Checkpoint Loader

The first honest thing to know about TMMCheckpointLoader is that it isn't really "the" checkpoint loader. It's the node the Tiny Model Manager dashboard drops onto your canvas when you click "+" on a checkpoint card - and behind the scenes it calls the exact same ComfyUI machinery as the core CheckpointLoaderSimple you already have. So if you've never touched this pack before, you've still used this node. It just had a different name on it.

Still, there's a reason the dashboard needs its own loader, and it's worth understanding what you're getting when it appears.

What it does

A checkpoint is the whole model - the diffusion UNet, the CLIP text encoder, and usually a VAE, bundled into one multi-gigabyte file. This node loads one from a dropdown and hands you three sockets:

• model (MODEL) → wire into your sampler. This is the thing that actually denoises.
• clip (CLIP) → wire into CLIP Text Encode (both the positive and negative prompt). This is the text encoder that turns your words into conditioning.
• vae (VAE) → wire into VAE Decode at the end. This is the pixel↔latent bridge that turns the sampler's latent back into an image.

One input, ckpt_name, which is a dropdown of every .safetensors / .ckpt in ComfyUI/models/checkpoints. That's the whole interface.

How it works

Under the hood the node calls comfy.sd.load_checkpoint_guess_config - the same function ComfyUI's own loader uses. The "guess" part matters: the function sniffs the file's architecture (SD 1.5, SDXL, Flux, Pony, whatever) and sets up the right config so the model, CLIP, and VAE all come out matching. It also points ComfyUI's embedding directory at your embeddings folder, so embedding: tokens in your prompt resolve while the checkpoint is loaded. You get Flux-scale complexity with a one-dropdown UI, because you're standing on the core loader's shoulders.

Why use this one instead of the core loader?

You don't have to - and if you build a workflow by hand, you probably shouldn't, because the core node is identical and one less thing to install. Reach for TMMCheckpointLoader when the Tiny Model Manager dashboard inserted it for you, which is the entire reason it exists: the dashboard's one-click "add to workflow" needs a matching node class to place. A nice detail: when you download a checkpoint through the dashboard, it refreshes ComfyUI's model list first, so the new file is already in this node's dropdown the moment it lands - no reloading the tab.

Install

The pack is one small repo - Zellione/comfyui-tiny-model-manager - that bundles a web dashboard plus these loader nodes. Install via ComfyUI Manager by searching Tiny Model Manager, or from a terminal:

cd ComfyUI/custom_nodes
git clone https://github.com/Zellione/comfyui-tiny-model-manager
cd comfyui-tiny-model-manager
pip install -r requirements.txt

Then restart ComfyUI. Python deps are light (aiosqlite, httpx, aiofiles, mistune) and the dashboard ships prebuilt, so no Node.js toolchain is required. The dashboard lives at http://localhost:8188/tiny-model-manager.

Where people get caught

• Expecting the loader to do more than it does. It's a loader, not a downloader. If the file isn't already in ComfyUI/models/checkpoints, it won't be in the dropdown - the downloading is the dashboard's job.
• Wiring vae into the sampler's VAE slot by reflex. Most current checkpoints bake their VAE, so a standalone VAE is often unnecessary; when it is needed, that output goes to VAE Decode, not into the KSampler.
• Heavy deps that aren't there. There are none - no extra torch packages, no frontend build step. If the dashboard at /tiny-model-manager complains it wasn't built, the loader nodes still work fine; only the dashboard needs the npm install && npx ng build rebuild.

One caveat, the honest kind: this pack is young, with basically no community footprint yet (you won't find a Reddit thread about it). Everything above is straight from the README and source, so it should age well - but treat it as a promising new tool, not an established one.