Load Iris-3B Model

Every Iris-3B graph starts here. This is the node that pulls the 12 GB text-to-image checkpoint off your disk, tells ComfyUI what precision to run it in, and hands the whole thing downstream as an IRIS_MODEL. If you've only run SDXL or Flux before, the thing to unlearn is that you'll be hunting for a matching VAE loader and CLIP loader next. You won't. Iris is pixel-space: the DiT predicts RGB directly, there's no VAE to decode through, and the sampler's output is already an IMAGE. That's the same architectural bet HiDream-O1 and Chroma Radiance made in 2026, and it's why these workflows look half-empty when you open them.

What it actually loads

Iris-3B is a 3B-parameter DiT from Sperid Labs. This node loads the safetensors weights and pairs them with the matching config that ships inside the pack - you don't pick a config, it's inferred from the filename. The weights are recognised by name: the dropdown lists anything in models/diffusion_models whose filename contains "iris" and doesn't say upscaler or depth. So the upscaler and depth checkpoints will never appear in this list even after you've downloaded them, which is deliberate - they belong to their own loaders.

The precision switch is the whole personality of this node.

• bf16 - the default, and the one you want. Roughly 6 GB on the GPU.
• fp32 - the precision the model actually shipped in, and 12 GB. The author's own tooltip says it needs well over 16 GB VRAM. On a 16 GB card, don't.

Even at bf16, the model doesn't sit on the GPU. Between runs it lives in system RAM and only visits the card for its own forward passes, which is the trick that lets a 16 GB card run text-to-image, upscaling and depth in the same graph without any of them spilling into each other's VRAM. The cost is RAM: budget around 22 GB of free system memory if you plan to have all three Iris models resident.

The inputs that matter

model_name is an enum, and it starts empty on a fresh install. The pack handles the awkward part for you: if you pick the default (iris-3b.safetensors) and it isn't there, the loader downloads it from the Hub the first time you run the graph, 12 GB. ComfyUI will also list it under Missing Models with a Download button. Either path works, you don't need to fetch it by hand first.

weight_dtype is the only real decision, covered above. Leave it on bf16.

The single output, IRIS_MODEL, goes straight into Iris Sampler (Simple), Iris Sampler (Advanced), or through Load Iris LoRA / Load Iris Turbo LoRA first if you're adding adapters.

Install

Search Iris-3B in ComfyUI-Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bani4kaskashka/Iris-3B-Comfy-Nodes
pip install -r requirements.txt

That requirements file is one line: omegaconf>=2.3. Manager installs it for you. You also want transformers 4.57 or newer, which recent ComfyUI already has - that one matters for the text encoder, not this node. Restart ComfyUI afterwards, then check the Iris-3B category in the node search.

Where it goes wrong

Empty dropdown. The name filter is the usual culprit. If you downloaded model.safetensors yourself and left it named that, it won't show up - rename it to something containing "iris". A Hub-layout folder (model.safetensors next to its config.yaml) also works, but the flat renamed file is the path the author documents.

OOM on first run. Almost always fp32. The fp32 path is meant for people with 24 GB and up, and even they'll feel it.

Nothing happens on first use. It's a 12 GB download with huggingface_hub. If the progress looks frozen, watch your disk, not the console - and make sure the drive holding models/ has room for the other two checkpoints if you're going to use upscaling and depth too. The whole set is about 44 GB.