In brief
- ComfyUI v0.46.0 removes DirectML support and the --directml launch option.
- The Video Deflicker node uses optical flow to reduce flicker between frames.
- Native support for PAN 2 has been added; the Split Image Channels node can separate color channels.
- A startup warning appears on Turing and newer NVIDIA GPUs if the PyTorch CUDA version is below 13.0.
ComfyUI core v0.46.0 was released on 18 November. The update removes the DirectML execution method preferred by some Windows users, while adding the Video Deflicker node for reducing flicker in video frames and native support for PAN 2. The changes, as reported by AI Creator Lab after reviewing the release notes, are particularly relevant to visualization and content production teams updating an existing ComfyUI installation.
What should DirectML users do before updating?
The new version no longer includes DirectML support or the --directml launch option. As a result, ComfyUI installations running on Windows with AMD or Intel GPUs via DirectML may fail to launch if updated as-is. Before updating, check whether this option appears in the launch file or shortcut you use.
According to the release’s README, an option is available to install PyTorch based on ROCm 10.0, distributed by AMD, for AMD graphics cards on Windows. This requires Windows 11, up-to-date AMD drivers, and 64-bit Python 3.13; supported architectures are listed as RDNA 2, RDNA 3, RDNA 3.5, and RDNA 4. v0.46.0 also includes portable Windows packages for AMD and Intel. Verify that the method you plan to use is compatible with your existing hardware and software environment before updating.
On NVIDIA systems, a startup warning appears if the system uses a Turing or newer GPU and has a PyTorch CUDA version older than 13.0. The warning notes that performance may drop significantly with older CUDA versions; the release notes recommend upgrading to PyTorch cu130 or newer. The standard NVIDIA portable package includes Python 3.13 and CUDA 13.0. The cu126 package, which uses CUDA 12.6 and Python 3.12, is reserved for GTX 10-series and older GPUs.
A new node for video workflows
Video Deflicker is designed to reduce flicker that can appear in videos processed or upscaled frame by frame. The node estimates motion between frames using optical flow, then evaluates neighboring frames together. An optional guide input lets you use a flicker-free original video to track motion; the source notes that flicker reduction may be more limited without this input.
Increasing the strength value applies more smoothing, but can also soften details. The default is 1.0, with an accepted range of 0–100. The direction setting selects bidirectional or forward processing: the bidirectional method combines forward and backward passes, but takes twice as long. The default threshold is 0.1; this setting determines which differences count as actual changes. The new ComfyUI implementation used for the node’s optical flow processing is reported to be about 2.3 times faster than the torchvision implementation at 1024 × 576, based on the developer’s measurement. Since the test hardware is not specified, this result cannot be generalized to every system.
Impact on image generation and architectural workflows
The Split Image Channels node separates an image into red, green, blue, and alpha channels. Save Image Advanced nodes can now also save single-channel grayscale PNG files. These changes can be used in ComfyUI workflows that process image layers separately or connect outputs to different nodes.
Core-node support for PAN 2 makes it possible to use the model within ComfyUI. For architecture and interior visualization teams, Video Deflicker is especially worth testing on animations processed frame by frame; however, image softening, processing time, and hardware compatibility may affect the results. The release summary does not include licensing or pricing information. Check whether frequently used custom nodes are compatible with the new version; making a copy of your existing ComfyUI folder before updating is also a useful precaution in case you need to roll back.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
The most tangible new feature in ComfyUI v0.46.0 for architectural visualization teams is the flicker-reduction node added to video workflows. For users enhancing animation frames individually, the option to use the original video as a guide is practical; on the other hand, the longer processing time for bidirectional mode and the possibility of softened details are points to test.
For offices in Turkey, it’s important to assess compatibility with your current setup before considering the cost of updating. DirectML users should not upgrade without changing methods, and NVIDIA users should check their CUDA version and hardware requirements. Since this source provides no details about custom-node compatibility or licensing terms, testing with a backup before updating production computers is the more cautious approach.
Frequently asked questions
When was ComfyUI v0.46.0 released?
According to the source, ComfyUI core v0.46.0 was updated on 18 November 2026.
Can you use DirectML in ComfyUI v0.46.0?
No. This version removes DirectML support and the --directml launch option.
What does the ComfyUI Video Deflicker node do?
Video Deflicker uses optical flow to reduce flicker between video frames. Motion in the original video can be tracked using the optional guide input.



