UniGen - Universal Generation Space
Image & Video Generation with multiple model families, variants, and full parameter control. Running on ZeroGPU (RTX PRO 6000 Blackwell).
| Tab | Sub-tabs | Model Families |
|---|---|---|
| Image | Text to Image, Image to Image | SDXL, SD1.5, SD3, FLUX, PixArt, Kolors, Playground |
| Video | Text to Video, Image to Video | Wan 2.1/2.2, LTX 2.3/2.5, CogVideoX, HunyuanVideo, Mochi, SVD |
Model Series
Model Variant
1 100
0 20
256 2048
256 2048
30 300
Model Series
Model Variant
1 100
0 20
256 2048
256 2048
30 300
0 1
Model Series
Model Variant
1 100
0 20
0 20
8 200
6 60
256 1280
256 1280
30 300
Scheduler
0.5 15
Model Series
Model Variant
1 100
0 20
0 20
8 200
6 60
256 1280
256 1280
30 300
Scheduler
0.5 15
Tips
- Seed -1 = random seed each generation
- ZeroGPU Duration: Set per-model GPU time quota. Lower = saves quota!
- Wan 2.2 Lightning / LTX Turbo: 45-60s
- Wan 2.1/2.2 / LTX: 90-120s
- CogVideoX: 120-180s
- HunyuanVideo: 180-240s
- Models load on CPU when clicking "Load Model" - no GPU quota used!
- GPU allocated only on generation via @spaces.GPU decorator
- Turbo/Lightning/Schnell: fewer steps (1-4), CFG=0
- Wan 2.2 Lightning (I2V-A14B): 4-8 steps, flow_shift=6.0
- LTX Distilled/Turbo: 4-8 steps, sampler: euler_ancestral
- LTX 10Eros: 12-25 steps for best quality
- Scheduler/Sampler selection for Wan and LTX
- Click Load Model (CPU) when switching models/variants
Adding New Models
Edit MODEL_REGISTRY in app.py to add new series/variants.
Add config to MODEL_CONFIG (video) or IMAGE_MODEL_CONFIG (image) for custom params including gpu_duration.