MinusZone - HYDiTInitWorkspace

MZ_HYDiTInitWorkspace is the entry point to the pack's native Hunyuan DiT path - "native" because it trains with Tencent's own HunyuanDiT repo rather than the kohya fork. It's the same bootstrap pattern as the other InitWorkspace nodes, but its workspace config type is different (MZ_TT_HYDiT_WorkspaceConfig), which matters more than you'd think: it's what keeps the native family separate from the kohya family in your graph. Wire the wrong init into the wrong dataset node and ComfyUI will tell you with a type error.

Hunyuan DiT, in case you found this node without the history: Tencent's open DiT image model, mid-2024, the SD3-adjacent architecture that the community briefly loved and then forgot when Flux showed up weeks later. The KB's forgotten-models doc has the arc, including the EU compliance scare that spooked adoption. It's a niche model today, and this node is for people who've deliberately chosen that niche.

How it works

On first run it clones Tencent/HunyuanDiT at a pinned commit into ComfyUI/models/minus_zone_models/train_tools/HunyuanDiT/, creates the workspace under ComfyUI/output/mz_train_workspaces/<train_name>/, and returns the workspace_config. Note the input is called train_name here, not lora_name - the native path names its workspaces after the training run.

The inputs:

• train_name - workspace name; required, and it raises if blank.
• branch - pinned commit of the HunyuanDiT repo. Default is what the author tested; changing it is an advanced move.
• source - github or the githubfast / 521github / kkgithub mirrors. The native path pulls the training repo from GitHub and the model weights from ModelScope, so if you're in an environment where one of those is slow, you'll feel it.
• seed - carried into the config for sample generation later.

Install & troubleshooting

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

Or Manager → search ComfyUI-TrainTools-MZ.

Two things to know before you build the rest of the graph. First, the big one: this path's trainer (MZ_HYDiTTrain) requires deepspeed installed manually - the code checks and raises a "please install a suitable deepspeed version" error if it's missing, which is the single most common first-run failure for this family. Second, the workspace config type mismatch trap: this node feeds MZ_HYDiTDatasetConfig and MZ_HYDiTTrain, and only those. The kohya-family nodes won't take it. The first clone takes a minute; after that, the heavy downloads happen on the train/generate nodes when their paths are set to auto.