# Model Testing Guide
# Wallpaper Studio - Cloudflare Workers AI Models

## Available Models (from Cloudflare)

### Text-to-Image Models

1. **@cf/black-forest-labs/flux-1-schnell** (CURRENT DEFAULT)
   - Description: Fast generation, 12B params
   - Use: Quick iterations, testing
   - Good: Atmospheric scenes, decent quality
   - Bad: Struggles with surreal/abstract concepts

2. **@cf/leonardo/lucid-origin** (TESTED - GOOD)
   - Description: Most adaptable, prompt-responsive
   - Use: Artistic concepts, stylized work
   - Good: Surreal ideas, artistic interpretation
   - Tested: 2026-06-13 NEONCYBER batch - good results

3. **@cf/leonardo/phoenix-1.0** (TO TEST)
   - Description: Exceptional prompt adherence, coherent text
   - Use: Complex prompts, text-in-image
   - Status: Not tested yet

4. **@cf/black-forest-labs/flux-2-klein-9b** (TO TEST)
   - Description: Ultra-fast, enhanced quality, distilled
   - Use: Speed + quality balance
   - Status: Not tested

5. **@cf/black-forest-labs/flux-2-klein-4b** (TO TEST)
   - Description: Ultra-fast, distilled
   - Use: Maximum speed
   - Status: Not tested

6. **@cf/bytedance/stable-diffusion-xl-lightning** (TO TEST)
   - Description: Lightning fast SDXL
   - Use: Rapid SDXL generation
   - Status: Not tested

7. **@cf/stability/stable-diffusion-xl-base-1.0** (TO TEST)
   - Description: Classic SDXL
   - Use: Reliable SDXL quality
   - Status: Not tested

## Current Status (2026-06-13)

**Daily Limit Reached:** Cloudflare free tier (10,000 neurons) exhausted
- Next generation available: ~24 hours after last use
- Last generation: 2026-06-13 18:35 UTC

## Testing Plan

### Phase 1: Model Comparison (when limit resets)
Test each model with same 3 prompts:

1. **Simple minimal:** Black sphere in void
2. **Complex surreal:** Arcade cabinet dissolving into digital noise
3. **Atmospheric:** Rain-soaked cyberpunk alley

### Phase 2: Best Model Selection
- Compare quality, prompt adherence, style
- Select best model for each series:
  - NEONCYBER: Artistic, atmospheric, surreal
  - IMPOSSIBLE-GEO: Geometric, precise, minimal

### Phase 3: Batch Generation
- Generate full batches using selected models
- Build portfolio for Steam Workshop

## Model Selection by Series

### NEONCYBER Recommendations
1. **@cf/leonardo/lucid-origin** (TESTED - GOOD)
   - Best for artistic interpretation
   - Handles surreal concepts well
   - Good for atmospheric cyberpunk

2. **@cf/leonardo/phoenix-1.0** (TO TEST)
   - Could be better for complex prompts
   - Good prompt adherence

### IMPOSSIBLE-GEO Recommendations
1. **@cf/black-forest-labs/flux-1-schnell** (CURRENT)
   - Decent for geometric concepts
   - Fast for iterations

2. **@cf/leonardo/lucid-origin** (TO TEST)
   - Might handle impossible geometry better
   - Could improve quality

## Testing Commands

```bash
# Test Phoenix 1.0
export CF_WP_MODEL="@cf/leonardo/phoenix-1.0"
bash /workspace/group/shared/scripts/cf-generate-wp.sh "[prompt]" [output] [prefix] 8

# Test FLUX.2-klein-9b
export CF_WP_MODEL="@cf/black-forest-labs/flux-2-klein-9b"
bash /workspace/group/shared/scripts/cf-generate-wp.sh "[prompt]" [output] [prefix] 8

# Test SDXL Lightning
export CF_WP_MODEL="@cf/bytedance/stable-diffusion-xl-lightning"
bash /workspace/group/shared/scripts/cf-generate-wp.sh "[prompt]" [output] [prefix] 8
```

## Next Steps

1. **Wait for limit reset** (~24 hours)
2. **Test Phoenix 1.0** with 3 comparison prompts
3. **Test FLUX.2-klein models** for speed/quality balance
4. **Select best models** for each series
5. **Generate full batches** with selected models
6. **Curate portfolio** for Steam Workshop

## Notes

- Lucid Origin showed good results for NEONCYBER (2026-06-13)
- FLUX.1-schnell is decent but struggles with surreal concepts
- Phoenix 1.0 might be best for complex prompt adherence
- Consider upgrading Cloudflare plan for more daily generations