ussoewwin/ComfyUI-QwenImageLoraLoader

What it solves

This project provides custom nodes for ComfyUI that allow users to apply LoRA (Low-Rank Adaptation) and ControlNet to specific high-performance image models, namely Nunchaku Qwen Image and Z-Image-Turbo. It bridges the gap for these models by enabling style/concept fine-tuning via LoRAs and structural guidance via ControlNet, which are not natively supported in the same way as standard models.

How it works

The project introduces several specialized loader nodes that handle the unique architecture of Nunchaku-quantized models:

• LoRA Stacks: It offers single and multi-LoRA loaders (V1, V2, V3) that can stack multiple LoRAs with adjustable strengths. V2 and V3 include experimental support for AWQ quantized modulation layers to fix noise issues using runtime monkey patching.
• Z-Image-Turbo Support: Dedicated stackers (V1, V4) are provided specifically for Z-Image-Turbo models.
• ControlNet Integration: It implements a DiffSynth route for ControlNet support and a dedicated Krea2 loader that auto-detects whether a LoRA is for depth or openpose control, applying specific injection routes (e.g., expanded first projection for depth or reference-latent injection for openpose).

Who it’s for

• ComfyUI users working with Nunchaku Qwen Image or Z-Image-Turbo models.
• Artists and developers wanting to use LoRAs and ControlNet (specifically Krea2 depth/openpose) with these specific quantized model architectures.

Highlights

• Multi-LoRA Stacking: Supports rgthree-style minimalist UIs for managing multiple LoRAs efficiently.
• AWQ Modulation Support: Experimental fixes for noise in quantized modulation layers.
• Krea2 Specialization: Automatic detection and routing for Krea2 depth and openpose ControlNet LoRAs.
• Broad Compatibility: Includes support for portable ComfyUI installations and automatic disabling of Nunchaku nodes on AMD/ROCm systems to prevent crashes.