagePadForOutpaint return a 3 channel mask. by @comfyanonymous in #8157
• Remove useless log. by @comfyanonymous in #8166
• Add missing category for T5TokenizerOption by @silveroxides in #8177
• Validate video inputs by @christian-byrne in #8133
• Node to add pixel space noise to an image. by @comfyanonymous in #8182
• Add delimiter to concatenate node by @LaVie024 in #8173
• Remove default delimiter. by @comfyanonymous in #8183
• Update AMD instructions in README. by @comfyanonymous in #8198
• Update README ROCm text to match link by @webfiltered in #8199
• This doesn't seem to be needed on chroma. by @comfyanonymous in #8209
• Revert "This doesn't seem to be needed on chroma." by @comfyanonymous in #8210

New Contributors

• @George0726 made their first contribution in #8013
• @LaVie024 made their first contribution in #8173

Full Changelog: v0.3.34...v0.3.35

What's Changed

• Add ACE Step to README. by @comfyanonymous in #8005
• Add --mmap-torch-files to enable use of mmap when loading ckpt/pt by @comfyanonymous in #8021
• move SVG to core by @thot-experiment in #7982
• Use normal ComfyUI attention in ACE-Steps model by @blepping in #8023
• Update node tooltips and validation by @christian-byrne in #8036
• Fix res_multistep_ancestral and res_multistep_ancestral_cfg_pp sampler by @pamparamm in #8030
• Fix issue with fp8 ops on some models. by @comfyanonymous in #8045
• Add method to stream text to node UI by @christian-byrne in #8018
• Add support for Comfy API keys by @christian-byrne in #8041
• ACE VAE works in fp16. by @comfyanonymous in #8055
• Remove extraneous aspect_ratio field from Kling I2V by @christian-byrne in #8062
• Fix torch warning about deprecated function. by @comfyanonymous in #8075

Full Changelog: v0.3.33...v0.3.34

What's Changed

• Change chroma to use Flux shift. by @comfyanonymous in #7961
• Better mem estimation for the LTXV 13B model. by @comfyanonymous in #7963
• Initial ACE-Step model implementation. by @comfyanonymous in #7972
• Refuse to load api nodes on old pyav version. by @comfyanonymous in #7981
• Experimental lyrics strength for ACE. by @comfyanonymous in #7984
• Detection code to make ltxv models without config work. by @comfyanonymous in #7986
• Make torchaudio not a hard requirement. by @comfyanonymous in #7987
• Adjust memory estimation code for ACE VAE. by @comfyanonymous in #7990
• Make japanese hiragana and katakana characters work with ACE. by @comfyanonymous in #7997
• Make ACE VAE tiling work. by @comfyanonymous in #8004

Full Changelog: v0.3.32...v0.3.33

What's Changed

• More API nodes by @Kosinkadink in #7956
• Add a way to disable api nodes: --disable-api-nodes by @comfyanonymous in #7960
• Switch mochi and wan modes to use pytorch RMSNorm. by @comfyanonymous in #7925
• Change cosmos and hydit models to use the native RMSNorm. by @comfyanonymous in #7934
• Change lumina to native RMSNorm. by @comfyanonymous in #7935
• Support saving Comfy VIDEO type to buffer by @christian-byrne in #7939
• Fix: Error creating video when sliced audio tensor chunks are non-c-contiguous by @christian-byrne in #7942
• This should not be a warning. by @comfyanonymous in #7946

Full Changelog: v0.3.31...v0.3.32

What's Changed

• [NodeDef] Add documentation on widgetType by @AustinMroz in #7768
• Add a T5TokenizerOptions node to set options for the T5 tokenizer. by @comfyanonymous in #7803
• Update portable package workflow to cu128 by @comfyanonymous in #7812
• Add experimental --async-offload lowvram weight offloading. by @comfyanonymous in #7820
• Make loras work with --async-offload by @comfyanonymous in #7824
• Prevent custom nodes from hooking certain functions. by @comfyanonymous in #7825
• Use stream for casting if enabled. by @comfyanonymous in #7833
• Add /api/v2/userdata endpoint by @benceruleanlu in #7817
• ltxv: add strength parameter to conditioning. by @kvochko in #7849
• Save V-Pred ZSNR metadata in CheckpointSave by @pamparamm in #7840
• Fix stream priority to support older pytorch. by @comfyanonymous in #7856
• ltxv: overwrite existing mask on conditioned frame. by @kvochko in #7845
• CFG++ for gradient estimation sampler by @chaObserv in #7809
• Add moderation level option to OpenAIGPTImage1 by @yoland68 in #7804
• Support HiDream E1 model. by @comfyanonymous in #7857
• Add support for VIDEO as a built-in type by @guill in #7844
• Latest desktop and portable should work on blackwell. by @comfyanonymous in #7861
• Add release process section to README by @huchenlei in #7855
• support more example folders by @jtydhr88 in #7836
• Per device stream counters for async offload. by @comfyanonymous in #7873
• Bump minimum pyav version to 14.2.0 by @comfyanonymous in #7874
• Better vace memory estimation. by @comfyanonymous in #7875
• Test updater in the windows release workflow. by @comfyanonymous in #7886
• Add updater test to stable release workflow. by @comfyanonymous in #7887
• Support for Chroma - Flux1 Schnell distilled with CFG by @silveroxides in #7355
• Fix content moderation not serialized properly bug by @christian-byrne in #7894
• Switch ltxv to use the pytorch RMSNorm. by @comfyanonymous in #7897
• Update frontend to v1.18 by @huchenlei in #7898
• Lint node method's self argument by @catboxanon in #7903
• Make clipsave work with more TE models. by @comfyanonymous in #7908
• Add node to extend sigmas by @catboxanon in #7901
• Fix outdated comment about Internet connectivity by @catboxanon in #7827
• Fix the bugs in OFT/BOFT moule by @KohakuBlueleaf in #7909
• [BugFix] Update frontend 1.18.6 by @huchenlei in #7910
• upstream Preview Any from rgthree-comfy by @jtydhr88 in #7815
• Fix updater issue with newer portable. by @comfyanonymous in #7917

New Contributors

• @AustinMroz made their first contribution in #7768
• @benceruleanlu made their first contribution in #7817

Full Changelog: v0.3.30...v0.3.31

What's Changed

• Add FreSca node by @bvhari in #7631
• [Type] Annotate graph.get_input_info by @huchenlei in #7386
• add support to output camera state by @jtydhr88 in #7582
• Replace CLIPType if with getattr by @city96 in #7589
• [BugFix] Update frontend to 1.16.9 by @huchenlei in #7655
• Add wanfun template workflows. by @robinjhuang in #7678
• Update CODEOWNERS and add christian-byrne by @yoland68 in #7663
• Hidream: Allow loading hidream text encoders in CLIPLoader and DualCLIPLoader by @power88 in #7676
• Slightly better fix for #7687 by @comfyanonymous in #7690
• CLIPTextEncodeHiDream. by @comfyanonymous in #7703
• Populate AUTH_TOKEN_COMFY_ORG hidden input by @huchenlei in #7709
• Support for WAN VACE preview model. by @comfyanonymous in #7711
• Add node type hint for socketless option by @webfiltered in #7714
• Update nodes_primitive.py by @Amorano in #7716
• Add strength control for vace. by @comfyanonymous in #7717
• Unified Weight Adapter system for better maintainability and future feature of Lora system by @KohakuBlueleaf in #7540
• Fix issue with WAN VACE implementation. by @comfyanonymous in #7724
• Add OFT/BOFT algorithm in weight adapter by @KohakuBlueleaf in #7725
• upstream MaskPreview from ComfyUI_essentials by @jtydhr88 in #7719
• Correct DoRA implementation with backward compatibility by @KohakuBlueleaf in #7727
• Add option for using fp8_e8m0fnu for model weights. by @comfyanonymous in #7733
• Replace aom-av1 with svt-av1 for saving webm videos, use preset 6 + yuv420p10le pixel format by @alexheretic in #7736
• [API nodes] Add backbone for supporting api nodes in ComfyUI by @robinjhuang in #7745
• Fix for dino lowvram. by @comfyanonymous in #7748
• Pytorch stable 2.7 is out and support cu128 by @comfyanonymous in #7749
• fp16 accumulation can now be enabled on the stable package. by @comfyanonymous in #7750
• Lower size of release package. by @comfyanonymous in #7751
• Lower size of release package more. by @comfyanonymous in #7754
• Add support for API Nodes in ComfyUI. by @robinjhuang in #7726
• [BugFix] Update frontend to 1.17.10 by @webfiltered in #7762
• implement multi image prompting for gpt-image-1 and fix transparency in outputs by @thot-experiment in #7763
• change timeout to 7 days because there's no reason to actually timeout eagerly by @thot-experiment in #7765
• [BugFix] Update frontend to 1.17.11 by @webfiltered in #7766
• throw away alpha channel in clip vision preprocessor by @thot-experiment in #7769
• fix multiple image return from api nodes by @thot-experiment in #7772
• Switch LTXVPreprocess to libx264 by @comfyanonymous in #7776

New Contributors

• @power88 made their first contribution in #7676
• @Amorano made their first contribution in #7716
• @KohakuBlueleaf made their first contribution in #7540
• @alexheretic made their first contribution in #7736

Full Changelog: v0.3.29...v0.3.30

What's Changed

• Support for HiDream models
• Support for Wan 2.1 First Frame Last Frame model.
• Support for Wan Fun control and inpaint models.
• Update frontend to 1.14.6 by @huchenlei in #7416
• ltxv: fix preprocessing exception when compression is 0. by @michaellightricks in #7431
• Add option to store TE in bf16 by @bvhari in #7461
• User missing by @lerignoux in #7439
• MLU memory optimization by @BiologicalExplosion in #7470
• Add activations_shape info in UNet models by @Slickytail in #7482
• Handle PNG-8 indexed transparency in LoadImage node by @christian-byrne in #7539
• Prevent custom nodes from accidentally overwriting global modules. by @comfyanonymous in #7167
• Align /prompt response schema by @huchenlei in #7423
• add VoxelToMesh node w/ surfacenet meshing by @thot-experiment in #7446
• Add prepare_sampling wrapper to allow modifying noise_shape by @Kosinkadink in #7500
• Deprecate InputTypeOptions.defaultInput by @huchenlei in #7551
• Dont cache workflow templates index.json file by @christian-byrne in #7569
• Update frontend to 1.15 by @huchenlei in #7564
• Dependency Aware Node Caching for low RAM/VRAM machines by @Chargeuk in #7509
• Only show image files in LoadImage file select combo by @christian-byrne in #7573
• Add CublasOps support by @catboxanon in #7574
• Update filter_files_content_types to support filtering 3d models by @christian-byrne in #7572
• Add SEEDS (stage 2 & 3 DP) sampler by @chaObserv in #7580
• add OptimalSteps by @bebebe666 in #7584
• [Type] Mark input options NotRequired by @huchenlei in #7614
• Update frontend to 1.16 (Install templates as pip package) by @christian-byrne in #7623

New Contributors

• @lerignoux made their first contribution in #7439
• @Chargeuk made their first contribution in #7509
• @bebebe666 made their first contribution in #7584

Full Changelog: v0.3.27...v0.3.29