ComfyUI quantized acceleration
last edited
ComfyUI is a node-graph tool for local image and video generation. Quantized acceleration, weight streaming, and ROCm support decide which models fit on a given GPU. These features link to Local Generation UIs and to the wider ComfyUI core project.
Precision formats
NVIDIA announced a ComfyUI optimization at CES in January. With NVFP4 on RTX cards, it gave up to 3x faster inference. VRAM use fell by 60%. FP8 reached 2x. These figures combined the formats with CUDA software optimizations.
A later NVIDIA update for current RTX cards gave 2.5x for NVFP4. That came with 60% less VRAM. FP8 gave 1.7x. It also cut VRAM by 40%.
ComfyUI supports NVFP4 and FP8 natively. Listed models include LTX-2, Qwen-Image, Z-Image, and FLUX.1. FLUX.2 is also listed. The Klein variant comes in 4B and 9B sizes.
NVFP4 acceleration needs PyTorch built with CUDA 13.0. Community reports describe sampling up to 2x slower than fp8 without it.
Memory and generation strategies
Weight streaming moves model weights into system memory when GPU memory runs short. Larger checkpoints then run on smaller cards, though repeated transfers can slow generation.
Consistency models cut compute through single-step generation instead. Video raises memory needs because of multiple frames. Pipelines such as AnimateDiff depend on careful memory management, and Optimizers also shape how much memory training and tuning need.
Official AMD ROCm support began on January 5 with desktop release v0.7.0. It uses ROCm 7.1.1 and covers portable and git installs. Compatible macOS computers use the MPS backend.
Workflow ecosystem
Subgraphs collapse node combinations into reusable nodes that creators can publish to the node library. A template library organizes workflows by creative task. The desktop app was rebuilt as a management tool and reached all users on June 8.
Version v0.26.0 arrived on June 18. It added a video model and a job cancellation API. Version v0.40.0 followed on October 13. It added 3D asset export and camera info nodes.
Guidance circulates through YouTube creators and educational platforms. Extensions from node creators combine quantized checkpoints and weight streaming. Operational limits appear in ComfyUI production infrastructure.
Sources
- NVIDIA RTX AI Garage CES post gives the NVFP4 and FP8 figures from January.
- NVIDIA RTX AI Garage GDC post on FLUX.2 Klein covers native NVFP4 support.
- ComfyUI release history on GitHub records versions and ROCm support.