image-to-psd Splits One Character Illustration Into a Layered PSD for Animation. Here Is the Step It Leaves to You
A free AI skill takes one finished character drawing and returns a Photoshop file with bangs, back hair, eye whites and irises on their own layers. What it gets right, what it guesses, and the one fix to make before the file goes anywhere near Live2D.
Every animated VTuber, every character that breathes and blinks in a visual novel, every 2.5D game portrait that tilts its head, starts life as a flat painting that somebody had to take apart. (2.5D here means a flat drawing animated so it seems to turn and move in depth.) Not traced. Taken apart. The bangs go on one layer. The forehead underneath them, which the artist never drew because the bangs were covering it, has to be painted in, so that when the hair sways the face behind it is not a hole. The same goes for the hair behind the neck, the whites of the eyes under the irises, the shoulder under the sleeve.
That job is called material separation, and it is tedious in a way that wears good illustrators down. It can take longer than the illustration did.
A small open project posted to GitHub on November 10, image-to-psd, claims to do it from one image. You hand it a character illustration (PNG, JPG or SVG), and it hands back a PSD in 19 named layers, each with its own pixels and its own editable mask, with the hidden areas filled in.
My position, after reading every file in it: this is the most useful kind of AI help an illustrator can get, because it keeps your actual pixels. It is also a draft, and the people who get the most out of it will be the ones who know which parts of the draft to distrust.
What it is, in plain terms
image-to-psd is an "agent skill." That is a folder of written instructions that a coding assistant, such as OpenAI's Codex or Anthropic's Claude Code, reads and follows step by step. It is not an app with buttons. You talk to the assistant, the assistant does the work on your computer, and a small script at the end writes the Photoshop file.
The install line from the README is:
npx skills add link-ding/image-to-psd --skill image-to-psd -a codex -g -y
You swap codex for the name of the assistant you use. The export script also needs three Python packages (pip install pillow numpy psd-tools), plus cairosvg if you feed it SVG files. Then you give the assistant your drawing and say: "Use $image-to-psd to split this character artwork into a layered PSD."
One oddity: the instructions inside the skill are written in Chinese. The README says the assistant follows them whatever language you prompt in, but reading the rules yourself takes a translator.
How it takes a drawing apart
The order of operations explains both why the results look like your work and where they go wrong.
First, it studies the picture and decides where each part ends. Which strands of hair belong to the bangs and which to the side locks. Where the sleeve stops and the hand starts. Which part is in front when the hand overlaps a prop. The instructions tell it to work on zoomed-in crops for the tricky borders: hands, sleeves and props first, then hair accessories and bangs against the face.
Second, it generates reference images of the parts. It asks an image generator (in Codex, the built-in image tool) to paint what each part should look like whole, including the bits your drawing hides.
Third, it cuts the visible pixels from your original, and by default uses the generated references only where something is hidden. This is the design choice that matters. The line you drew, your color, your brush texture, all come from your file. By default, generated paint goes only where the drawing never showed anything: the face under the bangs, the hair behind the head, the eye whites under the irises.
Fourth, it exports and then reviews its own file. The script writes each part as a full-canvas layer, lined up exactly with the original, then the assistant reopens the PSD and checks it layer by layer for paint that bled into the wrong layer, gaps at seams, and stray fragments. For eyes it has a specific test: look at the eye whites alone, hide the irises and check the composite, then slide the irises left and right and see whether anything breaks.
The README's examples are three characters at 1254 by 1254 pixels, between 10 and 12 MB each, and you can download the PSDs and open them in Photoshop, Krita or Clip Studio Paint before installing anything.
The 19 layers, and what they mean for rigging
The fixed layer list is the real strength of the tool. Every character comes out with the same names in the same order: back hair, front hair, face, ears, earwear, eyebrow, eyewhite, irides, eyelash, nose, mouth, headwear, neck, topwear, handwear, bottomwear, legwear, footwear, objects. Parts that are not in the drawing stay as empty layers rather than disappearing. If a hand and a held prop overlap in both directions, handwear can split into a back and a front layer, which makes 20.
If you produce more than one character, a template rig built for one keeps working for the next.
Now look at the list with a rigger's eye (the rigger is the person who sets up how each part moves). Both eyebrows sit on one layer, and so do both eye whites. The README says so for those two, and the skill's instructions do the same for irises and lashes. That is a sensible default for a cutout, and it is not how most people set up a face that winks. If you want each eye to blink on its own, or one eyebrow to raise, you will split those layers into left and right yourself. It is a five-minute job with the lasso tool, because the two halves are already separate shapes on the layer, but it is your job.
The mouth is the other gap. The instructions say it keeps the mouth exactly as drawn and does not add the extra mouth shapes (the open vowels) that lip sync usually needs. A character that talks still needs you, or a separate generation pass, to draw those.
Put this into practice: the hand-off to Live2D
This is the step the launch coverage, such as it is, has skipped, and it is the one most likely to cost you an afternoon.
image-to-psd gives every layer an editable mask. That is a good choice in Photoshop, where a mask lets you nudge an edge without destroying pixels. But Live2D Cubism, the program most VTuber and visual-novel characters are animated in, says otherwise. Its own documentation for preparing PSDs, the notes on PSD creation in the Cubism Editor manual, lists "Do not use layer masks" as a rule, and says: "If the PSD to be imported still has layer masks, execute 'Apply Layer Mask' to combine them into one layer."
Here is why that is safe to do with this tool's files. The skill keeps the real shape of each part, soft edges and semi-transparent bits included, in the layer's own pixels. The mask only marks which pixels belong to the layer at all. So applying the mask throws away nothing you need. You lose the ability to nudge the edge non-destructively, which is fine, because by then you are done editing.
The routine I would follow:
- Run the skill and read what the assistant tells you it filled in. The instructions require it to name the parts it inferred. Write that list down.
- Open the PSD in Photoshop or Clip Studio Paint. Hide the front hair and look hard at the forehead. Hide the irises and look at the eye whites. Hide the sleeves. These are the invented areas, and they are where style drift shows.
- Repaint anything in the filled areas that does not look like your hand. Usually it is small: a line weight, a shadow color.
- Split the eyes, brows and lashes into left and right if your rig needs it. Give every new layer a unique name. Live2D's manual warns that duplicate names "can cause confusion."
- Apply every layer mask.
- Check the file uses Photoshop's standard RGB color at 8 bits per channel (the everyday setting). Live2D says grayscale, CMYK, 16-bit and 32-bit files "cannot be imported." Set the color profile to sRGB, the standard screen profile, while you are there.
- Save a copy as your import file and keep the masked version as your editing master.
Photoshop and Clip Studio Paint are the two applications Live2D has confirmed load PSDs correctly, so do the final save in one of those rather than Krita. The skill's own troubleshooting notes make the same point from the other side: a PSD that opens correctly in one program does not prove it imports correctly into your animation tool, and the assistant is told not to claim it does until you have actually imported it.
Where it breaks
Be clear-eyed about these before you put it on a client job.
The hidden areas are guesses. They are guesses guided by your drawing and by generated references, and the review pass catches a lot, but a generated forehead is still a generated forehead. On a stylized character with a distinctive face shape, expect to repaint. The README says it plainly: hidden areas are inferred.
It does not promise a finished rig. In the README's words, it "does not promise one-click, production-grade rigging quality for any artwork." It cuts and fills. Setting up how each part bends, sways and responds to the animator's sliders is still the rigger's work, and that is most of the work.
Your results depend on someone else's image generator. The part references come from an image tool, so you pay whatever that tool costs, and its quirks show up in your fills. The README gives no timing and no cost per character.
There is no license file. GitHub shows no license on the repository. Under GitHub's own guidance, that means the author has not granted permission to reuse or redistribute the code. Your own drawing and the file you make from it are a separate matter, but if you plan to bundle the skill into a product or a paid service, ask the author first.
It was built for a particular kind of drawing. The examples are all stylized anime-style characters. My expectation, not a claim from the repo, is that a painterly piece with soft, lost edges or a character in a twisted pose will need far more cleanup.
Why this is the right kind of AI help
Most AI character tools ask you to give up your drawing and accept the model's version of it. This one asks for less. It takes the slowest, least creative step in preparing a character for animation, cutting and back-filling, and does a first pass you can check in ten minutes, while every visible line stays the one you drew.
The skill you need does not change. You still need an eye for a seam that will open when the head turns, for a forehead that is a shade too warm, for a mask that has to be applied before import. The tool just moves your time from the lasso to the review.
So pick one finished character you own. Download one of the example PSDs first and see whether the layer quality meets your bar. If it does, run your own drawing through it, hide the bangs, and look at the forehead. Whether you would ship that forehead is your decision, and it is the right one to be making.