Load Checkpoint (Online)

This is the node the whole pack is built around: Load Checkpoint (Online) is a drop-in replacement for ComfyUI's stock CheckpointLoaderSimple, except the model list comes from DoubTech's server and the file downloads itself the first time you pick it. You choose a checkpoint from the dropdown, the node pulls it into your models/checkpoints folder, and out come the same three sockets every ComfyUI user knows: MODEL, CLIP, VAE.

The catalog is where this gets interesting, and it's worth knowing what's actually in it. There are 13 entries, and it's a time capsule of roughly 2023–24: A-Zovya RPG, Deliberate v2, DreamShaper v7, Photon, Reliberate, WoW, Newdawn (the SDXL one), SDXL Base 0.9 / 1.0, SDXL Refiner 0.9 / 1.0, The Talos Project [SDXL] - plus Absolute Reality and DreamShaper v7, which are the two entries that will bite you, more on that below. These are the checkpoints the KB's concepts doc would call fine-tunes and merges of the SD 1.5 and SDXL generations. There is no Flux here, no modern base - the catalog was frozen in the era when DoubTech last updated it (the pack's last commit is late 2024).

The one input, three outputs

ckpt_name is the only input - the dropdown. Outputs:

• MODEL → the KSampler (the model socket).
• CLIP → CLIP Text Encode (Prompt) for both positive and negative prompts.
• VAE → VAE Decode at the end (SDXL checkpoints carry a VAE; the pack's guess-config loader handles that).

If you can wire a stock checkpoint loader, you can wire this.

How it works

On startup the pack fetches the checkpoint list from api.aiart.doubtech.com/comfyui/checkpoints. When you run the node, it resolves the filename, downloads the model if it isn't already in models/checkpoints, then calls ComfyUI's own load_checkpoint_guess_config - the same guess-the-architecture path the core node uses. Once the file is on disk, this is just a checkpoint loader; no more network.

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

• Two catalog entries are broken links. Absolute Reality and DreamShaper v7 point at CivitAI model pages, not direct file URLs. The node will try to download an HTML page and fail to load it. Don't pick those two.
• First downloads are gigabytes. DreamShaper-style SD 1.5 models are ~2 GB, SDXL models ~6.5 GB, all over Dropbox links with zero progress UI. The node looks dead while it works.
• Dropbox HTML trap. Scripted downloads from Dropbox can come back as an HTML bot-check page. A "downloaded" checkpoint that fails to load is usually a ~200 KB HTML file - grab the real file from the link and drop it in models/checkpoints, and the node will happily use it.
• Empty dropdown on fresh start. The list loads in a background thread; re-add the node after a moment.

Honest verdict: if you already have a stable local workflow, the stock loader beats this - you control your files, your sources, and your disk. Where Load Checkpoint (Online) genuinely helps is a fresh install where you want to be generating in minutes without managing model downloads. Just skip the CivitAI-page entries and expect a long first run.