Work great with NVFP4 - can't wait for the R2V lora
Amazing job. This work great with NVFP4 model on my 4090 card and my custom workflow where i only use custom node.
And, yes, it work well with R2V. But i can't wait for the dedicated R2V Lightx2v lora.
Thank you for this lora.
Amazing job. This work great with NVFP4 model on my 4090 card and my custom workflow where i only use custom node.
And, yes, it work well with R2V. But i can't wait for the dedicated R2V Lightx2v lora.
Thank you for this lora.
what do you mean/? it works already for REFERENCE to video
Amazing job. This work great with NVFP4 model on my 4090 card and my custom workflow where i only use custom node.
And, yes, it work well with R2V. But i can't wait for the dedicated R2V Lightx2v lora.
Thank you for this lora.
Hmm. NVFP4 requires blackwell (rtx 5000+), so something is not right here.
Amazing job. This work great with NVFP4 model on my 4090 card and my custom workflow where i only use custom node.
And, yes, it work well with R2V. But i can't wait for the dedicated R2V Lightx2v lora.
Thank you for this lora.what do you mean/? it works already for REFERENCE to video
Yes, it work for R2V, not as good as the FL2V, but at 4 step it works :p
Amazing job. This work great with NVFP4 model on my 4090 card and my custom workflow where i only use custom node.
And, yes, it work well with R2V. But i can't wait for the dedicated R2V Lightx2v lora.
Thank you for this lora.Hmm. NVFP4 requires blackwell (rtx 5000+), so something is not right here.
The 4090 isn’t executing native NVFP4 instructions; ComfyUI is transparently emulating them.
So ComfyUI keeps the compact NVFP4 model representation, then dequantizes/casts operations to BF16 as needed. That provides compatibility and potentially lower storage/VRAM usage, but not Blackwell’s native FP4 acceleration.
With NBFP4 on my 4090, it take 23 seconds per step. A 15s video at Preview resolution 864x480, it take arround 2mn.
Here my workflow for H3. I only make and use my custom node pack for all models 😉
Native NVFP4 matrix multiplication requires compute capability ≥10.0—meaning Blackwell-class hardware—according to https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/examples/fp8_primer.html
So the verdic is:
NVFP4 model works on RTX 4090 ✅
Native NVFP4 acceleration on RTX 4090 ❌
Clever ComfyUI emulation/fallback ✅
So, yes NVFP4 work on 4090 and can also work on other nvidia card. Only the native NVFP4 acceleration only work on Blackwell cards.
And I tested all models for H3, and the NVFP4 is the one that give me the best results, and the best speed 😃
@pat11 hi bro, there are so many NVFP4 model for H3, which one are you using? can you send me the download link H3 nvfp4?
Thank you