Using the custom nodes
A workflow stays the same once the nodes are installed. You keep the same sampler, steps, prompts, and connections. Only the loader nodes change.
In place of the standard loaders, use these nodes from the pack:
- UNet Loader (Dynamic VRAM) or Unet Loader (GGUF) to load the diffusion model
- CLIP Loader (Dynamic VRAM) or CLIP Loader (GGUF) to load the text encoder (Variations exist also for the Double/Triple/Quadruple CLIP Loader)
- VAE Loader (GGUF) to load the VAE
Default to Dynamic VRAM
Pick the Dynamic VRAM option by default. Use the plain GGUF loaders when a workflow needs the classic behaviour.
What is Dynamic VRAM
Dynamic VRAM is ComfyUI's system for managing GPU and system RAM during generation. Instead of pre-allocating memory for a whole model, it loads and offloads model layers as each step needs them.
This replaces rigid memory allocation with a flexible approach. Weights move in and out of VRAM layer by layer, so the process avoids slow SSD swapping while using as much of the available VRAM as possible. Large diffusion and video models can run on modest hardware this way.
Because model weights and LoRAs are handled more efficiently, out-of-memory errors become less common and switching between models is faster. The added layer management has a small cost: users with high-end GPUs running smaller workflows may see a slight performance hit.
Native quant execution
The loader nodes also pick the execution path for custom quants. A Q8_CR file runs on ComfyUI's native INT8 path, and a Q4_CR file runs on a int4 tensor-core path. Weights stay at their quantized bit depth during inference, so no dequantize step runs in the matmul. For format details and requirements, see Quant Formats.