MinusZone - HYDiTTrain

MZ_HYDiTTrain is the flagship of the pack's native Hunyuan DiT path - the node that runs LoRA training against Tencent's own HunyuanDiT repo, with the whole pipeline (auto-downloaded model components, CSV building, deepspeed training, per-save sample generation) handled for you. It's the most "train a real model in ComfyUI" node the pack offers, and it carries the most prerequisites.

Hunyuan DiT being a niche model, this is also the node where the pack's rougher edges show up. It's doable - people do run it - but plan for a setup session, not a five-minute install.

How it works

The full chain: MZ_HYDiTInitWorkspace → MZ_HYDiTDatasetConfig → this node, optionally with MZ_HYDiTAdvConfig on advanced_config.

On start, the node does a few things in order. It auto-installs its dependency list (pandas, pyarrow, diffusers, transformers, timm, peft, accelerate, loguru, einops, sentencepiece, polygraphy, protobuf) via pip, then raises if deepspeed isn't already installed - that one it refuses to auto-install because the right version depends on your CUDA/torch. Then, if your model paths are auto, it downloads the Hunyuan DiT components from ModelScope into ComfyUI/models/hunyuan/ (several GB - the DiT module plus CLIP plus VAE; t5_encoder_path is none by default here, so no mT5 unless you ask for it). Then it walks the workspace images, builds the image_text.csv the native repo trains from (captions read from .txt/.caption files next to each image), and runs training.

Inputs that matter

• workspace_config (native MZ_TT_HYDiT_WorkspaceConfig) and the optional workspace_images_dir (from MZ_HYDiTDatasetConfig; falls back to train_images/).
• unet_path, vae_ema_path, text_encoder_path, tokenizer_path, t5_encoder_path - all auto by default → downloads. Set any to a real path if you already have weights.
• ema_to_module (enable) - use the EMA weights as the module. Keep on.
• resolution (1024), batch_size (1), epochs (50), ckpt_every (500) - the run shape. Note ckpt_every here counts steps; drop it if you want frequent saves.
• rank (8) - LoRA rank. Same logic as everywhere: 8 for likenesses, up for styles.
• base_lora - latest resumes from the newest checkpoint in the workspace output; empty starts fresh; or pick any scanned LoRA. The resume behavior is the pack's signature and it's worth it on long native runs.
• sample_generate (enable) + sample_prompt - renders a sample at each save and displays it. Keep it on; Hunyuan DiT runs are slow, and you want the early-warning.

Install & troubleshooting

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

Then the one manual step this path can't do for you:

pip install deepspeed   # pick the build matching your CUDA/torch

Expect the first run to take a long while - dependency installs plus the ModelScope downloads happen before a single training step. If the run dies mid-epoch, the workspace survives; fix the env and re-queue with base_lora = latest to resume from the last save. If training errors on data, check the CSV it built - captions and image counts are the usual culprits. And remember the license context: Hunyuan weights are Tencent community-licensed, with EU/UK/South Korea excluded from the territory; fine for personal training, read the license before commercializing.