Kkaemii/fuzz-style-ltxv-lora

LTX-Video LoRA Collection

🎨 Available Styles

FUZZ_STYLE

A unique artistic style LoRA trained for LTX-Video generation.

• rank32/: Lightweight, 192MB per checkpoint - Fast inference, lower VRAM usage
• rank64/: Balanced, 384MB per checkpoint - Good quality/performance balance
• rank128/: High quality, 768MB per checkpoint - Maximum quality output

TEXT2CUTS_STYLE

Text-to-video style LoRA optimized for dynamic scene transitions.

• rank32/: Lightweight, 192MB per checkpoint - Fast inference, lower VRAM usage
• rank64/: Balanced, 384MB per checkpoint - Good quality/performance balance
• rank128/: High quality, 768MB per checkpoint - Maximum quality output

📊 Training Steps Available

Each rank includes models trained at different steps:

• 500 steps (FUZZ_STYLE only - initial checkpoint)
• 1500 steps - Early convergence
• 3000 steps - Balanced training
• 4500 steps - Well-trained
• 6000 steps - Maximum training

🚀 Usage with ComfyUI

1. Download the desired LoRA file(s)
2. Place in ComfyUI/models/loras/ directory
3. Use the LoRA Loader node in your workflow
4. Recommended LoRA strength: 0.8-1.0

📦 File Structure

├── fuzz_style/
│   ├── rank32/
│   │   ├── FUZZ_STYLE_rank32_500steps_FIRST.safetensors
│   │   ├── FUZZ_STYLE_rank32_1500steps.safetensors
│   │   ├── FUZZ_STYLE_rank32_3000steps.safetensors
│   │   ├── FUZZ_STYLE_rank32_4500steps.safetensors
│   │   └── FUZZ_STYLE_rank32_6000steps.safetensors
│   ├── rank64/
│   │   └── [5 files]
│   └── rank128/
│       └── [5 files]
└── text2cuts_style/
    ├── rank32/
    │   └── [4 files - 1500, 3000, 4500, 6000 steps]
    ├── rank64/
    │   └── [4 files]
    └── rank128/
        └── [4 files]

💡 Recommendations

• For testing: Start with rank32 at 3000 steps
• For quality: Use rank64 or rank128 at 4500-6000 steps
• For speed: Use rank32 with lower steps

🔧 Technical Details

• Base Model: LTX-Video
• Format: safetensors
• Precision: bf16
• Compatible with ComfyUI (with patched lora.py for LTXV support)

📝 Notes

• All LoRAs are converted to ComfyUI format (diffusion_model. prefix)
• Tested with ComfyUI and LTX-Video workflows