MiniT2I Text Encoder Loader

The text encoder has quietly become a model you load on its own, and this node is that idea applied to MiniT2I. Where SD 1.5 hid its CLIP inside the checkpoint, modern diffusion models treat the encoder as a separate component - Flux and SD3 bolted T5-XXL onto their backbones, and MiniT2I follows the same playbook with a smaller flan-t5-large. This node is the "load the encoder" half of that split: no inputs, no settings, just a MODEL output to hand to the sampler.

Why it exists

MiniT2I's sampler could, in theory, own everything. The reason it doesn't anymore is the same v1.0.2 change that created the model loader: splitting the pack into loader and inference nodes lets ComfyUI's VRAM manager hold each piece and offload it between runs. The text encoder is a real chunk of memory - flan-t5-large is ~770M parameters, loaded at fp32 here - so letting ComfyUI treat it as a managed component beats re-creating it inside the sampler on every queue. That's the entire design philosophy of this pack in one node: small model, and let ComfyUI do the housekeeping.

How it works

Under the hood it's a single call: T5EncoderModel.from_pretrained("google/flan-t5-large"), wrapped as a ComfyUI MODEL. The T5 encoder turns your prompt into the token vectors the diffusion transformer cross-attends to - it's the network that actually understands your words, which is why the sampler has a plain text prompt input and no separate conditioning wiring. One practical note: the node has no inputs at all, so there's nothing to misconfigure. You literally drag it in, wire its MODEL output to the text_encoder input on the MiniT2I sampler (next to the MiniT2I Model Loader feeding the sampler's model input), and it does its one job.

First use triggers a download of google/flan-t5-large into your Hugging Face cache (~/.cache/huggingface), not into ComfyUI/models. It's the same encoder the sampler would pull anyway, so you're not adding a new model to your life - you're just giving ComfyUI a handle on it. The download is a bit chunky, so that first run feeling like it's frozen is normal; the console will show progress. The encoder is T5-family, under the permissive Apache-2.0 license, which is why this whole pack is comfortable to run without licensing anxiety.

Installing

Same pack, same steps - ComfyUI Manager, search "MiniT2I", install, restart, or:

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

Then restart. Like its sibling nodes it ships with no declared pip dependencies (it leans on transformers and huggingface_hub already in your environment) and it's built on ComfyUI's newer comfy_api extension framework, so keep ComfyUI up to date or these nodes won't appear at all.

There's not much more to say, because there's not much more to the node - and that's fine. It's a loader. You'll spend ten seconds wiring it and never think about it again, which is exactly what a text encoder loader should be.