Shared Checkpoint Loader (Inspire)

If you flip between workflows a lot - or run a batch of different graphs that all use the same base model - you've felt this pain: every switch reloads the checkpoint from disk into VRAM, and that's several seconds of nothing while a 6GB file gets re-read for no reason. Shared Checkpoint Loader kills that. It's a drop-in replacement for ComfyUI's normal Load Checkpoint that caches the loaded model in Inspire's backend cache, so the second time you ask for it, it comes back instantly from memory instead of the disk.

That's the entire pitch, and it's a good one. Same three outputs as the vanilla loader (MODEL, CLIP, VAE), same checkpoint dropdown, but with a memory that survives workflow switches. For anyone iterating fast across graphs, it's a small quality-of-life win that adds up over a session.

How it works

The backend cache is a key-value store that Inspire keeps alive on the server between runs. When you load a checkpoint through this node, it stashes the loaded model under a cache key. Next time - same workflow re-run, or a different workflow asking for the same checkpoint - it checks the cache first and reuses what's there rather than loading again. The cache lives in the ComfyUI process, so it persists as long as the server is up.

The inputs and outputs that matter

• ckpt_name - the checkpoint to load, same dropdown you already know.
• key_opt - an optional custom cache key. Leave it blank and the node uses ckpt_name as the key, which is what you want most of the time. Set it explicitly only if you want a stable handle to delete later.
• mode (optional: Auto / Override Cache / Read Only) - Auto is the normal caching behavior. Override Cache forces a fresh load and replaces what's cached (use it if you swapped the file on disk). Read Only retrieves from cache but won't create a new entry.

Outputs are model, clip, vae - wire those exactly like a normal checkpoint loader - plus a cache key string. That last one isn't decorative: feed it to Remove Backend Data (Inspire) when you want to evict this checkpoint and free the VRAM.

Installing it

It's in the Inspire Pack by Dr.Lt.Data, the author of ComfyUI Manager and the Impact Pack - so this is mainstream, actively-maintained tooling. Via ComfyUI Manager: search ComfyUI Inspire Pack, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

then restart. No extra dependencies or model downloads for this node - it's pure backend plumbing.

Common issues

VRAM keeps climbing. This is the honest gotcha of caching: a cached checkpoint stays resident until you evict it. If you load several different big checkpoints through shared loaders in one session, they all sit in memory. Use the cache key output plus Remove Backend Data (Inspire) to clear ones you're done with, or restart the server to flush everything. The convenience isn't free - you're trading VRAM for speed on purpose.

I updated the checkpoint file but it's serving the old one. That's the cache doing its job too well. Set mode to Override Cache for one run to force a reload, then switch back to Auto.

Nodes won't load after installing. Check the terminal log - the usual Inspire Pack cause is a version mismatch. Update Inspire Pack and Impact Pack together and restart.