Guided Diffusion Loader

Before Stable Diffusion, there were no CivitAI checkpoints. There was Disco Diffusion, and it ran on a handful of old OpenAI "guided diffusion" models - the same family as GLIDE's ImageNet checkpoints, released in 2021. This node loads those. If you've never seen a diffusion model that isn't a latent-diffusion model, this is the door to that world, and it's the only way the Disco Diffusion Sampler gets anything to work with.

What it loads

A dropdown of 27 models, and they're a time capsule. The default is 512x512_diffusion_uncond_finetune_008100 - the big workhorse that produced most of the painterly Disco Diffusion images you've seen. Around it sit the 256x256_diffusion_uncond base, a couple of concept-art generators, four portrait generators, and a whole shelf of fine-tunes from the era: pixel art (hard, soft, expanded, and a 4K variant), watercolor, PulpSciFi, Medieval, Lithography, Floral, Textile, Isometric, and a couple of others.

These are not SD checkpoints. They're plain-image-space UNets - the diffusion runs directly on RGB pixels, no VAE, no latent space. That's a big part of why they were slow and why they look the way they do.

How it works

On load, the node picks the right architecture config for whatever model you chose (image size, channel counts, attention layout - all different per model family), then checks ComfyUI/models/Disco-Diffusion for the file. Missing it, it downloads - with SHA256 verification, from a primary Hugging Face mirror and a fallback if the first fails. It also always pulls the "secondary" helper model, a distilled network that cheaply predicts the denoised image during early steps so the CLIP guidance has something to grade. Everything loads with inference mode off, same as the rest of this pack, because the sampler needs autograd.

The inputs

Only two, and both are on the surface:

• model_name - the 27-model dropdown. For the classic Disco look, leave the 512x512 default. For the pixel-art and watercolor line, the author's own source notes the best pick is pixelartdiffusion_expanded.
• use_checkpoint - gradient checkpointing. It's a genuine trade: a little extra compute to keep memory use down during the backprop-heavy sampling. On a tight card, turn it on; it's the cheapest VRAM win this pack offers.

The output

GUIDED_DIFFUSION_MODEL - one wire, into the Disco Diffusion Sampler's guided_diffusion input. That's the whole job. There's no VAE to attach, no conditioning to encode; the model walks itself into the sampler as a bundle of config plus weights.

Installing it

Same pack, same story as the sampler: ComfyUI Manager (search "Disco Diffusion"), or:

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/space-nuko/ComfyUI-Disco-Diffusion
cd ComfyUI-Disco-Diffusion
pip install -r requirements.txt   # into the venv your ComfyUI uses

The --recursive matters - the pack vendors several submodules, and a plain clone leaves it half-dead.

The honest caveats

First-run downloads are big. Each model is a gigabyte-plus, and the SHA check means a corrupt partial download gets thrown away and retried rather than silently loading. If you're on an ancient link or a flaky connection, expect a wait. And the pack is legacy - last touched in 2023. On a current ComfyUI it may fail to import entirely; if your ComfyUI stops booting after you install this, it's the first thing to disable. When it does run, though, this loader is the easy part. The magic and the pain are both in the sampler.