Qwen-Image模型配置器

Qwen-Image is the model family this pack has clearly been tracking hardest - the README's update log shows new Qwen variants landing almost monthly, and QwenImageModelNode is the node that grew to hold them all. One node, five versions in a dropdown: qwen_image, qwen_edit, qwen_edit_2509, qwen_edit_2511, and qwen2512. If you're training a Qwen LoRA, this is the picker you want, and it's the most versatile model node in the pack.

The support table is generous across the family: Qwen-Image, Qwen-Image-Edit-2509 and 2511 all support LoRA and full fine-tune, and the README notes both Diffusers-format and ComfyUI-format models are accepted.

How it works

Unlike most model nodes, the required input here is just the version dropdown - all four paths are optional. The node builds a {"type": ...} config and only includes the path fields you actually filled in. The model type maps from your version choice (the source shows qwen_edit_2511 routing to type qwen2511, the other edit variants to qwen_image), and whichever paths you supply get normalized and folded in.

The path flexibility is the point: you can point it at a full diffusers folder (diffusers_path), or mix ComfyUI-format pieces - transformer_path for the transformer, text_encoder_path for the Qwen2.5-VL text encoder, vae_path for the VAE. For the edit variants you'll also want the multi-image edit dataset node (MultiImageEditDatasetPathNode) feeding the dataset side, since edit training needs target + control images.

The inputs that matter

• model_version - the dropdown: qwen_image (default), qwen_edit, qwen_edit_2509, qwen_edit_2511, or qwen2512. This one choice selects the training recipe.
• diffusers_path - full diffusers folder, easiest path if you have one.
• transformer_path / text_encoder_path / vae_path - ComfyUI-format pieces; use these when you don't have a diffusers folder.

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. One extra note specific to this family: the README warns you need the latest diffusers development build for some of the newest Qwen support:

pip install git+https://github.com/huggingface/diffusers

Where people get burned

The version dropdown is where most problems start - qwen_edit_2511 and qwen_edit_2509 are not interchangeable recipes, so pick the one that matches the checkpoint you actually downloaded. The "both formats accepted" flexibility also cuts both ways: if you supply a diffusers_path and ComfyUI-format piece paths, the config can end up with conflicting sources; pick one style and stick to it. And for edit variants, remember the dataset node matters as much as the model node - edit training fed with a plain image dataset node won't teach the edit behavior.