Full Changelog: v0.18.1...v0.18.2

What's Changed

• Fix canny node not working with fp16. by @comfyanonymous in #13085
• Fix sampling issue with fp16 intermediates. by @comfyanonymous in #13099
• Fix fp16 intermediates giving different results. by @comfyanonymous in #13100
• wan: vae: Fix light/color issue by @rattus128 in #13101

Full Changelog: v0.18.0...v0.18.1

What's Changed

• Revert "Revert "feat: Add CacheProvider API for external distributed caching"" by @deepme987 in #12915
• fix(api-nodes): Tencent TextToModel and ImageToModel nodes by @bigcat88 in #12680
• Bump comfyui-frontend-package to 1.41.19 by @comfy-pr-bot in #12923
• bump manager version to 4.1b4 by @ltdrdata in #12930
• comfy aimdo 0.2.11 + Improved RAM Pressure release strategies - Windows speedups by @rattus128 in #12925
• Update README. by @comfyanonymous in #12931
• comfy-aimdo 0.2.12 by @rattus128 in #12941
• fix: use no-store cache headers to prevent stale frontend chunks by @christian-byrne in #12911
• feat: Support mxfp8 by @kijai in #12907
• Add --fp16-intermediates to use fp16 for intermediate values between … by @comfyanonymous in #12953
• Update comfyui-frontend-package to version 1.41.20 by @comfyanonymous in #12954
• LTXV: Accumulate VAE decode results on intermediate_device by @kijai in #12955
• bump manager version to 4.1b5 by @ltdrdata in #12957
• ops: opt out of deferred weight init if subclassed by @rattus128 in #12967
• Make EmptyLatentImage follow intermediate dtype. by @comfyanonymous in #12974
• Enable Pytorch Attention for AMD gfx1150 (Strix Point) by @lostdisc in #12973
• feat: add essentials_category to nodes and blueprints for Essentials tab by @christian-byrne in #12573
• feat(assets): align local API with cloud spec by @luke-mino-altherr in #12863
• ci: add check to block AI agent Co-authored-by trailers in PRs by @christian-byrne in #12799
• Skip running model finalizers at exit by @blepping in #12994
• Add --enable-dynamic-vram options to force enable it. by @comfyanonymous in #13002
• [Trainer] FP4, 8, 16 training by native dtype support and quant linear autograd function by @KohakuBlueleaf in #12681
• Fix potential issue. by @comfyanonymous in #13009
• fix: atomic writes for userdata to prevent data loss on crash by @christian-byrne in #12987
• Disable SageAttention for Hunyuan3D v2.1 DiT by @paulomuggler in #12772
• Update workflow templates to v0.9.26 by @comfyui-wiki in #13012
• Mark weight_dtype as advanced input in Load Diffusion Model node by @christian-byrne in #12769
• Reduce LTX VAE VRAM usage and save use cases from OOMs/Tiler by @rattus128 in #13013
• Reduce WAN VAE VRAM, Save use cases for OOM/Tiler by @rattus128 in #13014
• bump manager version to 4.1b6 by @ltdrdata in #13022
• Inplace VAE output processing to reduce peak RAM consumption. by @kijai in #13028
• Fix case where pixel space VAE could cause issues. by @comfyanonymous in #13030
• cascade: remove dead weight init code by @rattus128 in #13026
• fix: run text encoders on MPS GPU instead of CPU for Apple Silicon by @k06a in #12809
• fix(api-nodes): add support for "thought_image" in Nano Banana 2 by @bigcat88 in #13038
• Update comfyui-frontend-package version to 1.41.21 by @DrJKL in #13035
• Make more intermediate values follow the intermediate dtype. by @comfyanonymous in #13051
• Further Reduce LTX VAE decode peak RAM usage by @kijai in #13052
• Fix regression. by @comfyanonymous in #13053
• fp16 intermediates doen't work for some text enc models. by @comfyanonymous in #13056
• ltx: vae: implement chunked encoder + CPU IO chunking (Big VRAM reductions) by @rattus128 in #13062
• memory: Add more exclusion criteria to pinned read (fixes corrupt outputs - rare cases) by @rattus128 in #13067
• Reduce tiled decode peak memory by @kijai in #13050
• Revert "fix: run text encoders on MPS GPU instead of CPU for Apple Silicon" by @comfyanonymous in #13070
• Fix VRAM leak in tiler fallback in video VAEs by @rattus128 in #13073
• ltx: vae: Fix missing init variable by @rattus128 in #13074
• [API Nodes] mark seedream-3-0-t2i and seedance-1-0-lite models as deprecated by @bigcat88 in #13060
• Add slice_cond and per-model context window cond resizing by @drozbay in #12645
• feat(api-nodes): add Quiver SVG nodes by @bigcat88 in #13047
• Make EmptyImage node follow intermediate device/dtype. by @comfyanonymous in #13079
• Move inline comfy.context_windows imports to top-level in model_base.py by @Kosinkadink in #13083

New Contributors

• @lostdisc made their first contribution in #12973
• @paulomuggler made their first contribution in #12772
• @k06a made their first contribution in #12809

Full Changelog: v0.17.0...v0.18.0

Full Changelog: v0.17.1...v0.17.2

Full Changelog: v0.17.0...v0.17.1

What's Changed

• refactor(assets): modular architecture + async two-phase scanner & background seeder by @luke-mino-altherr in #12621
• comfy-aimdo 0.2.9 by @rattus128 in #12840
• Import assets seeder later, print some package versions. by @comfyanonymous in #12841
• Fix text encoder lora loading for wrapped models by @kijai in #12852
• Update workflow templates to v0.9.18 by @comfyui-wiki in #12857
• fix(manager): improve install guidance when comfyui-manager is not installed and bump manager version to v4.1b2 by @ltdrdata in #12810
• Add a way for nodes to add pre attn patches to flux model. by @comfyanonymous in #12861
• model_detection: deep clone pre edited weights by @rattus128 in #12862
• Widen OOM_EXCEPTION to AcceleratorError form by @rattus128 in #12835
• main: switch on python faulthandler by @rattus128 in #12868
• Allow model patches to have a cleanup function. by @comfyanonymous in #12878
• Add pre attention and post input patches to qwen image model. by @comfyanonymous in #12879
• comfy-aimdo 0.2.10 by @rattus128 in #12890
• feat(api-nodes): add Reve Image nodes by @bigcat88 in #12848
• fix: guard torch.AcceleratorError for compatibility with torch < 2.8.0 by @adiborochov in #12874
• Fix issue with batch_size > 1 on some models. by @comfyanonymous in #12892
• Bump comfyui-frontend-package to 1.41.15 by @comfy-pr-bot in #12891
• Bump comfyui-frontend-package to 1.41.16 by @comfy-pr-bot in #12894
• Bump comfy-kitchen version to 0.2.8 by @comfyanonymous in #12895
• Support flux 2 klein kv cache model: Use the FluxKVCache node. by @comfyanonymous in #12905
• add painter node by @jtydhr88 in #12294
• fix: use frontend-compatible format for Float gradient_stops by @jtydhr88 in #12789
• Update workflow templates to v0.9.21 by @comfyui-wiki in #12908
• Lower kv cache memory usage. by @comfyanonymous in #12909
• Patch frontend to 1.41.18 (from 1.41.16) by @christian-byrne in #12910
• feat: Add CacheProvider API for external distributed caching by @deepme987 in #12056
• Revert "feat: Add CacheProvider API for external distributed caching" by @comfyanonymous in #12912
• Fix audio extraction and truncation bugs by @PxTicks in #12652

New Contributors

• @luke-mino-altherr made their first contribution in #12621
• @adiborochov made their first contribution in #12874
• @deepme987 made their first contribution in #12056
• @PxTicks made their first contribution in #12652

Full Changelog: v0.16.4...v0.17.0

What's Changed

• feat: add Math Expression node with simpleeval evaluation by @dante01yoon in #12687
• feat(api-nodes): add TencentSmartTopology node by @bigcat88 in #12741
• feat(api-nodes): add Gemini 3.1 Flash-Lite model to LLM node by @bigcat88 in #12803
• Fix fp16 audio encoder models by @rattus128 in #12811
• Don't run TE on cpu when dynamic vram enabled. by @comfyanonymous in #12815
• Fix requirements version. by @comfyanonymous in #12817
• Update workflow templates to v0.9.11 by @comfyui-wiki in #12821
• mm: reset_cast_buffers: sync compute stream before free by @rattus128 in #12822

New Contributors

• @dante01yoon made their first contribution in #12687

Full Changelog: v0.16.3...v0.16.4

What's Changed

• Add manual cast to LTX2 vocoder conv_transpose1d by @kijai in #12795
• LTX audio vae novram fixes. by @comfyanonymous in #12796
• feat: add causal_fix parameter to add_keyframe_index and append_keyframe by @tavihalperin in #12797

Full Changelog: v0.16.2...v0.16.3

What's Changed

• comfy-aimdo 0.2.7 by @rattus128 in #12791
• ops: Handle CPU weight in VBAR caster by @rattus128 in #12792
• Update workflow templates to v0.9.10 by @comfyui-wiki in #12793
• Refactor. by @comfyanonymous in #12794

Full Changelog: v0.16.1...v0.16.2

What's Changed

• feat(api-nodes-xAI): updated models and pricing by @bigcat88 in #12756
• feat(api-nodes): enable Kling 3.0 Motion Control by @bigcat88 in #12785
• Update workflow templates to v0.9.8 by @comfyui-wiki in #12788

Full Changelog: v0.16.0...v0.16.1