Clean-Plate Cutouts With AI: How PlateExtract Turns Two Photos Into a Transparent PNG With Soft Edges
The cleanest way to cut a person out of a photo has been known for decades. Shoot the frame with them in it. Shoot it again without them. Compare the two, and whatever changed is your subject.
Compositors call that empty frame a clean plate, and the comparison a difference key. It is simple, fast and wonderful right up until it meets hair. Or a wine glass. Or a soft shadow on the floor. Then it falls apart, because subtracting one picture from another can tell you a pixel changed, but it cannot tell you how much of that pixel belongs to the person and how much still belongs to the wall behind them.
A free add-on released on October 4 takes a real swing at that exact gap. It is called PlateExtract, it was made by Xavier Jara, and it does one job: you give it the composite and the clean background, and it gives you back the foreground as a transparent PNG. The README promises that the result "supports soft edges and partial transparency," which is precisely the part the old trick gets wrong.
My view after reading through it: the idea is right, the shoot discipline matters more than the model, and the license decides what you can do with it. Here is how to use it well.
What the old trick misses, and what this one adds
A traditional difference key does arithmetic. Pixel by pixel, it asks: is the composite different from the plate here? If yes, keep it. If no, make it transparent. Turn up the threshold and you lose fine detail. Turn it down and you keep noise, compression grain and the faint light change from a cloud passing over.
The worst cases are edges that are partly see-through. A strand of hair is mostly background with a little hair in it. A glass shows the wall behind it, slightly bent. A shadow is the floor, only darker. Arithmetic sees all of those as "changed" or "not changed." The truth is somewhere in between, and that in-between value is what makes a cutout look real when you drop it onto a new background.
PlateExtract approaches the problem as a picture-making task rather than a math one. It is an add-on for Viggle's fast build of Qwen Image 2.1, an image model with downloadable model files that can produce images with a built-in transparency channel. The add-on was trained to look at the two pictures together and paint the subject onto a transparent canvas, including the half-transparent edge. It finishes in four passes, which is quick by the standards of these models.
The two example files in the repository are an anime frame and a watercolor. That hints at something useful for illustrators and animators: if you have a background painting and the same painting with a character on it, you may be able to lift the character out without tracing it.
What you need before you start
Be honest with your hardware first.
The author tested PlateExtract on one NVIDIA card with 16 GB of memory, with part of the model parked in the computer's regular memory, and 64 GB of system RAM. That is a solid desktop, not a laptop. Macs are not part of the tested setup.
You also need a little comfort with a terminal. The README documents only a command-line script: no app, no plug-in for node-based tools like ComfyUI, no web demo. It is a download and a single Python command.
And you need the right two images, which is where most of the work actually lives.
Put this into practice
Here is the full workflow, from shoot to comp.
1. Shoot the plate like you mean it
The README does not spell out how closely the two images must match, so treat this as the classic clean-plate discipline rather than a rule from the author. Everything that changes between your two frames is a question the model has to answer, and you want it answering only one question: what is the subject?
- Lock the camera. Tripod, no zoom change, no refocus. If you can, trigger remotely so nobody bumps it.
- Lock the exposure. Manual exposure, manual white balance, same ISO. Auto settings will shift between frames when the subject leaves.
- Shoot the plate right after the subject steps out. Daylight drifts. A two-minute gap is better than a twenty-minute one.
- Watch for things the subject moved. A chair they nudged, a door they opened, a cable they kicked. Each one turns into an unwanted piece of foreground.
For illustration, the equivalent is simple: export the background layer and the full composite at the same size from the same file.
2. Download and run it
From a terminal with Python and an NVIDIA card set up:
hf download trmz/plate-extract-qwen-image-2.1 --local-dir plate-extract
cd plate-extract
pip install -r inference/requirements.txt
python inference/qwen_extract.py --composite picture.png --background clean_background.png --output extracted.png
The script loads the base model and the add-on for you. The add-on is a single model file, loras/extract.safetensors, if you want to keep it somewhere specific.
3. Check the edges before you trust it
Open extracted.png in your editor over three backgrounds: pure black, pure white and a mid grey. Zoom to 200 percent on the hair, any glass, and the ground contact. A bad edge hides against one of those backgrounds and screams against another.
Look for two specific failures. A halo, which is a faint outline of the old background clinging to the subject. And holes, where something in the subject happened to match the plate (a grey shirt against a grey wall is the classic) and got made transparent.
4. Comp it, then fix by hand
Drop the cutout onto your new background and judge it there, not on a checkerboard. Light direction and color temperature will sell or kill the comp long before the edge does. Fix any holes with a quick hand-painted mask. You are still the compositor. The model just did the slowest part.
5. Keep both originals
Save the composite and the plate together with the result. If the client asks for a change, or a better version of the model appears, you can rerun it in a minute instead of reshooting.
Clean plate or one-click remover?
You already have one-click background removers in most photo apps, so it is fair to ask why anyone would shoot a second frame.
A one-click remover has to guess what the background is. It has never seen the wall behind your subject's hair, so on every soft edge it is inventing. PlateExtract has seen that wall, because you photographed it. The guesswork drops from "what is back there?" to "how much of this pixel is hair?", which is a much easier question.
That makes the choice simple. If the subject has crisp edges against a plain background, a one-click tool is faster and good enough. If the shot has flyaway hair, a veil, glassware, smoke or a soft contact shadow you want to keep, and you can control the shoot, the clean plate is worth the extra ten seconds of camera time.
Where it breaks
The license rules out paid work. This is the big one. The add-on's model files are under the Qwen Research License (as is Qwen Image 2.1 itself), which allows use "FOR NON-COMMERCIAL PURPOSES ONLY" and says you "shall not use the Materials for any commercial purpose without obtaining a separate commercial license." Only the small script that runs it is MIT. So treat PlateExtract as a way to test the technique on personal work, mood boards and previs, not as a step in a client pipeline. If the result convinces you, that is useful information for choosing a commercial tool.
It is for stills. Nothing in the README covers video. You could run it on every frame of a locked-off shot, but each frame would be solved separately, and I would expect the edges to flicker from frame to frame. That is my expectation, not the author's claim. Test it on ten frames before you plan a shot around it.
It needs a real clean plate. No plate, no tool. That rules out most found photos, stock images and anything shot handheld. If you did not plan the plate at the shoot, a general background remover is still your option.
It needs a serious machine. Sixteen gigabytes of graphics memory and 64 GB of RAM is the tested setup. Smaller machines may work slowly or not at all, and the README does not promise either way.
The author published no accuracy figures. There is no comparison against other cutout tools in the README, just the method and the examples. Your own test is the only evidence that counts.
Why it is still worth an afternoon
Most AI image tools ask you to give up control: describe a picture and accept what comes back. PlateExtract asks for the opposite. It rewards the photographer who locked the tripod, the illustrator who kept their layers, the compositor who shot a plate out of habit. The better your craft at the shoot, the better the model does.
That is the right relationship between a creative and an AI tool, and it is rare enough to notice.
So here is a small experiment. Next time you set up a shot, take the extra frame. Subject in, subject out, nothing else changing. Run both through PlateExtract on a personal piece, put the cutout on a new background, and look hard at the hair. Whether it impresses you or disappoints you, you will know something concrete about where AI cutouts stand today, and you will have built the clean-plate habit that every better tool after this one will reward.
Medium metadata
Title: Clean-Plate Cutouts With AI: How PlateExtract Turns Two Photos Into a Transparent PNG With Soft Edges
Subtitle: A free add-on that gives the classic difference key the judgment it never had on hair, glass and shadow. The shoot checklist, the command, the edge checks, and why the license keeps it out of paid work.
Tags: Compositing, Photography, Photo Editing, AI Tools, Visual Effects
Estimated read time: 8 minutes