Nodes/Comfy UI Online Loaders/Load Diffusers (Online)
ComfyUI Node

Load Diffusers (Online)

Load Diffusers (Online)? The name is a lie — this one's local

By yolanother·Created 3 years ago·Updated 2 years ago· 1
Load Diffusers (Online)
    • MODEL
    • CLIP
    • VAE
    model_path

    Let's start with the part the title gets wrong: this node is not online, and doesn't download anything. Load Diffusers (Online) is the pack's one genuinely local node. Where every other DT loader pulls its list from DoubTech's server, this one scans your own models/diffusers folder and lists directories that contain a model_index.json - the signature of a HuggingFace-style diffusers model, where the weights are spread across a folder tree (text encoder, UNet, VAE as separate subfolders) instead of bundled into one file.

    The "(Online)" is a copy-paste from the pack's other loaders, or a dream. Either way, expecting a dropdown full of hosted models will leave you confused. What you actually get is a listing of what you put there.

    What it's for

    Diffusers-format models are the "unzipped" way HuggingFace ships models, and ComfyUI can load them - the core DiffusersLoader node does exactly this. This node is a thin re-implementation of that, with one nicety: it picks fp16 automatically based on your VRAM situation (should_use_fp16), rather than making you decide. The real reason you'd want it: you downloaded a model from HuggingFace as a diffusers folder and don't want to convert it to a single-file checkpoint, or you're loading a model that's only published in diffusers form.

    Inputs and outputs

    • model_path - the dropdown of local diffusers folders found under models/diffusers. Empty dropdown means you have no diffusers folders there yet.
    • Outputs: MODEL, CLIP, VAE - the standard trio, wired like any checkpoint loader (MODEL to the sampler, CLIP to the text encoder, VAE to the decoder).

    It's filed under DoubTech/Loaders/Advanced, which is fair - it's the niche one.

    How it works

    On execution it walks models/diffusers (following symlinks), finds every directory with a model_index.json, and loads the chosen one through ComfyUI's comfy.diffusers_load.load_diffusers, with fp16 decided from your hardware. It's the pack's only loader that doesn't touch the network at all.

    Install

    cd ComfyUI/custom_nodes
    git clone https://github.com/yolanother/DTAIComfyLoaders
    

    or search "DTAIComfyLoaders" ("Comfy UI Online Loaders") in ComfyUI Manager, then restart. Only dependency is requests.

    Where people get burned

    • Expecting downloads that never come. If the dropdown is empty, that's correct behavior - there's no online catalog. You must download the diffusers folder yourself and place it under models/diffusers, then restart (the scan happens when the node's inputs are built).
    • Confusing it with the checkpoint catalog. The pack's other loaders list DoubTech-hosted single-file models; this one lists only local folders. Different thing entirely.
    • The fp16 auto-pick. On most setups it's the right call, but if you're chasing precision-critical work and want fp32, this node won't let you choose - it decides for you.

    Honest take: this node exists because the pack author wanted parity with core ComfyUI's loaders, and it's the least interesting of the lot. The core DiffusersLoader does the same job without the misleading title. Still, if you're already running this pack and work with diffusers-format models, it costs you nothing to use.

    CategoryDoubTech/Loaders/Advanced

    Inputs (1)

    NameTypeDefaultDescription
    model_pathCOMBO0 options:

    Outputs (3)

    NameTypeDescription
    MODELMODEL
    CLIPCLIP
    VAEVAE