WVY-Cartoon-14m
A 14.7M-parameter cartoon character model trained from scratch on hand-drawn layers. Pick labels on the right and it draws the character right here in your browser. Almost every combination is one it never saw in training.
How it works
Labels in, picture out. The model has 110 on/off switches, one for every label in the panel. A character turns on exactly one switch per group.
labels → 110 switches → 768 neurons → 512 neurons → 96 × 96 × 3 pixels
Each switch is a learned recipe for one part. Turning on "heart eyes" adds that recipe. The next layers work out where each part goes and what covers what (hair over brows, a cup over a hand), then paint the pixels. That's why changing one label changes one part: swap one trait on a never-seen character and 94–99.8% of the pixel change lands inside that part's own area.
Model size follows output detail, not data. 96.7% of the 14.7M parameters belong to the final painting layer. Doubling the number of training combinations adds 0 parameters, and adding a brand-new part adds 768.
Fit rules
The training data always followed these, so the model learned them too:
- Shirts only go on the standing body. The other poses come without an outfit.
- Hand-held props (double cup, liquor bottle) only go on the standing body.
- The raised cup only goes with the "raise a cup" pose.
How it was trained
- Art: 55 pieces hand-drawn in Adobe Fresco, plus 77 new pieces drawn to match the style (new poses and body shapes in all 6 skins, outfits, hair, eyes, mouths, props). The artist reviewed and corrected the new pieces.
- Data: 10,000 characters stacked from the layers, each with exact labels. 300 random combos and 3 trait pairs (mohawk + heart eyes, pink hoodie + red-orange skin, waving + purple messy hair) were held out.
- Pretrain: 8,000 steps on all 10k. Fine-tune: 2,000 steps on 3,138 "good ones." Everything was written in numpy with hand-written backprop, on 2 CPU cores.
Benchmark (420 prompts it never saw)
For each prompt and each trait group, every other valid option is swapped in and re-rendered by hand. The check is whether the model's image sits closest to the correct one. A closest-label lookup, which returns the training image sharing the most traits, shows what pure memorizing gets.
| system | every trait right | traits right | PSNR |
|---|---|---|---|
| this model (fine-tuned) | 99.3% | 99.92% | 35.5 dB |
| this model (pretrained) | 99.5% | 99.95% | 35.8 dB |
| closest-label lookup | 0% | 56.6% | 15.5 dB |
What to look for (limits)
- Faint ghosts: blurry outlines of other hair or outfits can leak in. That's the average of everything the model saw in that spot.
- 96 × 96 pixels. Faces are readable but soft. Tick "show raw pixels" to see the real resolution.
- One image per prompt. The same labels always give the same picture.
- Closed vocabulary. It can only draw the 132 pieces it was given.
Learn more
Weights, code and the full model card are at StarpowerTechnology/WVY-Cartoon-14m. This page runs the same model in plain JavaScript: no server, no Python.
Draw the parts. Let it learn the grammar. · WVY · Starpower