MiniT2I Model Loader

This is the quieter half of niknah's MiniT2I pack, and it exists because of one changelog line: v1.0.2 "split up into loader & inference nodes so ComfyUI handles VRAM, quicker." The MiniT2I Sampler node can run on its own, but when you feed it a model from this loader, ComfyUI's own VRAM management gets to hold the weights and offload them to system RAM between runs. On a model that's already small enough to fit on modest cards, that's the difference between "it runs" and "it runs without clogging everything."

What it loads

One thing, and only one thing: the MiniT2I diffusion transformer, exposed as a standard ComfyUI MODEL. The node has a single input, a combo box:

• model_type - b16 or l16. The tooltip from the author is the whole manual: "l16=large, b16=normal."

That's it. Everything else about the model - the fixed 512×512 output, the flow-matching sampler, the flan-t5-large text encoder - lives elsewhere. This node is just "go get the transformer weights."

Which one should you pick?

The README's VRAM table is your real guide: b16 is about 2.6GB, l16 about 4.3GB. The 16 is the patch size; both are the same architecture at different scale.

Honest advice: if you're here, you're probably on the low-VRAM end of the spectrum, so start with b16. It's the default in the author's example workflow and it's the one that fits comfortably alongside everything else ComfyUI is holding. l16 trades quality headroom for another ~1.7GB of VRAM, and on a cramped card that doesn't buy you much - ComfyUI will just shove the excess into system RAM and run slower instead of better. If you're unsure, b16. It's not a downgrade you'll be embarrassed by.

How to wire it

Feed its MODEL output into the model input of the MiniT2I sampler node, then add the MiniT2I Text Encoder Loader's output to the sampler's text_encoder input - that's the exact graph from the author's bundled example workflow. There's no VAE anywhere in the picture because there's no latent space; the sampler turns out a ready IMAGE.

Installing

Same as the rest of the pack - ComfyUI Manager, search "MiniT2I", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/niknah/MiniT2I-ComfyUI

One heads-up: the pack ships with zero declared pip dependencies and uses ComfyUI's newer comfy_api extension framework, so make sure your ComfyUI is current. First time this loader runs, it downloads the transformer into the Hugging Face cache (~/.cache/huggingface), not into ComfyUI/models - so a long pause on the first run is normal, not a hang. Subsequent runs are faster because the weights stay cached on disk and ComfyUI handles the memory juggling.

That's the whole job. If you want the sampler to be easy on your VRAM, this is the node that makes it so.