Innovate Futures @ Benji
patreon
ComfyUI WaveSpeed Hunyuan Video - Optimize for Any Local AI Video Generate!
🕑 Added 2025-01-10 14:30:01 +0000 UTC
Video : https://youtu.be/wPiujsfr__Y
In this video, I’ll show you how to optimize your AI generation processing speed using Comfy WaveSpeed, a powerful set of custom nodes for ComfyUI. Whether you're working with image diffusion models like Flux or AI video models like Hunyuan Videos and LTX, this tool can significantly improve memory handling and speed up your workflow.
Resources & Links:
GitHub Project: https://github.com/chengzeyi/Comfy-WaveSpeed
Hunyuan Video Workflow Updates: https://www.patreon.com/posts/hunyuan-video-v2-119704026?utm_source=yt&utm_medium=vid&utm_campaign=20250110
Hunyuan Video GGUF Installation Guide: https://youtu.be/Q3WKoT_pLlE
Hunyuan Video Lora Specific Character Style https://youtu.be/i0AhvXihF74
What You’ll Learn:
How to install and set up Comfy WaveSpeed for ComfyUI.
Tips for optimizing memory usage with PyTorch and CUDA.
Step-by-step guide to using GGUF quantization and model caching for faster generation times.
Real-time examples with Hunyuan Videos and Flux to show the performance improvements.
Why This Matters:
If you're tired of long generation times and memory bottlenecks, Comfy WaveSpeed is here to help. It’s perfect for anyone working with AI video generation, image diffusion, or complex workflows. Plus, I’ll show you how to use Purge VRAM to keep your system running smoothly.
Comments
Steve Kiraly
I'm missing the Cuda toolkit and Visual Studio and probably cuDNN. I'll get back to you after I install those. I also should mention I get the following warnings when starting ComfyUI: WARNING: Failed to find MVSC. WARNING: Failed to find Windows SDK. WARNING: Failed to find CUDA. I think installing the above might fix all this. Don't really know what to install from Visual Studio, which must go before the Cuda Toolkit.
Steve Kiraly
Hrm. I have tried both versions of Triton for 311. I have a 3090. Can you tell me where I would set TORCH_LOGS and TORCHDYNAMO_VERBOSE? !!! Exception during processing !!! backend='inductor' raised: ImportError: DLL load failed while importing libtriton: A dynamic link library (DLL) initialization routine failed. Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information ...I wonder if I could be missing a requirement. Gemini said, "To run Triton Inference Server on Windows, you need a system with a compatible NVIDIA GPU, CUDA toolkit, cuDNN, and the appropriate Visual Studio Build Tools, essentially requiring a Windows system with the ability to compile C++ code, and the necessary dependencies outlined in the Windows Dockerfile for Triton installation; this includes Python 3, CUDA, cuDNN, and potentially TensorRT depending on your model requirements."
Benjamin Law
in 4090 I set 512X240 , I am able to get up to 16X length. But after I apply First Block Cache, I couldn't do 160 any more, I have to set length to 129. I think the cache do take some space when execute, in theory.
V V
i keep getting OOM. if residual is 0.0 it does nothing, so i set to 0.035 per the ParaAttention docs but i cant get it to work on low res 70 length on 4090. i have lora load after model load and before block cache. any tips?
