Delete Model Light (Passthrough Any)

The pack's flagship Delete Model node has a dirty secret in its source: it calls ComfyUI's free_memory() with a huge byte count, which means it evicts everything the memory manager is tracking, not just the model you pointed at. If you're running a multi-model workflow, that's a sledgehammer where you wanted a scalpel. Delete Model Light (Passthrough Any) is the scalpel.

Same interface, same pass-through behavior: data in, model in, output out - your data rides through untouched while the model gets destroyed. The difference is in the cleanup. The Light version removes only the target from ComfyUI's loaded_models() registry, nulls out its heavy attributes and detaches its parameters, then runs gc.collect() and torch.cuda.empty_cache(). Crucially, it skips that blanket free_memory() call. The author even added a guard that warns you if more than the target model got unloaded - which, honestly, tells you a lot about how the non-light version behaves.

When you'd reach for this one

Stacked-model workflows. You've encoded with a big CLIP, you're done with it, and now the diffusion model needs every byte. Rather than let ComfyUI park the CLIP in system RAM - which is fine until your RAM is the bottleneck - this node actually frees it, while your encoded CONDITIONING rides through on the data wire into the sampler. The other model you're mid-use (the ControlNet, the VAE) stays loaded, because the Light node isn't nuking the registry.

The same trade-offs the README is upfront about apply: the model is fully gone, so if anything later needs it, it's a slow reload from disk, and ComfyUI loses its cache benefit. And if another node still holds a reference to the model, you'll see the partial-free warning in the console - delete the reference or the memory won't actually come back.

Install

No surprises, same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Isi-dev/ComfyUI_DeleteModelPassthrough
pip install -r requirements.txt

Or just search "DeleteModelPassthrough" in ComfyUI Manager. Restart, and you're done - no model files, and requirements.txt is only torch and psutil. A final honest note: this is a niche utility from a WIP pack with essentially no community track record, and most people with a reasonable amount of RAM are better off letting ComfyUI's own caching do its job. This node exists for the genuinely tight setups - low VRAM and low system RAM - where you need a model gone, not parked. For most of us, the fact that you reached for it at all means you should also be looking at GGUF quantization and fp8 loaders, which solve the problem at the source instead of cleaning up after it.