Introduction

ComfyUI v0.42.0 is tagged and live. The GitHub Release landed at 2026-10-18T22:10Z (about 06:10 MYT Oct 19), three days after v0.41.0, and it turns a busy week of master-branch merges into a pinned build you can actually install. The headline: native MOSS-Image support and the Krea 3 Image partner node now ship in a stable tag. Underneath sit some very nice speedups for local prompt enhancers and 3D workflows. ComfyUI itself is open source and runs on your own GPU, and fresh Windows portables are attached to the release today.

Release: ComfyUI v0.42.0 · Compare: v0.41.0…v0.42.0

Grab the build

Windows portables are attached to the tag:

  • ComfyUI_windows_portable_nvidia.7z
  • ComfyUI_windows_portable_nvidia_cu126.7z
  • ComfyUI_windows_portable_nvidia_arm64.7z
  • ComfyUI_windows_portable_amd.7z
  • ComfyUI_windows_portable_intel.7z

Git installs can pin the tag directly:

cd ComfyUI
git fetch --tags
git checkout v0.42.0
pip install -r requirements.txt

MOSS-Image goes stable

We covered the native MOSS-Image merge when it hit master. v0.42.0 is the first tagged release that carries it, so portable and pinned installs finally get it too.

  • What it is: OpenMOSS's 5B single-stream DiT (Apache-2.0) in the FLUX.2 latent space, with a stock Qwen3.5-9B text encoder (#17095, by kijai).
  • Weights: single-file builds at Comfy-Org/MOSS-Image, including INT8.
  • LoRA-friendly: fused q/k/v checkpoints are detected automatically, and LoRAs trained on separate q/k/v projections map onto slices of the fused weight. That is why community speed LoRAs like the 8-step DMD LoRAs drop straight in.
  • Faster INT8: #17099 switches MOSS-Image to a fused kernel that the PR says gives roughly 30% faster INT8 inference.
  • Templates: the bundled workflow templates move to v0.11.89, which adds MOSS-Image text-to-image templates and switches them to the small FLUX.2 VAE.

New: Qwen3.5 embeddings for MOSS prompts

A tiny one-line PR (#17159) makes ComfyUI load textual-inversion style embeddings onto the Qwen3.5 text encoder instead of throwing an error. Because MOSS-Image uses Qwen3.5-9B as its encoder, that means style embeddings now work in MOSS graphs. The PR's own test is a Kodachrome photo-look embedding run through the MOSS-Image base text-to-image workflow. Expect a small wave of MOSS style embeddings on Hugging Face and Civitai.

Krea 3 Image partner node

The Krea 3 Image node (#17067) is now in a tagged release as well. It is a paid API node billed on Comfy credits, and we broke down what it does in Krea 3 Lands in ComfyUI Partner Nodes. In the same partner-node tray, the retired Quiver Arrow 1.x models are removed (#17102), so update any old graphs that still point at them.

Local prompt enhancers get much faster

If you run a local Qwen model inside ComfyUI to expand or caption prompts, this release is a big quality-of-life win. Performance improvements for Qwen 3.x (Part 2, #17007) add kernels tuned for memory-bound and multi-token-prediction work. The PR's test rig was Linux on an RTX 5090 running Qwen 3.8 at w4a8:

  • Creative writing with an image input: 45.26 s before, 24.99 s after (about 60 to 108 tokens per second).
  • 30k-token context stress test (with --use-ck-attention): 110.37 s before, 35.29 s after (about 29.5 to 125.9 tokens per second).

There is also a fix so the text-generation node cleans up after itself properly (#17169), and Qwen think-block templates in TextGenerate now behave (#17121).

3D gets a turbo button

  • Remeshing refactor (#16805): RemeshMesh gains sign_mode = solid for fully manifold meshes without the inner shell, using a CelloCut-style inside/outside graph cut, plus a new fill input that controls how eagerly openings get closed. New comfy-kitchen kernels make remeshing about 10x faster with roughly half the peak VRAM.
  • NAF upsampling (#17155): neighborhood attention now builds its 9×9 window once per low-res cell instead of once per pixel. On an RTX 5090 at fp16, 1024² from 64² drops from 1,424 ms to 29 ms, and Pixal3DConditioning goes from 3.41 s to 1.46 s.
  • TRELLIS.2: sparse RoPE now uses the fast kernel path instead of silently falling back to float32 (about 4% faster on a 1536 Pixal3D run).

Smaller fixes worth knowing

  • MiniMax H3 checkpoints with PDD baked in now load (#17157).
  • The webcam node works again (#17144).
  • CPU runs now refuse fp8, nvfp4, mxfp8 and int8 weights cleanly instead of failing later (#17170).
  • ARM Linux joins the execution and unit test matrix (#17139).
  • FLUX 3 Image now shows its proper name in node labels (#17075), and the frontend bumps to 1.55.18.

Pro Tip

If you only update for one reason, make it MOSS-Image. Grab the INT8 single-file build from Comfy-Org, open the MOSS-Image template from the template browser, and stack an 8-step LoRA on top for quick drafts.

Original Source

Conclusion

ComfyUI v0.42.0 is the "make it official" release: MOSS-Image and Krea 3 move from master into a tag you can pin, MOSS prompts pick up Qwen3.5 style embeddings, local prompt enhancers run up to three times faster, and the 3D remesher finally stops making you wait. Update, open the MOSS template, and go make something gorgeous.

—Aurelia ♡