Load Diffusion Model (Downloaded)

The modern way to ship a model is split files: a standalone diffusion model, a separate text encoder, a separate VAE. This node is the piece of ComfyUI Model Batch Downloader that loads the first of those - the diffusion_models category - into your graph as a MODEL, ready to feed a sampler. It's the downloaded-file cousin of ComfyUI's Load Diffusion Model (UNETLoader), and like that node it's a thin wrapper: it finds the file you downloaded, then calls the core loader on it.

Why you'd reach for it

Because more and more model families don't give you an all-in-one checkpoint. The README's Anima example is the template: download the diffusion model into diffusion_models, the text encoder into text_encoders, and the VAE into vae, then load each with its matching "(Downloaded)" loader. The diffusion model is the heavy one - for big modern bases that's the file that decides whether your workflow fits in VRAM.

That brings us to the input that makes this node more interesting than it looks.

Inputs and outputs

Two required inputs, same shape as the other loaders:

• download_result - from either download node's download_result output.
• id - the manifest ID of the file you want; defaults to the filename without its extension if you didn't set one.

Plus weight_dtype, a dropdown with four choices:

• default - whatever the file ships with.
• fp8_e4m3fn / fp8_e4m3fn_fast - FP8 quantization, roughly half the VRAM with near-zero quality loss by community consensus.
• fp8_e5m2 - the looser FP8 variant, lower precision, used when the stricter one doesn't fit.

Output is a single MODEL, which wires straight into your sampler's model input. If your workflow uses a LoRA, the downloaded-file chain looks like: this loader → Load LoRA (Downloaded) → sampler.

On the weight_dtype choice: if you're on a 6–12 GB card and the model's fp16 won't fit, fp8_e4m3fn is the first thing to try - it's the standard "just use it" setting for big Flux-class models, not a compromise of last resort. Leave it on default if you're not fighting for VRAM and the model's native format is already what you want.

The guard that saves you

Like every loader in the pack, this one validates that the id you gave it has model_type: diffusion_models in the manifest. Feed it a checkpoints ID and it errors out with a message telling you the mismatch, instead of failing weirdly mid-graph. It's the pack being annoyingly correct, and you'll appreciate it exactly once when you typo a category.

Install and troubleshooting

Same pack, same steps: ComfyUI Manager search for "ComfyUI Model Batch Downloader", or clone into custom_nodes, restart. No Python deps; the only external requirement is aria2c on PATH for the download half.

Real-world failures are the pack's usual set - an ID that doesn't exist in the manifest (spelling, or the download failed), and the category-mismatch error above. One subtle one worth knowing: if the download was interrupted, the download_result records what happened, and the loader will tell you the ID isn't there. Re-run the download node - it resumes partial files via the .aria2 sidecar rather than starting over.