ComfyUI 0.44 exports AI video to ProRes and upscales Flux to 4K

Generative video has a dull but costly problem: creating an attractive clip is not enough—the result still has to reach the editing suite without unnecessary conversion passes. ComfyUI 0.44 makes this part of the workflow shorter. The Save Video node now supports MOV output with the Apple ProRes codec, while Flux Video Upscale V2 can target not only 1080p but also 2K or 4K.

Here is what that means in a simple example. Previously, a team would generate a scene, save it in an available format, launch FFmpeg or a video editor to transcode it, and only then hand the footage to a post-production specialist. Every extra step took time, created another file, and could degrade the image. ProRes can now be selected directly in the standard video-saving node.

ProRes becomes part of the regular graph

For encoding, ComfyUI uses the FFmpeg prores_ks software encoder. No separate paid plugin is required. Users can select 4:4:4, 4:2:2, or 4:2:0 chroma subsampling, or leave the setting in automatic mode; the listed profiles include ProRes HQ and 4444.

The practical value is not the addition of another menu item. ProRes is useful as an intermediate format: editing systems handle it well, and footage can be handed to a colorist or video editor without manual preparation. The 4:4:4 mode is particularly useful when color transitions, masks, and subsequent compositing matter. For an ordinary preview, however, such a file will be unjustifiably large—the release does not promise compression miracles.

New models and upscaling to 4K

The core now supports PH3Image, while MiniMax-H3 gains Sol-H3-VAE loading, a latent upscaler, and context windows for video pipelines. The partner Flux Video Upscale V2 node from Black Forest Labs now offers target resolutions of 1080p, 2K, and 4K. This makes it possible to build a single graph that generates a scene, increases its resolution, and saves an editing-ready file.

One limitation is important to understand: the “4K” label describes the output dimensions but does not guarantee the creation of genuine detail. If the source is blurry or contains trembling hands or unstable objects, the upscaler may make artifacts more visible. A production workflow still requires frame review, comparison with the source, and testing on several difficult scenes.

The release also updates the web interface to version 1.58.1 and the template library to 0.11.101. The core eliminates redundant tensor comparisons, fixes FP8 computation fallback to BF16 under dynamic VRAM management, and bypasses Python for the memfd_create system call. These are not presentation-friendly features, but changes like these affect the stability of long graphs.

What developers and product teams should do

If ComfyUI is already integrated into a production pipeline, update the test environment first and run several existing graphs. Pay particular attention to third-party nodes, available VRAM, ProRes file sizes, and color accuracy after importing files into the editing application. Updating through Manager or git pull is not a substitute for backing up the environment and pinning dependency versions.

For new pipelines, a sensible sequence is to select a model, test upscaling on a short clip, choose the required chroma subsampling mode, and only then launch a long render. The main value of ComfyUI 0.44 is not any single new generator, but the reduction in manual operations between the model and a finished video-production workflow.