A high-performance custom node for LTX2 video generation in ComfyUI, designed to drastically reduce GPU usage, VRAM requirements, and temperature on consumer GPUs (especially 6GB cards like RTX 2060).

✨ Features

🎯 Core Optimizations

Frame-Stride Diffusion: Process only keyframes (e.g., 1 out of 4), interpolate the rest
Spatial Attention Freezing: Skip redundant attention computation after configurable threshold
LTX2-Specific Hooks: Precise targeting of attn1 (self-attention) vs attn2 (cross-attention)

🌡️ Thermal Management

Real-Time GPU Monitoring: Uses pynvml to read actual GPU temperature
Dynamic Throttling: Auto-adjusts processing speed to maintain target temperature
Emergency Protection: Automatic 2-second pause at critical temps (83°C+)
Preset Profiles: Quick optimization for different GPU tiers

🎬 Advanced Interpolation

Linear: Fast, basic blending
Slerp: Spherical interpolation - smoother for diffusion latents
Motion-Compensated: Optical flow-based warping for best motion quality

📦 Additional Nodes

LTX2 Temporal VAE Decode: Memory-efficient VAE decoding with temporal tiling
LTX2 Separate Audio/Video Latent: Extract video and audio from combined latents (for use with LTXVConcatAVLatent)
LTX2 Combine Audio/Video Latent: Recombine processed video with audio for downstream nodes
LTX2 Empty Audio Latent: Generate placeholder audio latent matching video duration

🎵 Audio-Video Workflow

For workflows using LTXVConcatAVLatent or MMAudio:

[LTXVConcatAVLatent] → [LTX2 Separate AV] → [LTX2 Efficient Sampler] → [LTX2 Combine AV] → [Decode]
                                         ↓                              ↑
                                  [Audio Pipeline (optional)] ──────────┘

📋 Requirements

ComfyUI (latest)
PyTorch with CUDA
pynvml (optional, for thermal monitoring)

🚀 Installation

Option 1: ComfyUI Manager (Recommended)

Search for "LTX2 Efficient" in ComfyUI Manager and install.

Option 2: Manual Installation

cd ComfyUI/custom_nodes/
git clone https://github.com/kakachiex2/LTX2-Efficient-Sampler
cd comfyui-ltx2-efficient
pip install -r requirements.txt

🎛️ Node Settings

Optimization Presets

| Preset | Freeze Ratio | Throttle | Frame Stride | Best For | | --------------------------- | ------------ | -------- | ------------ | --------------- | | Performance (RTX 3080+) | 70% | 0ms | 2 | High-end GPUs | | Balanced (RTX 3060/3070) | 50% | 25ms | 4 | Mid-range | | Aggressive (RTX 2060/2070) | 30% | 50ms | 4 | 6GB VRAM | | Ultra Low Power (GTX 1660) | 15% | 100ms | 6 | Older GPUs | | Extreme Cool (RTX 2060 6GB) | 10% | 150ms | 8 | Target 65°C | | Thermal Auto-Scale | 30% | Dynamic | 4 | Recommended | | Custom | Manual | Manual | Manual | Fine-tuning |

Key Parameters

| Parameter | Default | Description | | ---------------------- | ------------------ | ------------------------------------------- | | optimization_preset | Thermal Auto-Scale | Quick optimization profile | | target_temp | 70°C | Target GPU temperature (for auto-scale) | | frame_stride | 4 | Process every Nth frame (higher = faster) | | freeze_ratio | 0.3 | Start freezing attention at this % of steps | | attention_window | 4 | Temporal attention window size | | interpolation_method | slerp | linear, slerp, or motion |

Interpolation Methods

| Method | Speed | Quality | Best For | | -------- | --------- | ------- | -------------------------------- | | linear | ⚡ Fast | Basic | Quick previews | | slerp | ⚡ Fast | Smooth | Default - diffusion latents | | motion | 🐢 Slower | Best | Videos with significant movement |

📖 Usage

Add LTX2 Efficient Video Sampler node
Connect your model, latent video, and conditioning
Select an Optimization Preset (start with "Thermal Auto-Scale")
Set Target Temp to your comfort level (65-70°C recommended)
Choose Interpolation Method (slerp is a good default)
Run and watch console for temperature feedback!

Console Output

[LTX2Efficient] Using preset: Thermal Auto-Scale
[LTX2Efficient] -> THERMAL AUTO-SCALE enabled, target_temp=70°C
[LTX2Efficient] GPU Monitor active. Current temp: 47°C
[LTX2Efficient] Patched 96 attention layers (attn1=48, attn2=48)
[LTX2Efficient] Step 5: Temp=52°C, Delay=0ms
[LTX2Efficient] Freezing spatial at step 6 (30.0%)
[LTX2Efficient] Interpolating with method: slerp

🔧 Troubleshooting

"pynvml not installed" warning

pip install pynvml

Or install nvidia-ml-py (pynvml is deprecated but still works).

Still running hot?

Lower target_temp to 60-65
Use "Extreme Cool" preset
Increase frame_stride to 6-8
Lower video resolution/frame count

Quality issues?

Try interpolation_method: motion
Lower frame_stride to 2-3
Increase freeze_ratio to 0.5+

🏗️ Architecture

comfyui-ltx2-efficient/
├── __init__.py              # Node registration
├── nodes/
│   ├── ltx2_efficient_sampler.py  # Main sampler node
│   ├── ltx2_vae_decode.py         # Temporal VAE decode
│   ├── gpu_monitor.py             # pynvml temperature monitoring
│   └── interpolation.py           # Linear/Slerp/Motion interpolation
├── frontend/                 # Vue UI components (WIP)
└── Reference/               # Technical documentation

📄 License

MIT License - see LICENSE for details.

🙏 Credits

ComfyUI - The amazing UI framework
Lightricks LTX-Video - The video model
Inspired by research on efficient video diffusion

🤝 Contributing

Contributions welcome! Please open an issue or PR.

Made with ❤️ for the ComfyUI community