Deflicker in ComfyUI: Remove Flicker from AI Video for Free

Deflicker in ComfyUI: Remove Flicker from AI Video for Free

In qualità di Affiliato Amazon io ricevo un guadagno dagli acquisti idonei.As an Amazon Associate I earn from qualifying purchases.En calidad de Afiliado de Amazon, obtengo ingresos por las compras adscritas que cumplen los requisitos aplicables.

You upscale an AI-generated clip, play it back and something's off: the leaves tremble, the hair sparkles in jumps, the sky pulses like a faulty light bulb. That's flicker. It happens because many upscalers work on one frame at a time and invent slightly different details in every image. On November 15 a node built for exactly this landed in the ComfyUI code: Video Deflicker. Deflicker in ComfyUI used to mean third-party nodes or a trip through DaVinci Resolve or After Effects; now it's a core node, free, and it runs on your own computer.

Our take? We upscale video a lot, and flicker is the reason we so often end up back in the edit fixing things by hand. The node comes from Kijai, the same developer behind the Wan nodes and many of the video workflows we use, and the file you need weighs 21 MB. We haven't tried it yet on our 12 GB RTX 3060, and there's one detail you should know right away: as of November 17 it isn't in a stable release yet. Below you'll find what it does, the settings explained and a simple workflow to build.

What it isA native ComfyUI node (VideoDeflicker) that removes flicker from video processed frame by frame
Who made itKijai, with a change merged by comfyanonymous into the main code on November 15, 2026
VRAM requirednot stated
Does it run on a 12 GB RTX 3060?probably yes: the RAFT model weighs 21.1 MB, but nobody states it and we haven't tried it yet
LicenseGPL-3.0 like all of ComfyUI; the Comfy-Org RAFT file is Apache 2.0
Commercial use from Italyyes, no regional exclusions
Where to get italready inside ComfyUI, in the development version after November 15; the RAFT file is on Hugging Face (Comfy-Org/void-model)
Free or paidfree, local

What is the Video Deflicker node and why does AI video flicker?

It's a node that makes the frames of a video consistent with each other without wiping out the detail in each one. Flicker in AI video almost always comes from processing done image by image: an upscaler like Real-ESRGAN, a style filter, a correction applied to every frame. Each frame looks great on its own. Put them in a row and the invented details keep changing, and your eye reads that as a tremble.

This is different from the flicker of neon lights or LEDs shot with the wrong shutter speed. There the brightness of the whole frame changes, and the classic deflicker tools in editing software work on exactly that. The ComfyUI node follows the motion of the scene instead, point by point.

According to the description of the change on GitHub, the node uses a method published in 2015 by Bonneel and colleagues (Blind Video Temporal Consistency). For each frame it computes the optical flow, an estimate of how pixels move from one image to the next, with a model called RAFT. Then it warps the previous, already cleaned frame to line it up with the motion and looks for a middle ground between that and the new frame. The math runs on the graphics card.

How do you use deflicker in ComfyUI: which settings matter?

You need two nodes: Load Optical Flow Model, which loads RAFT, and Video Deflicker. The official node documentation explains the inputs, which we sum up like this:

  • images: the flickering frames, for example the output of your upscale.
  • guide (optional): a stable copy of the source video, even at lower resolution. Motion is read from here. If you don't connect it, the node uses the flickering frames themselves and removes less flicker.
  • strength (default 1.0, from 0 to 100): how closely each frame follows its neighbors. Higher removes more flicker but softens fine detail.
  • direction: bidirectional (default) runs a forward and a backward pass and averages them, so the image doesn't lag behind fast motion; forward runs a single pass and is quicker.
  • threshold (default 0.1): changes bigger than this are treated as real, say a lightning flash or a lighting change, and get smoothed less.

The most useful tip in the documentation is about the guide: connecting the original video from before the upscale gives better results. The guide doesn't end up in the image, it's only used to work out how the scene moves. Watch the frame count: if the guide is shorter, the extra frames pass through uncorrected.

Which version of ComfyUI do you need?

You need the development code from after November 15, because the stable releases that came out since don't include it. We compared the releases on GitHub: 0.45.4 from November 15 and 0.45.5 from November 17 only add partner nodes (the paid ones) and include neither Video Deflicker nor native PAN 2 support, both merged into the main branch that same day.

So you have two options. If you use the portable Windows version, the update_comfyui.bat script in the update folder brings you to the latest development code (update_comfyui_stable.bat stops at stable releases). If you use ComfyUI Desktop or would rather not take risks, wait for the next stable release: for a project you have to deliver, that's what we'd do.

What changes with the new, faster RAFT?

Optical flow is about 2.3 times faster than before, according to the measurement Kijai published at 1024×576. It's the only number given, so we don't know how long a whole clip takes. In the same change Kijai rewrote RAFT inside ComfyUI to replace the torchvision version, in half precision (fp16, which takes half the memory) and with CUDA graphs when dynamic memory management is on.

The deflicker isn't the only thing that benefits. The same RAFT model is used by the VOID nodes, Netflix's system for erasing objects from video that we covered in the article on Generative Remove in Premiere and the free alternative. If you use them, the optical flow part gets faster without changing anything.

Try it now: upscale and deflicker in one workflow

  1. Update ComfyUI to the development code (see above) and restart it.
  2. Download raft_large_C_T_SKHT_V2-ff5fadd5.safetensors (21.1 MB) and put it in ComfyUI/models/optical_flow. ComfyUI won't download it for you.
  3. Load the clip with Load Video and split the frames with Get Video Components.
  4. Upscale as usual, for example with Upscale Image (using Model) and an ESRGAN model.
  5. Add Load Optical Flow Model and Video Deflicker: the upscale output goes into images, the original frames into guide.
  6. Put the video back together with Create Video (same frame rate as the source) and Save Video.

Start with the default values and a 3 or 4 second clip. If some flicker remains, raise strength in small steps; if the detail turns mushy, go back. Always check the result in motion, never on a still frame. To pick an upscaler, the most used video models are in our guide to AI video models in ComfyUI, and for adding frames there's the article on the NVIDIA nodes for upscaling and slow motion. If you've never opened ComfyUI, start with the ComfyUI beginner's guide.

Other news

  • SenseNova U1.5 and U1.6 (November 17): ComfyUI adds reasoning before generation, interleaved text and images, and self-correction for these models with this change. The official model is Apache 2.0, but the BF16 file ready for ComfyUI weighs about 35 GB: not something for 12 GB for now.
  • Bria RMBG 3.0 (November 16): the Remove Image Background partner node can now use the new version for cutouts, at about 0.026 dollars per image according to the price badge. It's already in 0.45.5. Our free, local method is still cutting out with ComfyUI and WeVisOne-VL.

Frequently asked questions

Does ComfyUI's deflicker work on any video?

It's designed for video processed frame by frame, such as upscales or style filters. On LED light flicker in real footage it may help, but that's not its purpose: the deflicker tools in editing software are better suited there. On generated videos that change shape from one frame to the next it won't work miracles.

Do I need to install custom nodes?

No, Video Deflicker and Load Optical Flow Model are core ComfyUI nodes. All you need is the development code from after November 15, until a stable release includes them, and the 21.1 MB RAFT file in the models/optical_flow folder.

Deflicker before or after the upscale?

After, because the node is made to fix the flicker that per-frame upscaling introduces. Connect the original video to the guide input: motion read from a stable source gives better results. If you then add frames for slow motion, do that last.

What is RAFT?

RAFT is an optical flow model presented in 2020: it estimates how far and in which direction each pixel moves between two frames. In ComfyUI it's used by the Video Deflicker node and by the VOID nodes for erasing objects from video. The ComfyUI version weighs 21.1 MB.

On our YouTube channel @tutorial3d you'll find videos with free workflows to download.

Tutorial3d
Grafico - Videodesigner - VisualFX

Ti consigliamo anche:



#Devo averlo!

VIDEO

INSPIRATION IN YOUR INBOX

Una mail. Un tutorial. Niente rumore.

Ti scrivo un solo tutorial scelto, con file progetto e note tecniche. Senza promozioni, senza affiliate. Cancellabile in un click.

[newsletter_form]