How to Generate 4K AI Video Locally on RTX GPUs with ComfyUI and RTX Video Super Resolution
TL;DR
- Install the latest ComfyUI with App View for a simplified prompt-to-video interface and load NVFP4/FP8 versions of FLUX.2 Klein or LTX-2.3 for up to 2.5x faster generation and 60% less VRAM.
- Generate video at lower resolution then upscale instantly to 4K using the new RTX Video Super Resolution node (30x faster than traditional upscalers).
- Switch between App View and Node View seamlessly while benefiting from native RTX optimizations already built into ComfyUI.
Game developers, concept artists, and creators can now produce high-quality AI video entirely locally on NVIDIA RTX GPUs. At GDC, NVIDIA and the ComfyUI team released updates that dramatically lower the barrier for cinematic video generation and 4K upscaling directly on your RTX AI PC or DGX Spark desktop supercomputer.
Prerequisites
Before starting, ensure you have:
- An NVIDIA GeForce RTX 40 Series or RTX 50 Series GPU (RTX 5090 recommended for best performance)
- Latest NVIDIA drivers (Studio Driver preferred for creative workloads)
- At least 12 GB VRAM (24 GB+ ideal for 4K workflows)
- Python 3.10+ environment
- Git installed
- Hugging Face account (for downloading model checkpoints)
Step 1: Install the Latest ComfyUI with RTX Optimizations
ComfyUI now ships with significant RTX performance improvements: 40% faster since September, native NVFP4 and FP8 support, and a new App View interface.
- Clone the official ComfyUI repository:
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
- Create and activate a virtual environment:
python -m venv venv
venv\Scripts\activate # Windows
- Install PyTorch with CUDA support (use the latest stable version compatible with your driver):
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
- Install ComfyUI dependencies:
pip install -r requirements.txt
- Install the NVIDIA RTX optimizations and custom nodes (check official ComfyUI RTX fork or manager for latest):
pip install nvidia-pyindex
# Additional RTX nodes via ComfyUI Manager (recommended)
Launch ComfyUI:
python main.py
Open http://127.0.0.1:8188 in your browser.
Step 2: Download and Load NVFP4 / FP8 Models
The biggest performance gains come from the new quantized models.
Available models right now:
- FLUX.2 Klein 4B and 9B (NVFP4 and FP8)
- LTX-2.3 (NVFP4 support coming very soon)
Download the checkpoints from Hugging Face:
- Search for “FLUX.2 Klein NVFP4” or “FP8” on Hugging Face
- Place the
.safetensorsfiles inComfyUI/models/unet/orComfyUI/models/checkpoints/
Load a model in ComfyUI:
- Open the Template Browser
- Load a default video workflow (LTX-2 or FLUX video template)
- Double-click the model loader node
- Select your newly downloaded NVFP4 or FP8 checkpoint
- Save the workflow
Expected results on an RTX 5090:
- LTX-2: 512×768, 100 frames, 20 steps — significantly faster
- FLUX.2 Klein 9B: 1024×1024, 20 steps — up to 2.5x speedup and 60% less VRAM with NVFP4
Step 3: Use the New App View for Simplified Video Generation
Artists unfamiliar with node graphs can now use the streamlined interface.
- In ComfyUI, click the App View toggle (top right)
- You will see a clean prompt box and basic sliders instead of the node graph
- Enter your video prompt (e.g., “cinematic shot of a cyberpunk street at night, raining, neon lights”)
- Adjust parameters:
- Resolution (start with 512×768 or 768×512 for speed)
- Frames (16–100)
- Steps (20–30)
- Guidance scale
- Click Generate
The full Node View remains one click away for advanced customization. You can switch back and forth without losing your workflow.
Step 4: Add RTX Video Super Resolution Node for 4K Upscaling
This is the highlight for game developers and storyboarding artists.
Option A: Using the ComfyUI Node (Recommended)
- Install the RTX Video node via ComfyUI Manager or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/NVIDIA/ComfyUI-RTX-Video-Super-Resolution # placeholder — check official repo
- In your workflow, add the RTX Video Super Resolution node after your video generation node
- Connect the generated video output to the upscaler input
- Set output resolution to 4K (3840×2160)
- Run the node
The upscaler runs on RTX GPU Tensor Cores and is 30x faster than popular local upscalers while using far less VRAM. A 10-second clip that previously took minutes now finishes in seconds.
Option B: For Developers — Use the Python Package
Install the official package:
pip install nvidia-rtx-video-upscaler
Sample usage (simplified):
from nvidia.rtx_video import Upscaler
upscaler = Upscaler(model="rtx_video_sr_4x")
upscaled_frames = upscaler.upscale(low_res_frames, target_res=(3840, 2160))
Full sample code and VFX Python bindings are available on GitHub (see Sources).
Tips and Best Practices
- Start at lower resolution (512×768) and upscale to 4K — this is much faster and more memory efficient.
- Use NVFP4 models on RTX 50 Series for maximum performance.
- Keep workflows under 100 frames for real-time iteration during concepting.
- Save workflows in both App View and Node View so you can quickly iterate.
- Combine with LTX Desktop (open-source video editor) for local post-production.
- Monitor VRAM usage — the new quantized models reduce memory by up to 60%.
Common Issues
Why am I getting error X?
“Out of Memory” or CUDA OOM
Use FP8 or NVFP4 models, lower the resolution or number of frames, or enable --lowvram flag when launching ComfyUI.
Model not found after downloading
Make sure the checkpoint is in the correct folder (models/unet or models/checkpoints) and refresh the model list in ComfyUI.
RTX Video node missing
Install via ComfyUI Manager or clone the official RTX Video node repository. Restart ComfyUI after installation.
Slow performance on RTX 40 Series
FP8 still gives 1.7x speedup and 40% less VRAM. NVFP4 is optimal on RTX 50 Series but FP8 works well on previous generations.
App View not appearing
Update to the latest ComfyUI version. The feature was released at GDC and may require pulling the newest main branch.
Next Steps
After mastering basic video generation:
- Explore the NVIDIA Studio Sessions tutorial by visual effects artist Max Novak
- Attend the GTC session “Create Generative AI Workflow for Design and Visualization in ComfyUI” on March 17
- Try multi-GPU setups with the new LTX-2 Multi-GPU node for even longer videos
- Integrate generated 4K assets into your game engine pipelines
- Experiment with LM Link to offload heavier models to a DGX Spark while controlling from your laptop
Local AI video generation has never been more accessible for game developers and creators. With these tools, you can iterate on cinematic concepts and storyboards in minutes instead of hours — all while keeping your work private and offline.