A character sheet so rigorous it has acceptance tolerances

What it does Whale-chan is a fan-created mascot born from DeepSeek’s reasoning outputs — a whale-fin maid girl who treats white rice as compute currency. This project packages her into a machine-usable character bible: locked visual features (hair gradient, fin ears, cowlick, tail, maid uniform), a library of reference sheets, and agent skills that generate illustrations, posters, stickers, and multi-panel comics. It routes across image providers (Codex ImageGen, OpenAI, Nano Banana, Seedream) and ships local scripts alongside the skills.

The interesting bit Character consistency is enforced with actual engineering discipline: five body forms from 4.0-head standard down to 2.1-head super-deformed, each with a strict head-to-body acceptance range (e.g. semi-chibi must land in 2.695–2.995) verified by skeletal joint fitting. There’s also a “mindset engine” that classifies a source’s comedy type — reversal, escalation, loophole gag — and preserves the tone when converting chats or error logs into comics.

Key highlights

• Five quantified form profiles with measured skeletal ratios, not vibes
• Reference sheets covering costume layers, expressions, fin-ear language, and tail motion
• Comic pipeline that extracts a “Fact Anchor” from technical chats, CoT traces, or rants
• Multi-provider image routing with single-prompt delivery
• Trilingual README (English, Chinese, Japanese); MIT for code, CC BY-NC-SA 4.0 for assets

Caveats

• Explicitly unofficial — not affiliated with DeepSeek; trademark terms apply
• Asset license is non-commercial, so the pretty parts can’t ship in commercial work
• The README is heavy on showcase; how well the ratio verification holds up in practice isn’t demonstrated with numbers

Verdict Worth a look for anyone building consistent AI characters — the quantified-proportions approach is a genuinely transferable idea. If you don’t care about mascots or fan comics, the comedy engine won’t convert you.