Checkpoint Clip Loader Resident

Checkpoint Clip Loader Resident is the "I only need the text encoder" node. Load a checkpoint file and get back a single CLIP output - the diffusion model and VAE never get built, and if an equivalent CLIP object is already alive in memory, it's reused instead of rebuilt. For graphs that only need to encode a prompt, that's the whole game.

Why would you want just the CLIP from a checkpoint? Two situations come up constantly. First, prompt conditioning nodes - you have a checkpoint whose text encoder you like (or whose wording/tokenizer quirks you've tuned around) and you don't need its UNet at all. Second, the "staged checkpoint" workflow this pack is built around: load the model with Checkpoint Model Loader Resident, the CLIP here, the VAE with its own component loader, and each piece stays resident independently. ComfyUI's stock loader forces you to take all three together; this family of nodes lets you pay for exactly what you use.

The inputs are refreshingly small:

• ckpt_name - the checkpoint file.
• Optional policy_override - a residency policy string; connect Set Global Residency Policy here.

That's it. One output, CLIP, which feeds your CLIP Text Encode node.

The "reuse live equivalent" behavior is the point. Because the pack tracks loads in its residency registry and matches by loader key, loading the same checkpoint's CLIP twice - or switching between two checkpoints that share a text encoder - can hand you back the live object instead of re-reading and re-encoding from disk. It's a quiet optimization, but on big text encoders (Flux's T5-class models come to mind) it's real time saved.

The usual pack caveats still apply. The selective fast path is a safetensors thing; a .ckpt or .pt checkpoint still loads through CPU-first torch.load(), it's just tracked and reusable afterward. And this node inherits the pack's startup monkey patches on ComfyUI's load and residency paths whether you use the residency controls or not.

Install:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-GPU-Resident-Loader
cd ComfyUI-GPU-Resident-Loader
pip install -r requirements.txt

Restart ComfyUI, or find the pack by searching "comfyui-gpu-resident-loader" in ComfyUI Manager.

When would this not be the right tool? If your graph needs the full checkpoint anyway, the full Checkpoint Loader Resident handles it in one node. And if you don't have any VRAM-pressure or reuse problem, stock "Load Checkpoint" plus ignoring the extra outputs is simpler. This node earns its keep when you're deliberately staging a workflow so that only the pieces you actually use ever get loaded and held resident.