MinusZone - KohyaSS_KohakuBlueleaf_HYHiDSimpleT2I

MZ_KohyaSS_KohakuBlueleaf_HYHiDSimpleT2I is the generation companion to the pack's KohakuBlueleaf Hunyuan DiT training path: a self-contained text-to-image node that runs Hunyuan DiT with your LoRA loaded, and hands you an IMAGE batch. If you just trained a Hunyuan DiT LoRA with MZ_KohyaSS_KohakuBlueleaf_HYHiDLoraTrain, this is the node that lets you see it working without leaving the graph or touching a terminal.

It's the kohya-fork flavor of generation, which means a couple of differences from the pack's native Hunyuan DiT sampler (MZ_HYDiTSimpleT2I): the scheduler list is short (euler_ancestral or dpmpp_2m_sde), the default CFG is 5, the default resolution is 1024×1024, and there's a version selector for the model weights (1.1 or 1.2, default 1.2).

How it works

Like the train node in this family, it defaults its model paths to auto, which downloads the Hunyuan DiT components into ComfyUI/models/hunyuan/ on first run and reuses them after. It also reuses the cloned KohakuBlueleaf fork for generation, so the toolchain is shared with training - no second install. lora_path defaults to none and lists your ComfyUI models/loras/ plus workspace LoRAs, so you pick the LoRA you just trained and sample it directly.

The inputs you'll actually touch:

• version - 1.1 or 1.2 of the Hunyuan DiT weights. Match what you trained on.
• prompt / negative_prompt - multiline. Hunyuan DiT responds well to detailed prompts and is known for strong composition; negative prompting matters less here than on SD but still helps.
• lora_path - your trained LoRA, or none to test the base model alone (do this once before blaming the LoRA).
• steps, cfg, scheduler, width, height, seed - standard sampling knobs. Defaults are sane; dpmpp_2m_sde is a good starting scheduler.
• keep_device (default enable) - keep the model resident on the GPU between samples. Disable if you're juggling VRAM with other graph stages.

Output is images (IMAGE) - wire it to a preview or save node.

Install & troubleshooting

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

The first generation triggers the model downloads and the fork's dependency install - the same one-time cost as training, so the honest tip is to do your first training run and your first sample in the same session and let the downloads happen once. If a prompt produces nothing, check version matches your weights. And the usual Hunyuan caveat applies: the weights are Tencent community-licensed with EU/UK/South Korea excluded from the licensed territory - fine for personal sampling, a thing to read before commercial work. This node is also a decent sanity check before you train: generate a few base-model images with lora_path = none to make sure your Hunyuan DiT install is healthy.