Instant Load Diffusion Model

Modern workflows rarely load full checkpoints. For Flux, SD3, and most video models you load the diffusion model (the thing that actually denoises latents), the text encoders, and the VAE as separate files - and the diffusion model is the big one. Instant Load Diffusion Model is the pack's drop-in for ComfyUI's UNETLoader: same dropdowns, same MODEL output, same models/diffusion_models folder. The difference is that .safetensors weights stream straight into the model instead of sitting in a temporary state dict first - a real headroom win when the file you're loading is 12GB plus.

The two inputs

• unet_name - dropdown over everything in ComfyUI/models/diffusion_models. Same list as the stock node.
• weight_dtype - how the model's weights are stored, with four options: default, fp8_e4m3fn, fp8_e4m3fn_fast, and fp8_e5m2. This is the one people actually deliberate over.

On the fp8 question: fp8_e4m3fn is what everyone means when they say "just fp8 it" - half the VRAM of fp16 with a quality cost you'll struggle to see, and RTX 40-series and newer cards run the math natively. fp8_e4m3fn_fast is the same cast plus ComfyUI's fp8 optimizations for native compute. fp8_e5m2 is the less precise variant; you rarely want it. default keeps whatever dtype the file was saved in. For a Flux-scale model, fp8 is often the difference between fitting on 12–16GB and not fitting at all. If it fits, plain fp8 is the sensible default; only drop to default when you're chasing the last few percent of quality and have the VRAM.

Output

• MODEL - wire it into your sampler (and, for Flux-style graphs, into the DualCLIPLoader's CLIP pipeline and the VAE decoder).

Common issues

Same pack rules as the rest of this family: CUDA-only (CPU-torch ComfyUI errors immediately), instanttensor must be installed or you get "The instanttensor package is required for CUDA instant loading," and only .safetensors files get the streaming path - .gguf and other formats fall through to ComfyUI's normal loader. One thing to know: if the streaming path can't handle a particular safetensors diffusion model, the node raises an error rather than quietly doubling memory. That's by design - the fallback would recreate the peak-memory problem the node exists to fix.

How to install

ComfyUI Manager → search ComfyUI-InstantTensorLoaders → install, accepting the dependency install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/redstonewhite/ComfyUI-InstantTensorLoaders

Restart ComfyUI; the node is under advanced/loaders/instanttensor. If the dependency missed your environment, pip install instanttensor in the venv ComfyUI runs from and restart.