- name
- image-generation
- description
- Generate, edit, and upscale AI images. Use when creating visual assets for apps, websites, or documentation. FREE Cloudflare tier for iterate generation (~96/day), Fal.ai for paid tiers. Four quality tiers (iterate/default/premium/max). Supports text specialists, multi-ref editing, SVG, background removal. Triggers: generate image, create image, edit image, upscale, logo, picture of, remove background.
- metadata
- author
- ai-dev-atelier
- version
- 3.0
Image Generation
Generate, edit, and upscale images with standardized quality tiers and embedded best practices.
Quick Start
Need image?
├─ Text/Logo → bun scripts/gen.ts "..." --text [-t tier]
├─ Photo/Art → bun scripts/gen.ts "..." [-t tier]
├─ Edit existing → bun scripts/edit.ts <img> "..." [-t tier]
├─ Upscale → bun scripts/upscale.ts <img> [-t tier]
├─ Vectorize → bun scripts/svg.ts <img> ($0.01/img)
└─ Remove BG → bun scripts/rembg.ts <img> (FREE)
Tier selection:
├─ iterate → FREE drafts (~96/day via Cloudflare)
├─ default → Daily driver ($0.008/MP)
├─ premium → Final assets ($0.03/MP)
└─ max → Critical work, SOTA ($0.06-0.07/MP)Entry Points
| Script | Purpose |
|---|---|
bun scripts/gen.ts | Text → Image |
bun scripts/edit.ts | Image + Instruction → Image |
bun scripts/upscale.ts | Image → Larger Image |
bun scripts/svg.ts | Image → SVG ($0.01/img) |
bun scripts/rembg.ts | Remove background (FREE) |
Prompting Best Practices
CRITICAL: Good prompts are the difference between unusable output and production-ready assets.
The Universal Prompt Structure
[Subject] + [Action/Pose] + [Environment] + [Style/Medium] + [Lighting] + [Camera/Composition]Example:
"A cybernetic owl perched on a neon sign in a rain-soaked alley. Cinematic lighting with teal and orange highlights. Shot on 35mm film, shallow depth of field, hyper-detailed textures."
DO: Effective Prompting
| Technique | Example |
|---|---|
| Be specific | "middle-aged man with salt-and-pepper hair wearing charcoal turtleneck" NOT "a man" |
| Describe the result | "person with clear eyes" NOT "remove glasses" |
| Use camera terms | "Shot on Hasselblad, 85mm lens, f/1.8" |
| Specify lighting | "golden hour rim lighting with deep shadows" |
| Include textures | "weathered sandstone", "anodized aluminum", "iridescent silk" |
DON'T: Common Mistakes
| Mistake | Problem | Fix |
|---|---|---|
| Negative phrasing | "no glasses" often adds glasses | Describe what IS there |
| Vague subjects | AI interprets randomly | Be exhaustively specific |
| Keyword salad | "4k, trending, masterpiece" is noise | Use descriptive sentences |
| Short prompts | Under 20 words underperforms | Aim for 40-80 words |
Style Keywords That Work
| Category | Keywords |
|---|---|
| Lighting | golden hour, volumetric lighting, Rembrandt lighting, neon rim light, bioluminescent |
| Camera | 35mm anamorphic, macro photography, tilt-shift, fisheye, drone shot |
| Style | cinematic, photorealistic, concept art, ukiyo-e, baroque, impressionist |
| Quality | hyper-detailed, sharp focus, 8k resolution, raytraced |
Text & Logo Generation (--text flag)
Uses Recraft V3 (iterate/default) or Ideogram V3 (premium/max) - specialized for typography.
Text Prompting Rules
CRITICAL: Put text in "Double Quotes" at the START of your prompt.
# Correct - text first, then describe
bun scripts/gen.ts '"QUANTUM" in bold futuristic font, metallic silver, dark space background' --text
# Wrong - text buried in description
bun scripts/gen.ts 'A logo with the word QUANTUM on it' --textLogo Design Patterns
| Style | Prompt Pattern |
|---|---|
| Minimalist | "BRAND" minimalist vector logo, clean lines, simple geometry, flat design |
| Vintage | "EST. 1920" vintage badge logo, circular emblem, ribbon banner, ornate border |
| Negative space | "PEAK" logo where the letter A forms a mountain, negative space design |
| 3D/Modern | "TECHCORP" bold 3D chrome letters, gradient fill, dark background |
Font Specification
Use typography terms: modern sans-serif, elegant script, bold blocky, blackletter, neon tubing, retro 70s serif
DO/DON'T for Text
| DO | DON'T |
|---|---|
| "Three cats playing" (exact count) | "cats playing" (random count) |
| "wooden baseball bat" (specific) | "bat" (ambiguous) |
| Describe only what you want | "no cake" (will add cake) |
Image Editing
bun scripts/edit.ts <image> <instruction> [-t TIER] [--mask <mask.png>] [--ref <img>...]Writing Edit Instructions
Key: Describe the TARGET STATE, not the change.
| Bad Instruction | Good Instruction |
|---|---|
| "change car to blue" | "A sleek blue metallic sports car, reflections of neon lights on wet asphalt" |
| "add a hat" | "person wearing a vintage red fedora, matching the scene lighting" |
| "remove background" | Use rembg.ts instead (FREE and better) |
Mask Best Practices
| Task | Mask Strategy |
|---|---|
| Object removal | Mask LARGER than object (10-20px margin) for seamless fill |
| Object addition | Mask exact shape or slightly smaller |
| Outpainting | Overlap 10-20px INTO original image |
Feathering: Apply 12-16px blur to masks. Sharp masks = visible seams.
Multi-Reference Editing (--ref)
Using 2+ reference images auto-selects max tier (flux-2-flex).
# Style transfer: apply reference style to base image
bun scripts/edit.ts base.jpg "in the style of the reference" --ref style.jpg
# Multi-reference blending
bun scripts/edit.ts scene.jpg "forest sofa scene" --ref forest.jpg --ref sofa.jpgTip: When blending references, describe their relationship: "A velvet sofa placed in a misty pine forest"
Upscaling
bun scripts/upscale.ts <image> [-t TIER] [--scale 2|4]When to Use 2x vs 4x
| Source Quality | Recommendation |
|---|---|
| High (RAW, clean PNG) | 4x safe - AI infers detail accurately |
| Medium (standard JPEG) | 2x preferred - denoise first if possible |
| Low (compressed, blurry) | 2x max - noise gets magnified |
Use Case Guidelines
| Output | Scale | Notes |
|---|---|---|
| Web/UI | 2x | Reduces file size, improves perceived sharpness |
| Print (300 DPI) | 4x | Target 300 DPI for print quality |
| Icons/Logos | 2x | Use svg.ts instead for infinite scaling |
Common Artifacts & Fixes
| Artifact | Cause | Prevention |
|---|---|---|
| Haloing (white edges) | Aggressive sharpening | Use iterate/default tier |
| Plasticky skin | Over-smoothing | Reduce to 2x, use premium tier |
| Grid patterns | Tile processing | Use higher tier models |
Rule of Thumb: If image looks "crunchy" at 100% zoom, don't exceed 2x.
Tier Selection Guide
| Scenario | Tier | Why |
|---|---|---|
| Exploring 10+ variations | iterate | FREE, fast iteration |
| Daily work, 3-5 variations | default | Best cost/quality balance |
| Client deliverables | premium | Higher fidelity |
| Critical assets, multi-ref | max | SOTA quality, advanced features |
| Text/logos (any) | default | Recraft V3 already excellent |
| Text/logos (critical) | premium | Ideogram V3 for perfect typography |
Cost Optimization
EXPENSIVE WORKFLOW (avoid):
Generate at max tier → iterate on max → deliver
COST-EFFECTIVE WORKFLOW (recommended):
Generate at iterate (FREE) → find best concept
→ Regenerate winner at default/premium → deliverEnvironment
# For FREE iterate generation (Cloudflare)
CLOUDFLARE_ACCOUNT_ID=xxx
CLOUDFLARE_API_TOKEN=xxx
# For paid tiers (Fal.ai)
FAL_API_KEY=xxxQuota: Cloudflare FREE tier allows ~96 images/day at 1024x1024.
Exit Codes
| Code | Meaning | Action |
|---|---|---|
| 0 | Success | Image saved to .ada/data/images/ |
| 1 | General error | Check error message |
| 2 | Config/auth error | Verify API keys in .env |
| 3 | Resource limit | Quota exceeded - wait 24h or use paid tier |
CRITICAL: Exit code 3 does NOT fall back to paid tier. This prevents accidental charges.
Integration
| Skill | When to Use Together |
|---|---|
ui-animation | Animate generated images for web/mobile |
docs-write | Document image assets and parameters used |
search | Find prompting resources and style references |
code-quality | After modifying skill scripts |
References
references/usage-guide.md- Extended prompting guide, error codes, testingREADME.md- Architecture diagrams, model reference, CLI details- Fal.ai Docs - Official API documentation
Output
Images saved to .ada/data/images/ with timestamped filenames:
20260118_gen_default_cyberpunk_city.jpg
20260118_svg_default_logo_vector.svg