MinusZone - HYDiTSimpleT2I

MZ_HYDiTSimpleT2I is the native Hunyuan DiT sampler: text-to-image, one node, IMAGE out. It's the generation half of the pack's native HYDiT path, and unlike the kohya-fork sampler (which limits you to two schedulers and has a version selector), this one exposes the fuller sampling surface of Tencent's own script - five schedulers, default CFG of 8, and keep_device control. Same auto model-path behavior, same ModelScope downloads.

If you trained a Hunyuan DiT LoRA with MZ_HYDiTTrain, this is the natural check-your-work node. If you haven't trained anything, it's still a way to run Hunyuan DiT in ComfyUI at all - the model never got the first-class ComfyUI support Flux did, so a self-contained sampler like this is genuinely useful for poking at a forgotten model.

How it works

Defaults matter here: unet_path, vae_ema_path, text_encoder_path, and tokenizer_path are all auto, meaning the node downloads the Hunyuan DiT components from ModelScope into ComfyUI/models/hunyuan/ on first run and reuses them after. t5_encoder_path is also auto by default on this node (unlike the kohya-fork version, where it's none) - so the native sampler pulls the mT5 encoder too, another chunk of download. lora_path defaults to none and lists your ComfyUI models/loras/, so you can load a trained LoRA straight in.

The sampling knobs:

• scheduler - ddpm, ddim, dpmms, uni_pc, or dpmpp_2m_karras. Default ddpm is the conservative choice; dpmpp_2m_karras is usually faster to converge.
• cfg (8), steps (20), width/height (512, step 16, max 8192), seed - standard. The 512 default is small; Hunyuan DiT is happier around 1024.
• prompt / negative_prompt - multiline. This is the model known for strong prompt adherence and composition, so write real prompts, not tag soup.
• keep_device (enable) - keep the model resident on GPU between samples. Disable to free VRAM for other graph stages; you'll pay a reload tax on the next queue.

Output is images (IMAGE) - preview or save as usual.

Install & troubleshooting

cd ComfyUI/custom_nodes
git clone https://github.com/MinusZoneAI/ComfyUI-TrainTools-MZ
# restart ComfyUI

First run is the download session - several GB from ModelScope, so budget the time and a stable connection. If sampling dies with an OOM, drop resolution or disable keep_device. If it dies at import, the dependency auto-install for the native path did its thing but a mismatch (usually transformers/diffusers versions against the pinned HunyuanDiT repo) is the culprit - the pinned commit is from the author's debug era, and newer libraries can drift. That's the eternal tradeoff of this pack: the pinned toolchains are reproducible but frozen, and you're responsible for making today's Python play nice with them. Also worth remembering, the sampler is Hunyuan DiT's image model - if you came here expecting the video model, that's the other Tencent repo entirely.