README.md

metadata

title: UltraPixel Multi-Stage (Community Fixed)
emoji: 🎨
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: apache-2.0

🎨 UltraPixel Multi-Stage Generator (Community Fixed)

A properly working UltraPixel-style high-resolution image generator that actually respects your parameter inputs.

What's Different From Original UltraPixel Spaces?

The original public UltraPixel spaces have a critical flaw - they hardcode CFG and timesteps inside the generation function, making the UI sliders meaningless:

# Original broken code:
extras.sampling_configs['cfg'] = 4          # ← Always uses 4!
extras.sampling_configs['timesteps'] = 20   # ← Ignores your slider!

This Space Fixes That ✅

• Real CFG Control: Your slider values are actually passed to the model
• Real Steps Control: Set your own timesteps (10-100) per stage
• Memory Optimized: Won't OOM on ZeroGPU (max 3072×3072 with tiling)
• No Login Required: Public access for easy testing

Features

• 🎯 3-Stage Pipeline: Stable Cascade architecture (Stage C → B → A)
• 🔧 Independent Controls: Separate CFG/steps for each stage
• 💾 Memory Safe: Aggressive cleanup between stages, forced tiling
• ⏱️ 120s Per Stage: Each stage gets fresh GPU allocation
• 🔓 Public Access: No authentication needed

How to Use

Standard Workflow (3-4 minutes total)

1. Stage C - Generate Initial Latent (~30-60s)

   • Enter your prompt
   • Set CFG (recommended: 7.5) and Steps (recommended: 30)
   • Click "Generate Stage C"
   • Wait for completion

2. Wait for GPU availability (if needed during high traffic)

3. Stage B - Upscale Latent (~30-50s)

   • Adjust CFG (recommended: 5.0) and Steps (recommended: 15)
   • Click "Generate Stage B"
   • Uses the latent from Stage C automatically

4. Wait again if needed

5. Stage A - Final Decode (~60-90s)

   • Keep "Use Tiling" checked (prevents OOM)
   • Click "Generate Final Image"
   • Download your high-res result!

Optimal Settings 💡

• Stage C: CFG 7-8, Steps 30-40
• Stage B: CFG 4-6, Steps 15-20
• Stage A: Always use tiling
• Resolution Limits: Max 3072×3072 for stability (1536×1536 per stage C/B)
• For Training Data: Generate at 3072px, then downscale to 1024px for optimal quality

Technical Details

Memory Management

Each stage runs in isolated @spaces.GPU(duration=120) calls:

• Models loaded only when needed
• Aggressive torch.cuda.empty_cache() after each stage
• Latents stored in-memory (automatically cleaned after 1 hour)
• VAE tiling enabled for Stage A decode

Resolution Scaling

• Stage C Input: 512-1536px (base resolution)
• Stage B Output: 2× Stage C (1024-3072px)
• Stage A Output: Full decode to target resolution
• Memory Usage: ~20-30GB peak per stage (safe for ZeroGPU)

Why This Matters

Many public AI spaces claim to offer "full control" but secretly override your parameters. This leads to:

• ❌ Inconsistent results despite changing settings
• ❌ Users wasting time tweaking sliders that do nothing
• ❌ Frustration when trying to reproduce outputs

This space guarantees that your inputs = actual model parameters.

Deployment Notes

Built specifically for:

• ZeroGPU compatibility (120s duration per stage)
• Public/unlogged access
• High-resolution output (up to 3072×3072 stable)
• Proper parameter control

Credits

• Stable Cascade: Stability AI
• Original UltraPixel Concept: Various community implementations
• This Implementation: Community-fixed version with proper parameter control

License

Apache 2.0