ComfyUI v0.44.0, released on November 4, 2026, adds a more practical set of video and audio controls than its version number alone suggests. The release includes six MiniMax H3 Max editing nodes, Sonilo Music 1.5, ProRes export and explicit chroma-subsampling choices. It also adds a branch-specific VAE compute-dtype control and support for the Sol-H3-VAE decoder. These changes address different stages of production: revising a shot, creating its soundtrack, decoding local video latents and choosing what leaves ComfyUI for an editor.
The official release record dates publication to 01:12 UTC, or 08:12 WIB, on November 4. Its changelog compares v0.43.0 with v0.44.0, so it should not be read as proof that every listed feature first appeared that morning. Some functionality was already available through development builds or intervening versions. The news is the tagged release that collects these changes, not a claim that ComfyUI has just invented video editing or music generation.
There is also a boundary readers should establish before trying the new tools. MiniMax H3 Max editing and Sonilo Music 1.5 are hosted Partner Nodes. ProRes export, VAE dtype selection and local H3 model support belong to the local software path. A node appearing on the same canvas does not make its processing local, free or available offline.
MiniMax H3 Max adds six specific editing operations
The MiniMax integration change adds Recast, Extend Video, Reframe Video, Insert Video, Lip Sync and Camera Control. The tagged node implementation provides a useful distinction between them. Some revise existing footage; others generate a new shot from an image. Treating all six as interchangeable video-to-video filters would lead to the wrong inputs and expectations.
MiniMax H3 Max Recast takes a source clip and one to four reference photos. Its description says it replaces people while keeping the source motion, camera, cuts and sound. The node validates a source length of five to 30 seconds and reference images of at least 256 by 256 pixels. Its tooltip also requires each shot to be 15 seconds or shorter. The exposed resolution choices are 768P and 1080P. Those choices are more reliable than broader resolution claims in automatically generated template descriptions.
For a producer, the useful question is whether an existing performance can survive a cast change without rebuilding the whole shot. That is an intended use, not a verified result from this article. Facial identity, hands, clothing boundaries and continuity still need review in the returned footage. Reference photos also introduce a consent question that a technically valid workflow cannot answer for the operator.
Reframe Video addresses a different problem: changing the canvas around a finished clip. The implementation describes fitting the source onto a new canvas and generating surrounding areas while retaining timing and sound. It accepts clips from half a second to 300 seconds and offers landscape, portrait, square and wider aspect ratios. This is generative outpainting, not evidence that a crop has recovered real information outside the original camera view.
The published reframe template connects Load Video to the MiniMax reframe node and then Save Video. Its example prompt asks the model to continue the existing background on both sides and keep the pilot centered. That is a more useful starting point than an unrelated style prompt: it defines what the added area should contain and what should remain stable. The template demonstrates the intended graph, but downloading its JSON does not establish that the hosted service has executed successfully on a reader’s account.
Extension and insertion change the timeline differently
Extend Video accepts an existing clip and a prompt describing what happens next, with optional audio references. Its controls distinguish an extended output containing the whole clip from a continuation containing the new footage. The continuation option starts with a 1.6-second repeat of the source ending, according to the node tooltip. An editor assembling multiple continuations should account for that overlap rather than assume every returned frame is new material.
The implementation also rejects a full extended output at 2K when source duration plus requested extension exceeds 45 seconds. That is a concrete compatibility limit, not a general statement that all MiniMax H3 clips are capped at that length. A lower resolution or continuation-only output follows a different route. Reading the error in relation to the selected operation is more useful than repeatedly changing the prompt.
Insert Video replaces the source interval between start and resume times with a generated scene. It can use reference images and videos, and its new-scene duration is a separate setting. The removed interval and replacement scene therefore need not have the same length. Anyone relying on fixed downstream timecodes should check the returned edit before placing subtitles, music cues or other synchronized material.
Lip Sync starts from an image and an audio track, rather than requiring an existing speaking video. Camera Control also starts from an image, with azimuth, elevation and relative camera distance controls, plus an optional camera-info input. Its tooltip explicitly calls the requested orbit a target: a 90-degree request does not guarantee an exact 90-degree result. These are generative controls, not a substitute for a deterministic camera rig.
Sonilo Music 1.5 adds a lyrics-aware audio branch
The new Sonilo Music 1.5 integration generates songs with sung vocals from text. The node supports a prompt describing genre, mood, instruments and vocals. A Lyrics: block with verse and chorus tags supplies the words; without that block, the model writes lyrics itself. The tagged implementation limits the prompt to 2,000 characters.
Duration has an automatic mode and a custom mode. Automatic length follows the lyrics; the custom control exposes ten to 360 seconds. The tooltip warns that a requested length can cut lyrics that do not fit or repeat short lyrics. This makes duration an arrangement decision as well as a file-length setting. A chorus written for a longer song does not become a complete short cue simply because a smaller duration was selected.
A seed field should not create a false sense of reproducibility. The Sonilo node source says its seed controls whether the node reruns, while results remain nondeterministic. Saving a graph and seed is still useful for recording intent, but it does not guarantee that a later request will reproduce an approved vocal take. Keep the actual accepted audio file alongside the workflow.
The release also lists VEED Dubbing, which translates speech, clones speakers’ voices and synchronizes lips to the translation according to its node description. Its standard and pro options differ in language choices and audio handling; the pro description includes regional accents and stereo audio. This serves localization rather than songwriting. The source rejects video without an audio track, so a silent generated shot is not ready for dubbing merely because its video socket connects.
ProRes and chroma controls change the export decision
The ProRes export change affects the local handoff from generated material to video files. In v0.44.0, Save Video exposes ProRes among its codec choices and supports a MOV container route. With the format set to auto and the codec set to ProRes, the implementation chooses MOV. MP4 does not list ProRes as one of its available codecs in that node.
Create Video separately exposes bit depth, color space and chroma subsampling. The subsampling menu offers auto, 444, 422 and 420. Its tooltip explains that 444 keeps full color resolution, 422 halves horizontal color resolution, and 420 halves it on both axes. This gives users an explicit choice about color sampling rather than leaving every export decision to an implicit default.
That choice still has to match the encoder. The node warns that unsupported combinations use the closest supported mode and print a console warning. A follow-up change in the same release validates chroma input. Selecting a menu value therefore is not sufficient evidence of the final file’s properties. For a delivery with prescribed specifications, inspect the encoded file and test it in the receiving application.
Bit depth and color space are independent controls. The automatic bit-depth selection uses eight bits for sRGB and ten for HDR modes, while explicit eight-bit and ten-bit choices are independent of color space. A ten-bit setting alone should not be described as proof that an ordinary source has gained HDR detail. The controls describe how ComfyUI interprets and encodes its inputs; they do not prove the quality or provenance of those inputs.
The practical benefit is a more deliberate export stage. A quick review file and a file intended for further editing can now take different routes inside the graph. This article did not encode a ProRes sample, validate a particular editor or measure file sizes. The verified claim is that the tagged implementation exposes these controls and defines container and fallback behavior.
VAE dtype and Sol-H3-VAE need separate evaluation
The new VAE Compute Dtype node offers default, fp32, fp16 and bf16. Its description says it changes VAE encoding and decoding without affecting other branches. That scope matters in a graph where the same loaded VAE feeds multiple paths: the control is intended to change the selected branch rather than silently redefine the entire workflow’s numerical behavior.
The release includes a related fix for a crash when that control selects fp32 on the LTX 2.5 diffusion VAE decoder. The fix description traces the failure to linear-layer weights remaining bf16 while the latent became fp32. It changes the affected layers and casts a modulation parameter. This is a compatibility correction, not evidence that fp32 is always the best precision setting or that all VAE models support every dtype equally well.
Sol-H3-VAE is a separate decoder implementation, not another name for that dtype selector. Its model card describes a 24-layer distilled video decoder replacing the H3 decoder while leaving the audio decoder unchanged. The upstream ComfyUI integration says it supports the released decoder-only file or a file merged with the H3 encoder.
The most consequential limit is whole-clip decoding. The contributor says the decoder cannot tile, so invoking VAEDecodeTiled still decodes the whole clip. The model card likewise describes noncausal whole-clip behavior and memory growth with clip length. A workflow that previously depended on tiling to fit a long clip should not assume the same recovery option remains available after swapping decoders.
Both the model card and integration report performance measurements, but their engines, precision modes and test conditions differ. They are source-reported measurements, not interchangeable promises for every ComfyUI installation. The decoder is distilled and is not identical to the original. Evaluate visual differences and memory on the actual clip before choosing it for an approved render; no local decoder benchmark was performed for this article.
Upscaling now has several meanings on the same canvas
The release adds a local MiniMax H3 latent-upscaler path and basic H3 context-window support. Those changes concern model execution and latent processing. They should be distinguished from hosted pixel-video enhancement: the release also adds Pruna P-Video Upscale and new named output targets for Flux Video Upscale V2. A shared word such as upscale does not establish that these tools accept the same representation or preserve the same information.
Pruna’s node offers 720p, 1080p, 2K and 4K targets, plus source or selected frame rate. Its frame limits vary by target, and its implementation checks clip duration against those limits before upload. Flux’s named targets fit within specified bounding boxes and remain subject to scaling limits. A 4K label consequently should not be presented as a guarantee that any arbitrary small source will return at a particular full-frame dimension.
For existing users, the useful adoption order follows the job rather than the changelog. Choose the operation that fixes a specific shot or audio problem, confirm whether media leaves the machine, and then check the decoded and encoded output separately. Comfy’s Partner Nodes documentation requires account access and credits for hosted calls and describes those nodes as optional. None of those services was invoked for this analysis.
ComfyUI v0.44.0 is worth attention because it exposes decisions that previously required separate tools or less visible defaults. The trade-offs remain explicit: hosted editing requires uploads, generative controls need output review, whole-clip decoding changes memory planning, and export settings require file-level verification. The release record establishes fresh upstream activity on November 4; it does not establish viral popularity or measured improvements on a reader’s system.
Sources
- ComfyUI v0.44.0 official release and linked changes
- Tagged MiniMax node definitions and input validation
- Tagged Sonilo Music 1.5 implementation
- Tagged Create Video and Save Video controls
- Sol-H3-VAE model card and limitations
- Official ComfyUI workflow templates
- Partner Nodes requirements and local-versus-hosted distinction
Tags: ComfyUI, MiniMax H3, Sonilo, AI video, AI audio, ProRes, VAE