ComfyUI v0.42: MOSS‑Image Support, Quiver Arrow 1.x Removed

ComfyUI v0.42.0: MOSS-Image, Qwen Updates and Removals #

MOSS-Image text and layout-guided generation #

The ComfyUI v0.42.0 changelog lists native support for MOSS-Image for text to image and layout-guided generation, with strong bilingual text rendering. A second entry adds fused situ attention for MOSS-Image and layout prompts driven by bounding boxes. A day later, v0.42.1 fixed crashes that some users hit when running the MOSS image template workflow, so MOSS users should skip straight to a later 0.42 release.

Quiver Arrow 1.x removal #

The same release lists the retired Arrow 1.x models removed from the Quiver nodes. Any workflow that still references those models will no longer have them available in the Quiver nodes. Search your saved workflows for Quiver nodes and update each one to a current model before the next run. This is a breaking change for those workflows, so schedule the update rather than discovering it during a production job.

Qwen 3.x, Qwen3.5 and PDD checkpoint support #

For text generation, the notes list faster text generation for the Qwen 3.x family. Qwen3.5 embedding models can now be loaded, and models that ship with PDD baked into their weights are supported. The GitHub notes also list a fix for Qwen think-block templates in TextGenerate (#17121).

Runtime and frontend fixes #

Smaller items in the v0.42.0 notes are grouped here. They include a refactor of 3D model remeshing (#16805) and a fix for a crash with solid mode and an empty mesh (#17140), frontend 1.55.18, ARM Linux execution (#17139), a webcam node fix (#17144) and a text-generation cleanup fix (#17169). The GitHub release notes for v0.42.0 describe the earlier behaviour: the webcam node was no longer working, and text generation was not properly cleaning itself up. They also state that the CPU device does not support fp8, nvfp4, mxfp8 or int8 (#17170), which matters for anyone running quantized models on CPU.

ComfyUI v0.42.4: Partner Video, Audio and Image Nodes #

VEED and Mirelo partner nodes #

The ComfyUI v0.42.4 changelog adds a set of partner nodes. VEED Fabric 1.0 turns a still photo and a speech track into a talking, lip-synced video. VEED Lipsync 2.0 re-syncs a speaker's mouth in existing footage to a new speech track. VEED Remove Video Background and Clean Audio provides alpha-matte subject cutouts and speech cleanup, and Mirelo SFX 2.0 generates sound effects from a text prompt or timed to video action. Teams building media pipelines should test these nodes with their own footage before relying on them.

Krea 3 Image ultra thinking mode #

The notes add an ultra thinking mode to the Krea 3 Image node. The entry describes it as:

A slower, higher-quality mode that generates several candidates and returns the best one

ComfyUI v0.42.4 changelog

Because this mode generates several candidates, it takes longer than the standard mode. Expect higher run times in batch jobs that use it, and budget for that before enabling it in automated workflows. The notes do not state how many candidates are generated or how the returned result is chosen.

HeyGen Voice 1.0 price badge and template update #

A bug fix in v0.42.4 means that the HeyGen Voice 1.0 node now shows the current price. Users who run this node should check the price badge before executing paid jobs. The same release also updates Templates to v0.11.94. The notes list no further detail for the template update, so review your templates after upgrading.

1. Record your current ComfyUI version and list the workflows you run.
2. Search those workflows for Quiver nodes and replace any Arrow 1.x model references, since v0.42.0 removes them.
3. Upgrade to v0.42.1 or later rather than v0.42.0, since v0.42.1 fixes crashes in the MOSS image template workflow.
4. Re-run 3D remeshing jobs that crashed in solid mode on an empty mesh, and webcam workflows that stopped working.
5. Upgrade to v0.42.4 and check the HeyGen Voice 1.0 price badge before running that node.
6. Test the new VEED, Mirelo and Krea 3 ultra thinking mode nodes on sample inputs before adding them to production.