OmniGen2模型配置器

OmniGen2 is the unified image model - text-to-image, image editing, and a few other tasks in one architecture - which makes it an interesting but unusual LoRA target. OmniGen2ModelNode is the pack's model picker for it: one required path to a complete checkpoint folder, one optional toggle, and you're wired into the training pipeline.

The support table lists OmniGen2 as LoRA yes, full fine-tune no, so this node's job is narrower than some - you're here to teach the model a style, a character, or a specific task behavior, not to re-train it. Given how much weight an OmniGen2 checkpoint carries, LoRA-only is honestly the sensible posture.

How it works

It builds a {"type": "omnigen2", "diffusers_path": ...} config that flows to ModelConfig, gets the dtype/timestep stamp, and eventually lands in the TOML that the Train node runs through deepspeed. As with every model node here, no weights are loaded at this stage - the node is pure config assembly.

The one required input is diffusers_path, and the tooltip is emphatic about what that means: it needs a complete official checkpoint directory, with the example being /data/models/OmniGen-v1. Not a single safetensors, not a partial folder - the full model structure. The optional flux_shift toggle, default off, applies the resolution-dependent timestep shift in the style of Flux. It's the same knob you'll find on the SD3 node, and it's a "try it and compare" option rather than a required step.

The inputs that matter

• diffusers_path - required, the complete official OmniGen2 checkpoint directory.
• flux_shift - optional, off by default; the Flux-style timestep shift.

Output: model_path → ModelConfig.

Install

The standard pack install - ComfyUI Manager (search "Diffusion_pipe_in_ComfyUI") or:

cd ComfyUI/custom_nodes
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
cd Diffusion_pipe_in_ComfyUI
git submodule init && git submodule update
pip install -r requirements.txt

Linux/WSL2 only - on plain Windows the pack won't train at all. The deepspeed/diffusers stack comes with the requirements.

Where people get burned

The complete-folder requirement is where almost everyone trips. OmniGen2's official repos ship as structured diffusers checkpoints, and pointing this at a single-file conversion or a mid-download directory produces a load failure deep inside the training subprocess - the error won't say "your path is wrong," it'll say something cryptic about missing keys. Also remember this is a LoRA-only model in this pack; if you wire it up expecting full fine-tune options, they won't be there, and that's per the support table, not a bug. On flux_shift: default off is the author's default for a reason. Only flip it if you have a concrete resolution problem you're trying to fix.