Zuanfilm Retrolab: A Free ComfyUI Node That Gives AI Images a Real Film Finish
AI images rarely look fake because of the details. They look fake because of the finish. A new open-source ComfyUI node brings 50 film looks, grain, halation and light leaks into the graph where the image was made, and it changes nothing until you tell it to.
Look at a hundred AI images in a row and you start to see the same thing in all of them, even when the hands are right and the text is spelled correctly. The skin glows a little too orange. The shadows never quite go black. Every highlight lands soft and clean. And the surface of the picture is perfectly smooth, like a print that never touched paper.
None of that is a detail problem. A better model with more steps will not fix it, because the model is giving you exactly what it learned to give: a polished, even, slightly overcooked average of every image it saw. It is a finishing problem. Photographers and colorists solved it decades ago with film stocks, grain and the small optical accidents of real lenses.
That is the bet behind Zuanfilm Retrolab, a free ComfyUI node released on November 17 under the Apache 2.0 license. (A node, if you are new to ComfyUI, is one box in the visual graph where you wire a model, a prompt and a few processing steps together.) Retrolab is one box with twelve finishing stages inside it, from a white balance wheel to film grain. It does not generate anything. It takes the image you already made and gives it a surface.
My position after reading the project closely: this is the right fix for the right problem, and the part most people will skip past, a slider literally called "Anti-AI," is the most useful thing in it.
What it actually does to a picture
Retrolab runs your image through a fixed order of steps, the way a lab would.
It starts with corrections. A white balance wheel shifts the picture warmer or cooler and greener or more magenta, a two-axis puck in the style of DaVinci Resolve's color wheels. The Anti-AI slider comes next (more on it below). Then exposure, a simple brighter or darker control.
Then come the optical effects. Bloom gives bright areas a soft glow. Halation adds the reddish ring that real film shows around bright lights, caused by light bouncing back through the film base. Light leaks spill colored light in from an edge or corner, in five shapes with names a film shooter will recognize, including "First-Frame Burn" and "Holga 135."
Then the look itself: a menu of 50 film stocks (the README calls them emulsions, after the light-sensitive coating on real film) plus a neutral setting. They come in four families: 10 cinema negatives (Kodak VISION3 500T, CineStill 800T and others), 10 instant films (Polaroid SX-70, Fujifilm instax), 10 slide films (Velvia, Ektachrome, a Kodachrome-inspired look) and 20 everyday 35mm stocks (Portra, Ektar, Gold, Superia). An Emulsion Response slider runs from half strength to triple strength, so you can lean on a look or barely tint with it.
After that: solarization (a partial tone reversal in the highlights, a darkroom effect), a soft focus that blurs toward the corners and leaves the center sharp, a vignette that darkens the corners like real lens falloff, an optional texture layer with options like dust, scratches and paper staining, and finally film grain with controls for amount, size and how colorful it is.
Every one of those stages starts switched off. A fresh Retrolab node hands back your image exactly as it came in. That sounds like a small design note. It is the reason the tool is pleasant to use: nothing happens unless you choose it, so you build the look up from zero instead of fighting a preset back down.
The slider called Anti-AI
The README describes the Anti-AI control in plain terms. It "gently compresses contrast, reduces red/orange chroma, softens the black toe, opens midtones and rolls highlights more gradually." Translated: it lowers the punchiness, pulls back the orange that AI skin tones drift toward, lets the shadows breathe, lifts the middle of the picture and stops highlights from slamming into white.
Read that list again and compare it with the complaints at the top of this article. It is almost the same list. Somebody looked at a pile of AI renders, named the habits and wrote a correction for each one.
Two things about it are worth knowing before you get excited. First, the authors say it is "not an AI detector" and not a local retouching tool. It is a tone correction, applied evenly across the whole frame, the same way every time. Second, it works on the whole picture at once. It will not fix one waxy face in a crowd or one blown-out window. It shifts the whole image toward a more photographic balance, and you judge whether that helped.
I think it belongs at the start of nearly every finish, at a low setting, before any film look. The film stocks all assume a reasonably photographic image going in. Give them an overcooked render and they tint the problem instead of fixing it. Take the edge off first and the stock has something to work with.
Why doing this inside ComfyUI matters
You could get many of these effects in Lightroom, Photoshop or a paid film plugin for Resolve or Premiere. So why want them in the graph?
Because AI work comes in batches. You rarely make one image. You make forty variations, pick six, then come back next week for twenty more in the same style. Every round trip to another app is an export, an import, a preset applied by hand and a re-export. Put the finish inside the graph and it travels with the workflow. Save the workflow once and every future render comes out with the same stock, the same grain and the same leak, without you touching it.
The grain behaves the way you need for that. Retrolab draws it from a fixed seed (a starting number for the random pattern), so the same image with the same settings gets the same grain every time. Across a batch, each frame gets its own slightly different grain, and the light leak holds its shape instead of jumping around. The README says this is "to avoid spatial flicker," which tells you the authors had image sequences in mind, not only stills. If you work with video in ComfyUI as a stack of frames, the node takes the stack as it is.
It is also light to install. It needs no model download, no account and no extra Python packages beyond what ComfyUI already has, and it runs on a normal processor, using an NVIDIA card automatically when one is free.
Put this into practice
Here is a first session that should take about twenty minutes.
- Install it. In ComfyUI Manager, search for Zuanfilm Retrolab and install it, or run
comfy node install zuanfilm-retrolabin your ComfyUI folder. Restart ComfyUI. - Wire it in. Put the Retrolab node right before your Save Image node. To test on finished work, use a Load Image node as the input instead.
- Queue once. The live preview needs one run to have something to show. After that, slider changes update the preview almost immediately, without re-running the whole graph. The small preview leaves grain off on purpose; grain shows in the actual-size crop.
- Start with Anti-AI only. Set it somewhere low, then flip the Before and After switch. If you cannot see a difference, raise it a little. Stop as soon as the image looks like a photograph rather than a render.
- Pick one stock, then turn it down. Try Portra 400 for people, Velvia 50 for landscapes, CineStill 800T for night scenes. Then pull Emulsion Response below 1.0 until the look feels like a choice and not a filter.
- Add grain last, and check it at full size. The preview includes a 240-pixel crop at actual size for exactly this. Grain that looks right on a shrunken preview is often too heavy at full resolution.
- Save the workflow under the look's name. "Portra-soft-anti-AI-20" is a better file name than "test3," and in a month you will want to find it.
Then do the honest test. Run the same ten renders with and without your finish, mix them up, and see which ones you pick when you do not know which is which.
Where it falls short
The film names are labels, not measurements. The README says it directly: the looks are "original, independently authored creative RGB transforms" that borrow stock names as vocabulary. They are not lab-measured copies of Portra or Velvia. If a client asks for a true Kodak match, this is not the tool, and the authors do not pretend it is.
It finishes; it does not repair. A render with a bad hand, a melted background or wrong light direction will still have them after Retrolab, just with nicer grain on top. Heavy settings are tempting precisely because they hide problems. Use them as a look, not as camouflage.
It is one day old. The repository had 9 stars on the morning after release. It ships with tests and unusually thorough documentation, which is a good sign, but nobody has lived with it on a real project yet. Keep your untouched renders.
The controls are many. Twelve stages with several sliders each is a lot of surface. The interface comes in ten languages and groups things well, but plan on a session of play before you trust your own settings.
Grain and screens do not always get along. Fine grain can turn to mush or shimmer after a social platform recompresses your upload. Test an export on the platform you post to before you commit a whole series to one setting.
The finish is the part you own
Most AI image work goes into the front of the process: the model, the prompt, the reference images. The back end, what happens to the picture after the model is done, usually gets a quick crop and an export. That is backwards. The model's habits are shared by everyone using it. The finish is where your images start to look like yours.
Install Retrolab, take the last batch you were almost happy with, and give it twenty minutes. If the Anti-AI slider and one well-chosen stock make you stop noticing that the images were generated, you have found a step worth adding to every workflow you build from now on. If they do not, you have lost an afternoon and learned exactly what your renders are missing.
- Sources: Zuanfilm Retrolab on GitHub · Retrolab README