Skip to content
trendoviz

ComfyUI v0.42.0–v0.46.0 widens nodes for image, video, audio

ComfyUI releases v0.42.0 through v0.46.0 add or extend nodes for image, video, audio, voice, text and 3D work, including MOSS-image support in v0.42.0 and a Video Deflicker node in v0.46.0. The sequence also introduces a local-versus-hosted split in which some model families run as local support and others are reached through partner nodes, while removing DirectML support in v0.46.0.

55 out of 100Emerging7D: Insufficient history ProvisionalPublished

Overview

ComfyUI is a graph-based interface with a GUI, API and backend for diffusion-model work [1]. The project describes itself as "the most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface" and claims it is "the fastest local inference engine in the world" — a vendor claim that appears on each release page and is not supported by an independent benchmark in the release records [1][2][3][4][5][6].

Across five version tags, the release notes show the same pattern in each: a mixture of partner nodes that reach hosted services, local model support, workflow and template maintenance, and memory or precision handling. The notes carry no release dates, so the interval between versions is unknown, and no adoption or download figures appear [1][3][4][5][6].

Why it matters

For people assembling visual-media workflows, the release sequence widens the set of operations reachable inside one graph: image generation and editing, video upscaling and cleanup, voice and music generation, text generation, and 3D and mesh handling [3][4][5][6]. A workflow can now combine, for example, MOSS-image support [1], HeyGen voice cloning and text-to-speech [5], Flux Video Upscale V2 at 2K and 4K targets [4], and ProRes with selectable chroma subsampling for video saving [4].

The local/hosted split is the other structural change. Model work that runs in the user's own environment includes MOSS-image, FIBO 2.0, MiniMax-H3 and Qwen 3.x support [1][3][4][5], while partner nodes reach external services [1][3][4][5][6]. Reading the notes together, this appears to be a deliberate separation between models executed locally and models reached through hosted APIs; the project does not state such a strategy in these records, so that reading remains an inference [5][6].

Opportunities

Reusable media pipelines. Because workflows are graphs of connected nodes and templates, a maker can assemble a repeatable sequence rather than a one-off prompt — for example, generate an image, upscale it, remove a background, convert to video, and save with a chosen codec and chroma subsampling using nodes added across v0.44.0 and v0.45.0 [3][4].

Audio and voice inside the same graph. HeyGen Voice instant clone and text-to-speech arrived in v0.43.0, Pika Soundtrack, SFX and Music nodes and Breeze TTS 2 nodes in v0.45.0, and HeyGen Video 1.0 gained 1080p, 30-second clips and auto duration in v0.46.0 [3][5][6]. The notes do not confirm which of these can be wired directly into a video node in one graph, so the practical integration path is unknown.

Video finishing passes. Video Deflicker with a faster RAFT optical flow implementation in v0.46.0, 2K/4K targets for Flux Video Upscale V2 in v0.44.0, and the Crystal Video Upscaler node in v0.43.0 form a set of post-processing capabilities [4][5][6].

Mixed local and hosted workflows. Local model support includes MOSS-image, FIBO 2.0, MiniMax-H3 and Qwen 3.x, while partner nodes expose hosted services such as Krea 3 Image, Kling 4.0, HeyGen and GPT Image [1][3][4][5]. Moving node families to Comfy Router in v0.45.0 and v0.46.0 extends that hosted path [3][6].

Lower-memory operation. Dynamic VRAM KV caches, adaptive SeedVR2 VAE slicing sized to free memory, a VAE compute dtype node, and a fix for fp8_e4m3fn_fast falling back to bf16 under dynamic VRAM all target memory or precision handling on local hardware [3][4][5]. The effect on a specific setup is not established by the notes.

Documented integration conventions. Model integration instructions were added to AGENTS.md in v0.43.0 and AGENTS.md was updated with more specific instructions in v0.46.0, giving contributors and studios maintaining their own wrappers a maintained reference for documenting integrations [5][6].

Risks

Reliability of newly added paths. v0.42.1 shipped a single stated change: fixing crashes that some users experienced running the MOSS image template workflow [2]. A later release also carried a fix for some workflows not working properly [6]. Users adopting a workflow on release day may encounter this class of issue.

Hardware and platform discontinuities. v0.46.0 removes DirectML support and adds a warning when a GPU supports CUDA 13 but the user is not running it [6]. The notes give no detail on which configurations are affected, how many users are affected, or what replaces DirectML.

Cost exposure from partner nodes. The release notes list many hosted-service nodes — HeyGen voice and video, Kling, Topaz, Pika, FishAudio — and include occasional price-badge updates, including one for retired Grok image models and one showing the full price for FishAudio Drama 3 [3][5][6]. Pricing, API contract details, rate limits and availability terms are not given, so a per-run cost cannot be inferred from a workflow.

Third-party dependencies in the execution path. Partner-node sign-in token renewal was added in v0.43.0, and several node families moved to Comfy Router in v0.45.0 and v0.46.0 [3][5][6]. A provider's authentication, routing and uptime become operational dependencies for workflows that use them.

Model and provider churn. The notes remove retired Arrow 1.x models in v0.42.0, deprecate Seedance 1.x video nodes in v0.44.0, and remove the retired Claude Sonnet 4.5 model and Qwen3.6 Max Preview model in v0.46.0 [1][4][6]. A workflow pinned to one provider's model can break without notice.

Licensing and commercial terms are unaddressed. The notes contain model names and node lists only. License terms of the underlying models, training-data status, and the commercial rights attached to partner-service outputs are not covered, so paid work requires separate review for each model and service involved.

Ecosystem-specific structure. Graph structure, workflow templates and model integration conventions are specific to this project; moving a workflow elsewhere means rebuilding the graph and any provider-specific logic [5][6].

Context

The node graph is the organizing idea: workflows are built from connected nodes rather than run as single prompts, and templates are versioned separately from the application (template updates appear in every release from v0.11.85 through v0.11.111) [1][3][4][5][6]. Some additions extend what a graph can do locally, such as loading Qwen3.5 embedding models, rope scaling and color-depth control for MOSS-image, context windows for MiniMax-H3, and a VAE compute dtype node [1][4][5]. Others add access to hosted services, marked in the notes as "Partner Nodes," including Krea 3 Image, HeyGen voice cloning, Kling 4.0 and GPT Image V3/GIF/SVG [1][3][5].

Several releases also move existing node families to Comfy Router: Qwen Image in v0.45.0, then Wan, Krea and Luma in v0.46.0, alongside a general move of more partner nodes to the router [3][6]. Partner-node sign-in token renewal was added in v0.43.0 [5].

Known, claimed, inferred

Known. The release notes are official project records, and each version's contents are directly stated: v0.42.0 at commit 355cd6b [1], v0.42.1 at 63be2f9 [2], v0.43.0 at 426f5ee [5], v0.44.0 at 2c75dc1 [4], v0.45.0 at 989d93d [3], and v0.46.0 at a277b5b, marked "Latest" [6]. v0.42.1 lists only one change, and v0.46.0 removes DirectML support [2][6].

Claimed. The project describes ComfyUI as a modular diffusion-model GUI, API and backend with a graph/nodes interface, and states it is the fastest local inference engine in the world [1]. The same tagline appears on the other release pages [3][4][5].

Inferred. The addition of partner nodes, the relocation of several node families to Comfy Router, and the addition of model integration instructions to AGENTS.md indicate a working distinction between models run locally and models reached through hosted APIs [5][6]. Separately, the breadth of node types across these versions — image, video, audio, voice, text, 3D and mesh — suggests ComfyUI is functioning as a general media-processing front end rather than a single-model tool [3][4][6]. Both readings follow from the listed changes rather than from explicit project positioning.

Limitations & uncertainties

The evidence contains no adoption, download, install or usage figures for any version, and no information about how many users run the affected workflows or on which hardware [1][3][4][5][6].

The v0.42.0 notes list Qwen 3.x performance work without quantifying it, and later parts of that series across v0.43.0 to v0.45.0 are also unquantified [1][3][5]. The size of the gain, and the hardware it was measured on, are unknown.

Release dates are absent from every record, so neither the cadence between versions nor the interval from v0.42.0 to v0.46.0 can be established [1][2][3][4][5][6].

Breaking changes, migration requirements and known limitations for an upgrade from v0.41.0 to v0.42.0 are not documented beyond the v0.42.1 crash fix [2]. The full scope of changes in v0.42.1 is also unknown because its notes list only that one fix.

The "MOSS-image," "MOSS-Image" and "MOSS" labels appear across releases, and the evidence does not define their relationship or confirm that they refer to one model family [1][5].

All records come from official GitHub release pages for a single project. Records [3], [4], [5], [1] and [6] are truncated duplicates of the fuller entries and add no independent detail. No item is an independent third-party test report or benchmark, so there is no published performance result to report for any of these releases.

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementComfyUI v0.42.0 added MOSS-image support, a Krea 3 Image partner node, further Qwen 3.x performance work, workflow template updates, better documentation for the comfy quant format, and ARM Linux coverage in execution and unit tests.[1]
  • Primary-source or vendor statementv0.42.1 contains a single stated change: fixing crashes that some users experienced when running the MOSS image template workflow.[2]
  • Primary-source or vendor statementv0.43.0 through v0.46.0 broaden node coverage across image, video, audio, voice, text and 3D work, including HeyGen voice cloning, Crystal Video Upscaler, MiniMax-H3 support with context windows and a latent upscaler, ProRes and chroma subsampling for video saving, a Video Deflicker node, a mesh input for Create Camera Info, and support for Qwen3-1.7B in Generate Text.[3][4][5][6]
  • Primary-source or vendor statementSeveral node families were moved to Comfy Router across v0.45.0 and v0.46.0, including Qwen Image, Wan, Krea and Luma.[3][6]
  • Primary-source or vendor statementv0.46.0 removes DirectML support and adds a warning when a GPU supports CUDA 13 but the user is not running it.[6]
  • HypothesisThe addition of partner nodes, relocation of node families to Comfy Router, and model integration instructions in AGENTS.md indicate a working separation between models run locally and models reached through hosted APIs; the project does not state this strategy in the release notes.[3][5][6]
  • Primary-source or vendor statementThe project describes ComfyUI as a modular diffusion-model GUI, API and backend with a graph/nodes interface and claims it is the fastest local inference engine in the world.[1]

Sources

  1. 1
    Comfy-Org/ComfyUI v0.42.0
    ●github.comretrieved Nov 18, 2026
  2. 2
    Comfy-Org/ComfyUI v0.42.1
    ●github.comretrieved Nov 18, 2026
  3. 3
    Comfy-Org/ComfyUI v0.45.0
    ●github.comretrieved Nov 18, 2026
  4. 4
    Comfy-Org/ComfyUI v0.44.0
    ●github.comretrieved Nov 18, 2026
  5. 5
    Comfy-Org/ComfyUI v0.43.0
    ●github.comretrieved Nov 18, 2026
  6. 6
    Comfy-Org/ComfyUI v0.46.0
    ●github.comretrieved Nov 19, 2026

Updates & corrections

  1. First published · approved by an editor

Found an error? Report it.