Fix Blurry Photos From Camera Shake for Free: ComfyUI-PhotoDeblur and the Two-Model Test That Tells You If It Worked
There are two kinds of blurry photo, and almost every "fix blurry photos" tool pretends there is one.
The first kind is shake. The camera moved while the shutter was open, so every bright edge got smeared along the same path: a neon sign with a ghost beside it, a window frame drawn twice. The detail is still in the picture. It is just spread out. The second kind is focus. The lens never put the detail on the sensor at all. Nothing is smeared; it simply is not there.
A new free tool for ComfyUI, ComfyUI-PhotoDeblur, is built for the first kind and says so plainly. That honesty is the reason it is worth an hour of your week, and the reason the way you use it matters more than the button you press.
What it is, in plain terms
ComfyUI-PhotoDeblur appeared on GitHub on November 7 under the MIT license. It takes two published research models for removing motion blur and puts them inside ComfyUI, the free node-based app many people already use for AI images.
The two models come from university labs:
- HI-Diff, presented at NeurIPS 2023 and released under Apache 2.0. It is a diffusion model, the same family of AI that draws images from noise, used here to rebuild sharp detail from a smeared photo.
- MISCFilter, presented at CVPR 2024 and released under MIT. It works differently: it estimates how each part of the picture moved and undoes that motion in one pass, with no repeated drawing steps.
Both licenses let you use the results in paid work. The weights download from the original authors' Google Drive, and the node checks each file against a known fingerprint before it uses it.
The practical upgrade is size. Research code like this usually runs on small test images. PhotoDeblur cuts your photo into overlapping squares, fixes each one, and blends the seams, so a full-resolution file from your camera comes back at full resolution.
Why "which blur is it?" is the whole game
Photographers lost a familiar tool here a few years ago. Photoshop's Shake Reduction filter disappeared in version 23.3, and Adobe's own community forum still carries a long thread of people asking for it back. So if you shoot handheld at night, at events, or through a car window, you have probably been living with smeared keepers, or paying for a separate app.
Here is the part most tools skip. Motion-deblur models learn by studying pairs of photos: one shaken, one sharp, same scene. They learn what a smear looks like and how to fold it back. Shown a soft-focus portrait, they have nothing to fold. At best they do nothing. At worst they invent crisp texture that was never there, which looks fine at a glance and wrong on a print.
The PhotoDeblur README is refreshingly direct about this. It says defocus, detail lost in blown highlights and very small faces "are usually harder to restore than ordinary camera shake." It reports its own mixed results: a night bookstore sign improved noticeably, while a small, heavily blurred stage face and a cat's face "showed limited improvement." It calls these "subjective observations from a small photo set." That is the voice of someone who tested the thing, and it tells you how to read your own results.
The two-model test
Since you cannot know in advance which model suits a given photo, the tool's own advice is to run both and compare. I think that comparison is the real product. It turns "AI fixed my photo" into a judgment you make with your own eyes.
Setup, once
- Install ComfyUI if you do not have it. You need a computer with an NVIDIA graphics card for the full setup; MISCFilter in particular needs extra NVIDIA software the README walks you through.
- In a terminal, go to
ComfyUI/custom_nodesand rungit clone https://github.com/porfeng/ComfyUI-PhotoDeblur.git. - From that new folder, using ComfyUI's own Python, install the two requirement files the README lists (
requirements.txtandrequirements-download.txt). For MISCFilter on NVIDIA, also installrequirements-misc-cu12.txt. - Restart ComfyUI and drag
ComfyUI-PhotoDeblur.jsonfrom the repo'sexample_workflowsfolder onto the canvas. It carries notes in English and Chinese.
The download node fetches the model files the first time. If you prefer, the README links each official download for manual install.
Each photo
- Start from a clean file. Develop your RAW first in your usual editor, then export a standard-color (sRGB) 16-bit TIFF or PNG. The tool does not read RAW files directly.
- Use the suggested starting settings: model variant
RealBlur_J, tile size 768, overlap 96, extra averaging (TTA) off. - Run MISCFilter, then HI-Diff, on the same photo with the same settings. Save both.
- Compare at the same zoom, 100 percent or closer, on three things: lettering and signs, faces, and high-contrast edges like window frames. The README names these because that is where smears show and where invented detail shows.
- Try the other variants only if neither result convinces you.
RealBlur_RandGoProare versions of each model trained on different example photos. The GoPro set was built by blending frames from high-speed video, while RealBlur pairs come from a real camera rig, so on true handheld shots I would start with RealBlur and treat GoPro as the second opinion. - Save the winner as 16-bit through Save Image (Advanced) if you will keep grading. The standard Save Image node writes 8-bit.
Keep the original in the race
Treat the untouched photo as a third option. If both repairs trade a soft edge for a brittle, overcooked one, the original wins, and you have learned something about that kind of shot.
Two settings worth understanding
TTA (test-time augmentation). With this on, the tool flips and rotates your photo eight ways, fixes each version, and averages the results. It can smooth out odd artifacts. It also costs roughly eight times the computing. Leave it off until a result is close but noisy, then try it on that one photo.
Match Input Tone. Repair can shift colors slightly. This optional node pulls the broad colors back toward your original while keeping the new sharpness. The README used it, with its sigma dial at 16, on a night city shot (setting the dial to 0 turns the correction off) with light trails. If your repaired photo looks a little off in color, turn it on before you start grading by eye.
Where it falls down
It cannot fix focus. Worth saying twice. If the eyes were soft because the lens focused on the ear, this is the wrong tool.
It fixes blur you wanted. A panning shot of a cyclist, a shallow-depth portrait, a long exposure of traffic: the model sees blur and removes it. The README's advice is to blend the result with the original through a mask in your photo editor, so only the subject gets sharpened. Plan on doing that for any photo where motion was part of the idea.
Small faces are a gamble. A face a few dozen pixels across has little real information. The README's own stage-face test barely improved. Do not promise a client a rescued group shot from the back row.
Setup is not one click. You need a working ComfyUI, an NVIDIA card for the full path, and patience with Python installs. If the install of NVIDIA's graphics software (CUDA) fails for MISCFilter, the README's setup-help section lists the usual causes. If you run out of graphics memory, drop the tile size to 512 and overlap to 64.
No speed promises. The README does not publish timing, and neither will I. It depends on your card, your photo size and the model. Time your first run and you will know.
The training photos are not yours. Each variant learned from a specific set of example photos. If your camera, lens or kind of motion is far from those, expect weaker results. That is one more reason to run the comparison instead of trusting a single pass.
Is it worth it?
For anyone who shoots handheld in low light, yes, today. Not because it rescues every frame, but because it rescues some frames you had already given up on, and it tells you honestly which ones it will struggle with.
The bigger lesson travels beyond this one tool. Every AI repair, deblur, upscale or cleanup, makes a guess about detail it cannot see. The habit that keeps you safe is the same each time: name the problem first, run more than one fix, compare at full zoom against the original, and let the original win when it deserves to.
So pick the three shaky photos you have kept the longest out of stubbornness. Run them through both models this week. You will come away knowing exactly which of your shots this tool can save, and that is worth more than any before-and-after on a product page.