github.com-gykim80-perfectpixel-studio_-_2026-06-16_12-52-56

gykim80

AI-powered animation sprite studio — generate character sprite sheets with 8 directions and 100+ actions from a single text prompt (Wails + Go + React) # PerfectPixel**An AI-powered animation sprite generation studio**Turn a one-line character description into a base character, then auto-generate 100+ motionanimations (walk, run, attack, magic, …) and complete 8-direction sprite sets — exported informats your game engine can import directly.**[🌐 Homepage](https://pp.andrew-ai.app/)****English** · [한국어](README.ko.md) A power-up animation generated end-to-end by the actual pipeline (`gemini-3-pro-image`). Why this is a hard problem Plenty of tools draw a pretty picture. Game developers need something else: a 6-frame walkmust be exactly 6 frames, every frame must read as the same character, the background mustbe fully alpha-transparent, and the character's anchor must not jitter between frames — all in aformat an engine can import as-is. AI models are bad at these constraints. Frame counts come out wrong, poses bleed into eachother, hair color drifts mid-strip ( identity drift ), and the chroma-key background spillsonto the character's silhouette. PerfectPixel's core idea: let the AI render, but enforce quality, consistency, and precisionwith a deterministic post-processing pipeline. Even when the AI's output is non-deterministic,the result converges to a consistent quality bar. That is the engineering moat — see thebefore/after comparisons below, all produced by the real pipeline code on real AI output. Features Text → character : describe a character, pick a style (pixel art, chibi, cartoon …), andgenerate a base character with the background auto-removed. 100+ motion presets : a categorized keyword catalog — walk, run, jump, attack, magic,emotes, and more. 8-direction sets : 5 directions are generated by AI and the 3 mirror directions arederived by horizontal mirroring, for a consistent set at 37.5% lower generation cost . Self-correcting quality loop : generate → matte → extract → inspect → correctiveregenerate, up to 3 passes, converging frame count and motion quality. Real pixelization : shared-palette quantization + pixel-grid snapping for authentic dot art. Engine-friendly export : sprite sheet + manifest.json + Aseprite-compatible JSON +per-state GIF/APNG + individual frame PNGs, all in one pass. Multi-provider : choose Gemini, OpenRouter, fal.ai, or BytePlus as the backend. Sessions & gallery : work state is persisted to disk and results are auto-archived. Sample Output | Walk | Idle | Cheer | Power Up ||------|------|-------|----------|| | | | || Dance | Victory | Block | Low HP || | | | | How it works Generating one state (animation) runs this pipeline in the Go backend: description + style + motion preset │ ▼ build prompt ──► AI image generation (horizontal filmstrip) │ ▼ background detect & matte ──► frame extraction (cell segmentation) │ ▼ inspect (frame count / identity drift / motion presence) │ ├─ pass ─────────────► pixel quantization ──► done │ └─ fail ─► corrective retry hint ──► regenerate (up to 3×) Two things make this more than a naive retry loop: Best-candidate scoring — every attempt is scored score = Found*100 − errors*10 , and thebest candidate is kept. A perfect result returns immediately; if 3 passes still aren't perfect,the best-so-far is returned (never an empty hand). API errors and cancellations bail early. Measurement-driven retry hints — defects detected during inspection are converted intoprecise English correction instructions injected into the next prompt (e.g. "the previousresult read as 7 poses but exactly 6 are required; split the canvas into 6 even columns…" ).Combined with user feedback, each pass becomes a closed-loop self-correction that convergeson the defect rather than rolling the dice again. The engineering: deterministic correction The design philosophy in one line: signal processing, not heuristics. Three signal-processingaxes — wrapped in a self-diagnostic, self-correcting closed loop. 1. Background removal — chrominance-based matting Instead of thresholding RGB, colors are converted to YCbCr and the background is separatedusing only the chrominance (Cb, Cr) components, discarding luma (Y). This treats shaded andbright magenta as the same color, and is inherently robust to JPEG's 4:2:0 subsampling (whichpreserves luma but crushes chrominance). The background key is estimated as the mode of a CbCrhistogram (not the mean, so gradients/noise don't shift it), sampled from the four cornerswhere the character rarely intrudes. Soft alpha matting uses a Hermite smoothstep for edgefeathering, despill projects out only the key-direction color spill (keeping the character'sown colors), and a 4-connectivit