Diffusion Model Selector

Diffusion Model Selector is a dropdown for diffusion models that records a choice instead of loading it. You pick a UNet (and a weight dtype) from a combo box, and the node hands back an IPT-Model value describing that choice - no weights touched, no VRAM spent. The actual loading happens later, when that value reaches a loader node. It's the same "separate selection from loading" idea the pack applies to checkpoints, LoRAs, CLIPs, and VAEs, and it exists to kill a specific ComfyUI habit: re-typing model names into a fresh loader at the top of every new workflow.

How it works

The combo lists the diffusion models ComfyUI can see in its diffusion_models directory. Alongside the model you pick weight_dtype - the dtype that will be passed through when the model is actually loaded by Load Diffusion Model, so you can decide at selection time whether a heavy model gets fp16, fp8, or the default treatment. There's also an optional sha256 string: if you've cached the model's hash, it feeds the right-click View Model Info... action, which fetches metadata from Civitai and shows it in a model-info window. Without a cached hash it falls back gracefully.

The model output is the pack's IPT-Model type - a description of the selection (file, folder, dtype) rather than a loaded runtime. The second output, unet_name, is the raw model name for when a node just wants the string.

Where the selection goes

IPT-Model plugs into the pack's image-info machinery, and that's the real point. Common targets, straight from the README:

• Image Info Defaults / Image Info Context - so the model becomes part of the image_info bundle that gets saved to the image's metadata.
• Load New Model / Use Loaded Model - the pack's split of loading from reuse, which keeps a process-local cache so a model selected in five places is actually loaded once.
• XY Plot Modifier - so a model can be one axis of a comparison grid.

Because the selection rides inside image_info, it survives the save/load round trip: save an image with Image Saver, load it back with Image Reader, and the diffusion model choice comes back with it, ready to feed Load New Model again.

Why you'd bother

If you build one-off graphs, this adds a step for nothing. If you keep rebuilding the same workflow with a different base model, a selector means the model choice is one visible dropdown instead of a loader buried mid-graph - and it makes metadata that much more reusable. The tradeoff is that it only speaks the pack's own IPT-Model type, so it's a commitment to the pack's conventions, not a universal adapter.

Installing it

Part of the ComfyUI-Info-Prompt-Toolkit pack. ComfyUI Manager - search "Info Prompt Toolkit" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/kinorax/comfyui-info-prompt-toolkit.git
cd comfyui-info-prompt-toolkit
pip install -r requirements.txt

Restart ComfyUI. No model downloads - this node only lists models you already have. Small single-author pack (GPL-3.0) with a modest community footprint; the selector-plus-metadata pattern is its main draw.