One AI-generated waifu, exploded into a full character sheet

What it does

The project takes character art of “AI 娘” (anthropomorphized AI assistants — GLM, GPT, DeepSeek, Qwen, Kimi, Claude, and more, mostly from Bilibili artist ZipZipPipe) and turns each single illustration into a full horizontal character design sheet. Each character gets 22 assets: 3 views (front/side/back), 8 expressions, 4 clothing items laid out flat, and 7 prop close-ups, all composited at native pixel scale with no resampling.

The interesting bit

The side and back views don’t exist in the source art — they’re inferred by an image model from visible structure, including things like the back of a skirt panel or the inside of a cape. The pipeline is also unusually honest: every generated asset has a JSON record with model, quality, dimensions, and sha256, and a separate verify_board.py script re-verifies the composite from disk rather than trusting the composer’s own claims — byte-by-byte, checking for zero overlap, zero out-of-bounds, zero scaling.

Key highlights

• 17 character entries (18 sheets — DeepSeek Q-version has two outfits), each 22 assets, all passing byte-level verification
• Full prompt engineering per asset: each of the 22 tasks has its own JSON prompt specifying what to keep, what to change, and what this image alone should contain
• Native-pixel 1:1 slices kept per section, usable directly as line-art tracing references
• Failed generations are kept as v01 records and regenerated as v02 — the audit trail survives
• Entries 15–17 use a built-in image_gen tool whose underlying model isn’t exposed; the records only claim what’s actually confirmable

Caveats

• CC BY-NC-SA 4.0 — non-commercial only, and several source images (entries 12–14) have unattributed authors with “unknown” licensing status
• The full-resolution master sheets are deliberately not in the repo (.gitignore); older entries’ raw assets live in a local project, so rebuilding those masters requires supplying the assets yourself
• The “side/back views” are model inference, not canon — treat them as plausible reconstruction, not ground truth

Verdict

Worth a look for anyone building image-generation pipelines that need provenance and verification discipline — the workflow documentation is the real artifact here. If you just want the pretty sheets, they’re there too, but you can’t ship them in anything commercial.