Hunyuan DiT Checkpoint Loader

Tencent's Hunyuan DiT landed in mid-2024 with real momentum - it was the SD3-alternative people were actively looking for at a moment when SD3's own release had disappointed, and the launch thread pulled hundreds of comments. It got Diffusers support, Kohya and webui integration work started, and comparison threads pitted it directly against PixArt Sigma and SD3 Medium. Then an EU data-compliance controversy surfaced around it, and - combined with Flux arriving a few weeks later and consuming essentially all community attention - the image model faded fast. (Tencent's next model, HunyuanVideo, had a completely different trajectory and became one of the most-discussed open video models of its era - worth knowing if you're wondering whether "Hunyuan" as a name is obscure. It isn't; the image model specifically just didn't stick.) This node is the pack's loader for that faded-but-real model.

The README flags it plainly as still a work in progress - initial groundwork came from a community PR before the maintainer built it out further - so treat this corner of the pack as less polished than the PixArt side.

Inputs. Two, both required. ckpt_name is a dropdown from your checkpoints folder. model selects the architecture size, defaulting to G/2, with G/2-1.2, XL/2, L/2, and B/2 also available - G/2 is the flagship variant Tencent actually shipped weights for.

Output. MODEL, into your KSampler.

Getting the checkpoint. Per the README: download the model file from HuggingFace, rename it to HunYuanDiT.pt, and place it in your checkpoints folder. That renaming step isn't cosmetic - the loader expects that filename convention.

The rest of the pieces you'll need. Hunyuan DiT runs on two text encoders simultaneously, loaded via HYDiTTextEncoderLoader: a Chinese RoBERTa model (download, rename to chinese-roberta-wwm-ext-large.bin, place in ComfyUI/models/clip) and an mT5 model (rename to mT5-xl.bin, into ComfyUI/models/t5 - or grab the smaller fp16-converted version for faster loading). VAE-wise, use any standard SDXL VAE - the README suggests madebyollin's fp16-fix version as a solid default.

Installing the pack. ComfyUI Manager: search "Extra Models for ComfyUI." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI_ExtraModels

then pip install -r requirements.txt in your venv and restart.

Worth knowing before you invest time here. This corner of the pack is genuinely less battle-tested than the PixArt side - smaller user base, WIP status acknowledged by the author, and the model itself never built the kind of ecosystem momentum PixArt did before Flux arrived. If you hit rough edges, that tracks with its history rather than being a sign you've done something wrong. If your goal is just "run a capable pre-Flux DiT model that isn't PixArt," it's worth trying; if you're troubleshooting an obscure error and short on patience, PixArt is the more thoroughly worn path in this same pack.