Checkpoint Save (No Workflow) (DEPRECATED)

DonutCheckpointSave is a save node with a very specific job: take a model, a clip, and a vae and write them to disk as a complete checkpoint file - no workflow embedded, no extra metadata ceremony. The "No Workflow" in its display name is the whole point. ComfyUI's core save node embeds the generating workflow in the PNG metadata; this node's lineage exists for people saving models, not images, who want a clean checkpoint they can load back through a normal Load Checkpoint.

The inputs are what you'd expect: model, clip, vae, a filename_prefix (default checkpoints/ComfyUI - the file lands in your models/checkpoints folder), and a dtype dropdown (default original, with options for casting the save). No outputs - it's an output node; when it runs, it writes the file. Set the prefix, wire in your merged or patched model, and run.

How it fits the pack

This is the natural end of the pack's model-merging story. You merge a couple of checkpoints (with the pack's merge nodes), calibrate them (Donut Block Calibration), tune them (Donut Detailer family), and when the result is worth keeping, this is the node that bakes model + text encoder + VAE into one loadable file. The dtype control matters here: original keeps whatever precision your tensors are in, which for a freshly merged model is usually fp32 - fine for a working checkpoint, bigger than you'd want to distribute. Saving at fp16 (or the pack's quantized options, if present) shrinks the file considerably at negligible quality cost for most SDXL use.

The honest verdict

The node is DEPRECATED - the pack folded the family into a single DonutSave node that handles both model-only and full-checkpoint saves, and DonutCheckpointSave is now just an alias kept so old workflows still load. If you're building new graphs, use DonutSave (from the same pack) and pick the full checkpoint inputs; the behavior is identical, this node is just the legacy name.

One caution that applies to any checkpoint save: you're writing a file that can be 2–7 GB depending on dtype. Save with original while you're still iterating, switch to fp16 for anything you intend to keep or share, and confirm the file actually lands in models/checkpoints - the default prefix points there, but if you changed ComfyUI's output layout, a save "succeeds" into a folder you never look at.

Installing

Ships with the pack: ComfyUI Manager → search DonutNodes, or git clone the repo into custom_nodes/ and pip install -r requirements.txt with the same Python that runs ComfyUI. No model downloads - it writes, it doesn't fetch. And as a save node it only fires when the graph runs; if you wired it but saw no file, check that it's actually connected to the model you merged, not sitting bypassed.