Model Pruner (mtb)

Checkpoints are fat, and a lot of that fat is stuff you don't need to keep on disk. Model Pruner is MTB's node for slimming one down - convert the weights to a smaller precision, strip the EMA copy, drop junk keys, and save a lighter file. If you've got a 6GB fp32 checkpoint that could be a 2GB fp16, this is how you cut it without leaving ComfyUI.

Fair warning up front, straight from the node's own description: this is experimental. It works, people use it, but treat the output as something to verify rather than trust blindly. Prune a copy, load it back, make sure it still generates before you delete the original.

How it works

It's an output node - it doesn't hand you a MODEL to keep chaining, it writes a file to disk and stops. You wire in whichever components you want to keep (the UNet, the CLIP, the VAE - all optional), tell it what to do with each, and it saves a new checkpoint. The savings come from three places: dropping the EMA weights (a second copy of the UNet used during training that you don't need for inference), converting precision (fp16 is half the size of fp32 for a negligible quality difference in most cases), and removing junk metadata keys.

The inputs and outputs that matter

The three model inputs - unet (MODEL), clip (CLIP), vae (VAE) - are all optional; wire the ones you want in the output. For each, there's a matching pair of knobs:

• precision_unet / precision_clip / precision_vae - full, fp32, fp16, bf16, or fp8. fp16 is the standard "make it half the size" choice. fp8 is aggressive and can cost quality - test it.
• operation_unet / operation_clip / operation_vae - copy (keep as-is), convert (change precision), or delete (drop this component entirely).
• ema_mode - remove_ema (the usual choice, drops the training copy), disabled, or ema_only.
• fix_clip and remove_junk (both default true) - clean up a broken CLIP and strip junk keys.
• save_folder (default checkpoints/ComfyUI) - where it lands. A relative path is relative to ComfyUI's output directory. save_separately splits the components into separate files instead of one checkpoint.

There are no node outputs - the result is the saved file.

How to install it

Via ComfyUI Manager: search MTB Nodes, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/melMass/comfy_mtb

then restart. No extra models.

Common issues & troubleshooting

The pruned model looks worse. Precision is the usual culprit. fp16 is almost always safe; fp8 trades real quality for size and isn't a free lunch. If quality dropped, re-prune at fp16 (or bf16) before blaming anything else.

I can't find the output file. Check save_folder - a relative path saves under ComfyUI's output folder, not models/checkpoints. Move it into your checkpoints folder afterward, or point save_folder at an absolute path.

Nothing happens / it's an output node. Right - you can't feed its result into a sampler in the same graph. It writes a file; you load that file in a fresh run. Queue the prompt, then go find the checkpoint on disk.

It's experimental, so verify. Load the pruned checkpoint and generate a test image before you throw away the original. Dedicated prune scripts exist too; this node is the convenient in-Comfy option, not the only one.