Describe the bug
When a Qwen-Image LoRA's keys are already named lora_A/lora_B, _convert_non_diffusers_qwen_lora_to_diffusers discards each module's .alpha instead of applying it. A LoRA trained with alpha ≠ rank therefore loads at the wrong strength. The lora_down/lora_up path and the Z-Image converter already fold alpha / rank into the weights.
For example, e-n-v-y/Qwen-Image-2.1-Fix-v2.0 is rank 32 with alpha 83 and 96, so it loads at roughly a third of its intended strength. ai-toolkit and ComfyUI LoRAs commonly use this layout. It affects Qwen-Image (20B) and Qwen-Image-2.1 alike.
Reproduction
import torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from diffusers import QwenImage21Pipeline
repo, weight = "e-n-v-y/Qwen-Image-2.1-Fix-v2.0", "qwen2.1-detail-fix-2.0.safetensors"
raw = load_file(hf_hub_download(repo, weight))
sd = QwenImage21Pipeline.lora_state_dict(repo, weight_name=weight)
print(sum(k.endswith(".alpha") for k in raw), "alpha keys in the file,",
sum(k.endswith(".alpha") for k in sd), "after conversion")
a = next(k for k in raw if k.endswith("lora_A.weight") and k.replace("lora_A.weight", "alpha") in raw)
converted = "transformer." + a.replace("diffusion_model.", "")
print("unscaled:", torch.equal(sd[converted].float(), raw[a].float()))Logs
30 alpha keys in the file, 0 after conversion
unscaled: True
# first module: alpha 83, rank 32 -> should load at 2.59x what it does
System Info
diffusers main at 8b33bfc, Python 3.14, torch 2.14.1, peft 0.21.2.
Who can help?
@sayakpaul
Describe the bug
When a Qwen-Image LoRA's keys are already named lora_A/lora_B, _convert_non_diffusers_qwen_lora_to_diffusers discards each module's .alpha instead of applying it. A LoRA trained with alpha ≠ rank therefore loads at the wrong strength. The lora_down/lora_up path and the Z-Image converter already fold alpha / rank into the weights.
For example, e-n-v-y/Qwen-Image-2.1-Fix-v2.0 is rank 32 with alpha 83 and 96, so it loads at roughly a third of its intended strength. ai-toolkit and ComfyUI LoRAs commonly use this layout. It affects Qwen-Image (20B) and Qwen-Image-2.1 alike.
Reproduction
Logs
System Info
diffusers main at 8b33bfc, Python 3.14, torch 2.14.1, peft 0.21.2.
Who can help?
@sayakpaul