sprite-generation
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 6
- Forks
- 0
03 SKILL.md 正文
Sprite Generation Skill
PROBLEM
LLM-based sprite generation (PIL, Grok, MiniMax without proper workflow) produces garbage:
- 2/10 quality sprites
- Horizontal gaps between body parts
- Checkerboard artifacts
- Misalignment
- General-purpose image gen ≠ pixel art
RIGHT APPROACH (from wiki research)
Tier 1: BEST QUALITY - ComfyUI SDXL + Pixel Art LoRA
System: RTX 4090, ComfyUI 0.20.1, 22GB VRAM
Checkpoint: sd_xl_base_1.0.safetensors
LoRA: pixel_icon_v1.safetensors (or Pixel Art Stylizer SDXL)
Workflow:
CheckpointLoaderSimple → LoraLoader (strength=1.0)
→ CLIPTextEncode (positive + negative)
→ EmptyLatentImage (128x128)
→ KSampler (steps=25, cfg=8.0, euler)
→ VAEDecode
→ Quantize (cap at 64-256 colors)
→ SaveImage
Generation time: ~2 seconds on RTX 4090
Tier 2: BEST FREE - Pollinations.ai
- 256x256 sprites
- Respects size parameters (unlike Grok)
- Free, no GPU needed
- Use k-means post-process: alpha-binarize FIRST, then k-means k=24-32
Tier 3: HYBRID (what we've been trying)
- Generate body parts separately via LLM
- Composite with LPC sprites
- THIS APPROACH FAILS - wiki explicitly says "body parts pipeline = failed approach"
- Don't use PIL-based generation - produces simple geometric shapes, not pixel art
WHAT DOESN'T WORK
- ❌ PIL ImageDraw generation (simple shapes, not art)
- ❌ Grok API (ignores size params, returns 1408x768)
- ❌ General LLM image gen without pixel art LoRA
- ❌ k-means without alpha-binarize first (destroys colors)
- ❌ Body parts pipeline (failed approach per wiki)
WORKING ASSETS ALREADY AVAILABLE
- Cordon CC0: 205 sci-fi sprites at 128x128 in
assets/cordon/temp_cordon/sprites/ - LPC Sprites:
assets/generated/body_parts/(if properly generated) - Procedural drawPixelCharacter(): Works in game_demo.html
DECISION TREE
- Need sprites NOW? → Use Cordon CC0 assets or procedural fallback
- Want quality pixel art? → Use ComfyUI SDXL + pixel_icon_v1 LoRA
- Free only? → Use Pollinations.ai with proper prompting
- Don't use: PIL generation, Grok, body parts pipeline
PROPER PROMPTING FOR POLLINATIONS
"pixel art, 128x128, knight walk cycle, 8 frames, side view,
transparent background, limited color palette 24 colors,
pixel-perfect edges, no anti-aliasing, no gradients,
medieval armor, fantasy RPG sprite"
POST-PROCESSING (if using LLM output)
- Alpha-binarize (threshold transparency)
- k-means palette reduction (k=24-32 for 128x128)
- Nearest-neighbor upscale if needed
- Manual cleanup in Aseprite (5 sec/frame)
REFERENCES
- Wiki: ~/llm_wiki/wiki/sources/comfyui-pixel-art-sprite-pipeline-2026-05-01.md
- Wiki: ~/llm_wiki/wiki/sources/sprite-generation-learnings-2026-05-01.md
- Wiki: ~/llm_wiki/wiki/sources/pixel-art-ai-generation-research-2026-05-01.md
04 同仓库更多 Skills
instinct
Use when the user invokes /instinct or /in, or asks to "message Instinct", "ask Instinct", "tell Instinct", "check Instinct", or read replies from the configured peer assistant. iMessage first, WhatsApp Web as fallback.
hr
Compatibility alias for history-resume; produce the same fresh-state, provenance-aware resume briefing.
history-resume
Resume briefing for the current repository branch or worktree using fresh state and admitted history sources.
grok
Alias for the canonical grokbot browser conversation workflow; /grok talks to the Grok bot the same way /dot talks to Dot.
codexa-bench
Compare `codexa` and normal `codex` on a fair split of comparable work by default; use `--strict` for duplicate runs of the exact same task.
frontend-design
Create distinctive, production-grade frontend interfaces with high design quality. Use this skill when the user asks to build web components, pages, artifacts, posters, or applications (examples include websites, landing pages, dashboards, React components, HTML/CSS layouts, or when styling/beautifying any web UI). Generates creative, polished code and UI design that avoids generic AI aesthetics.