LoadLoraQwenImage (Selective Qwen Image layers from ai-toolkit)

LoadLoraQwenImage is a ComfyUI custom node for Qwen Image Edit pipelines. It selectively applies Qwen Image edit LoRAs to specific UNet transformer layers based on keyword filters, keeping the rest of the network untouched for precise control.

---

Feature Overview

• 🎯 Layer-level control – Match target transformer layers via the layer_filter keyword list (e.g. transformer_blocks.0). You can control the layers of Lora applications.
• 🔒 Exclude specific layers – Use exclude_filter to explicitly block layers from being patched. This is applied after layer_filter and takes precedence when both are provided.
• 🧩 Qwen Image Edit friendly – Optimized for ai-toolkit Qwen Image edit LoRAs and compatible with existing workflows.

---

Inputs & Outputs

| Name | Type | Description | |------|------|-------------| | model | MODEL (optional) | Diffusion model to be patched. Connect the output of a Qwen Image checkpoint loader or another upstream node. | | lora_name | Combo | LoRA file from ComfyUI/models/loras. | | strength_model | FLOAT | Scaling factor for the injected LoRA. Supports negative values; default is 1.0. | | layer_filter | STRING (multiline) | Optional comma/newline separated keywords to match target weight paths. Leave empty to apply to all available layers. | | exclude_filter | STRING (multiline) | Optional comma/newline separated keywords to exclude matched layers (e.g. attn, transformer_blocks.1). Applied after layer_filter; when both are set, exclusion wins. |

Output:

• model – The model patched with the selected LoRA. Attach it to samplers or downstream nodes.

---

Usage Scope

• Recommended – Qwen Image Edit / Qwen Image-VL LoRAs, especially ai-toolkit releases focused on instruction or style edits.
• Compatible – Any LoRA that follows Stable Diffusion / SDXL naming conventions, as long as the layer names match.

Qwen Image Model Notes

Qwen Image and Qwen Image Edit UNet backbones contain 60 transformer blocks, indexed from transformer_blocks.0 through transformer_blocks.59. When you populate layer_filter, you can point at any subset of these blocks—for example:

transformer_blocks.0,
transformer_blocks.5,
transformer_blocks.0, transformer_blocks.12

Leave the filter empty to apply the LoRA to all available transformer layers.

Include vs Exclude

• layer_filter selects a candidate set (leave empty to select all).
• exclude_filter removes any candidates whose names contain the given tokens.
• If both are provided, exclusion takes precedence.

---

How to Use

1. Copy Comfyui-QwenLoraLoaderSelective into ComfyUI/custom_nodes.

2. Launch ComfyUI. The node appears under loaders as LoadLoraQwenImage (Selective Qwen Image layers from ai-toolkit).

3. In your workflow: 1. Connect the upstream MODEL output (e.g. from Checkpoint Loader (Qwen Image)). 2. Pick a Qwen LoRA file from lora_name. 3. Enter layer keywords in layer_filter, for example:

    ```text
    transformer_blocks.0
    cross_attention
    ```
   

   3b. Optionally, add exclude_filter to block certain layers even if included above, for example:

    ```text
    attn,
    transformer_blocks.1
    ```
   

   4. Tune strength_model like any standard LoRA node.

4. Feed the resulting model into a sampler (KSampler, etc.) to continue your editing pipeline.

Example：

Example Workflow

Checkpoint Loader (Qwen Image) → LoadLoraQwenImage → KSampler → VAE Decode → Save Image

Using layer_filter = transformer_blocks.0 limits the LoRA to the first transformer block, enabling fine-grained adjustments to localized styles or directives.

Troubleshooting

• LoRA not applied – Verify that layer_filter hits valid layer names. Leave it empty to test global coverage.
• Nothing changes with both filters – If using exclude_filter together with layer_filter, ensure they don’t cancel each other out (e.g. excluding everything). Try clearing exclude_filter first.
• Loading error – Ensure the LoRA resides in models/loras and targets a Qwen Image-compatible architecture. If logs show missing keys, refine the filter keywords.

---

License

MIT License. Please respect the individual licenses of LoRA assets you deploy.

Note: The node has been validated with LoRAs trained via the ai-toolkit training framework. LoRAs produced by other toolchains may not be compatible—please test them on your own setup. If you encounter issues, feel free to open an issue for assistance.