Lumina2模型配置器

Lumina Image 2.0 is one of those models that's good enough to have a real following but not famous enough to have beginner-friendly tooling everywhere. Lumina2ModelNode is the pack's bridge for it: point the training run at the three files that make up the model and you're set. It's one of the rarer model pickers here, in a good way - most training stacks still don't cover Lumina 2 at all.

The support table lists Lumina Image 2.0 with LoRA and full fine-tune support, so this node gets you more headroom than the LoRA-only model nodes in the pack.

How it works

Standard model-picker behavior - it assembles a {"type": "lumina_2", ...} config with the three required paths, then feeds ModelConfig for the dtype/timestep merge. Where other nodes need one folder, Lumina 2 is a three-part model and the node is blunt about it:

• transformer_path - the transformer weights, typically a single-file lumina_2_model_bf16.safetensors.
• llm_path - the language-model text encoder. Lumina 2 uses an LLM for text conditioning rather than CLIP, which is exactly the architectural shift the modern training docs keep hammering on: you don't train these text encoders, you just feed them through.
• vae_path - the VAE.

All three are required; the node returns an error if any is empty. On top of them sits lumina_shift, which defaults to on - it applies the resolution-dependent timestep shift that Lumina uses. The interesting contrast with the SD3/OmniGen2 nodes is the default direction: those ship flux_shift off, this one ships its shift on, because Lumina 2 expects it.

The inputs that matter

• transformer_path / llm_path / vae_path - all required.
• lumina_shift - default on; leave it on unless you have a specific reason to disable it.

Output: model_path → ModelConfig.

Install

Pack-wide - 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, as always with this pack.

Where people get burned

Mostly by the three-required-paths structure - leave one blank and the node errors out with exactly which path is missing, so the failure mode is at least honest. The subtler issue is format: the paths in the tooltip point at single-file bf16 weights (lumina_2_model_bf16.safetensors), and if you've got a differently-quantized or folder-based download, the training load can fail in ways that don't point back at the node. Match the format the tooltips show. And since the text encoder is an LLM here, don't expect the SDXL-style encoder-LR knobs - that whole family of settings is gone for this model, and that's correct, not a bug.