Simple Checkpoint

Simple Checkpoint (zsqcheckpoint) from the zsq_prompt pack is a no-frills checkpoint loader: pick a model from a dropdown, get model, clip, and vae out the other side. It's ComfyUI's core Load Checkpoint with the fat trimmed - same outputs, fewer widgets, and one behavior difference that matters: it caches the loaded model by name and only reloads when you switch checkpoints.

How it works

The only required input is ckpt_name, a dropdown populated from your ComfyUI/models/checkpoints/ folder. It loads through ComfyUI's load_checkpoint_guess_config, which auto-detects the architecture (SD 1.5, SDXL, Flux-era, whatever the file is) and hands you the three standard outputs.

The caching detail is the real difference from core. The node keeps its loaded model in memory; as long as you don't change the checkpoint name, running the graph again reuses the cached model instead of reloading from disk. That's a genuinely nice speedup for batch runs and iterative tweaks - you pay the load cost once per checkpoint, not once per run. If you're on a modest machine, that can be the difference between "tolerable" and "waiting on disk I/O every queue."

Inputs and outputs that matter

• ckpt_name - the checkpoint dropdown. The one input.
• Outputs: model (MODEL), clip (CLIP), vae (VAE) - standard, wire them to a sampler, an encoder, and a VAE decode respectively.

Where it fits

It's a convenience loader inside a grab-bag pack. If you're already using zsq_prompt's other nodes, it makes a tidy three-node start to a workflow (loader → loraStack_2 for LoRAs → sampler). If you're not, core Load Checkpoint does the same job - the caching is the only real edge this one has, and honestly it's a small one.

The VAE output deserves the usual reminder: "baked in" vs "separate" VAEs vary by checkpoint. load_checkpoint_guess_config will often extract the VAE from the checkpoint file, but if your images come out grey or washed out, the fix is a proper VAE, not this loader - that's a model-file problem, and no loader can paper over it.

Installing zsq_prompt

Shared pack install. ComfyUI Manager → search zsq_prompt → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/windfancy/zsq_prompt

Restart ComfyUI. Heavy shared deps (transformers, opencv-python, opencv-contrib-python, scikit-image, timm, onnx, ultralytics) come with the pack; if Manager flags "conflicting with zsq_prompt" on a workflow, that's dependency-overlap noise, only meaningful once the pack is installed. No model files are downloaded - the dropdown reads your existing checkpoints folder.

Gotchas

• The caching means switching checkpoints mid-session triggers a reload, which can take a while on big models - expected, not a bug.
• It's guess_config, so very new or exotic architectures might not be recognized until ComfyUI itself knows them. When in doubt, use core Load Checkpoint which carries the same machinery more visibly.
• Fewer widgets than core loader - no separate VAE override here. If you need a specific VAE file, load it separately with a VAE loader and bypass this node's vae output.