Vetted Checkpoint Loader

PseudoVettedCheckpointLoader is the pack's answer to a real annoyance: which checkpoint should a workflow load, anyway? The pseudotools team curates a shortlist of models that work well in their architectural-rendering pipeline, and this loader gives you that list as a dropdown instead of the grab-bag of everything you've ever dragged into models/checkpoints.

"Vetted" here means Pseudotools' editorial pick - not an intrinsic quality badge. The current shortlist is six models across two generations:

• SDXL line: Juggernaut_X_RunDiffusion_Hyper.safetensors, albedobaseXL_v21.safetensors, realvisxlV50_v50LightningBakedvae.safetensors, sd_xl_base_1.0.safetensors, sd_xl_refiner_1.0.safetensors
• SD 1.5: aiAngelMix_v30.safetensors

Those are sensible picks, and they map to the README's own recommendations: aiAngelMix for SD 1.5 and albedobaseXL for SDXL. Juggernaut XL is the community's long-standing default realistic SDXL, and the two Lightning/RealVis entries are the fast-generation distilled variants - handy for iteration-heavy parametric work where you generate a lot and keep little. If you're following the pack's workflow templates, you'll see this loader at the front of the graph.

How it works

Under the hood it's a wrapper around ComfyUI's own checkpoint loader. The dropdown is the interesting part: at startup the node phones home to https://tools.pseudotools.com/api/models, pulls the current vetted list, and builds the dropdown from what the API returns. Pick a model, and it loads the file with the standard load_checkpoint_guess_config call, which auto-detects the architecture and hands back the three things you need.

Outputs:

• model (MODEL) → into the sampler (and into the LoRA loader if you stack one)
• clip (CLIP) → into the CLIP Text Encode nodes
• vae (VAE) → into the VAE Decode

There's also a hidden-ish record_id STRING input that tracks which pseudotools record a model belongs to - it's populated for you and you can safely ignore it unless you're scripting against their API.

The catch: it doesn't download anything

This is where people get burned. The dropdown lists filenames that are approved - it does not fetch the model for you. If Juggernaut_X_RunDiffusion_Hyper.safetensors isn't already sitting in ComfyUI/models/checkpoints, the node raises an error the moment you run. The vetted list is a filter, not a delivery service. Grab the models yourself from Civitai or HuggingFace (the README links the ones it recommends) and put them in the checkpoints folder with exactly the filenames the dropdown shows - the name has to match.

Second gotcha, straight from the source code: the dropdown only populates if the node can reach the Pseudotools API at startup. If you're offline, or the API is having a bad day, the dropdown collapses to (no vetted checkpoints available) and the node has nothing to load. A restart with a working connection fixes it. If your ComfyUI has no internet at all, this loader is dead weight - use the regular Checkpoint Loader instead.

Installation

Part of the Pseudocomfy pack:

# ComfyUI Manager: Custom Nodes Manager > search "Pseudocomfy" > Install

# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/Pseudotools/Pseudocomfy.git

Restart after installing. On Windows portable, run the README's first step to install diffusers into the embedded Python - the pack's pipeline wants it. Then make sure your chosen checkpoint is in models/checkpoints under the exact filename the dropdown expects.

Verdict

If you run the Pseudocomfy architectural workflow, this is a convenient front door that keeps the whole team on the same curated models. If you're a general ComfyUI user, the regular Checkpoint Loader lists every model you own, needs no network, and does the same job - the only reason to switch is if you want someone else's shortlist enforced on your graph.