modelscope
gradio.exceptions.Error: '没有人物LoRA，请先训练(There is no character LoRA, please train first)
问题描述： 在 Windows web-ui 本地部署时， 人物形象训练阶段，基础模型选择 “AI-ModelScope/stable-diffusion-v1-5”，则在 “无限风格形象写真” 阶段只能选择 “写实模型（Realistic sd_1.5 model）”，选择 “写真模型（Photorealistic_sd_1.5 model）” 则会报错： gradio.exceptions.Error: '没有人物LoRA，请先训练(There is no character LoRA, please train first)
观察 app.py 中的代码发现，当选择写真模型时，它就会去 AI-ModelScope/stable-diffusion-xl-base-1.0/ 目录下查找 pytorch_lora_weights.bin 文件，当然就找不到了。 我的疑问是：
- 为啥 sd 1.5 model 的基础模型，选择写真风格时，要去 AI-ModelScope/stable-diffusion-xl-base-1.0/ 目录找模型文件？
- 为啥只有本地部署时才会有这个问题？我使用官网的在线体验，基础模型是 sd 1.5 model 时，在无限风格形象写真那里是可以选择写真模型的。
此外， 而如果在人物形象训练阶段，基础模型选择 AI-ModelScope/stable-diffusion-xl-base-1.0，则在无限风格写真那里选择写真模型时，由于存在 AI-ModelScope/stable-diffusion-xl-base-1.0/ 目录，所以报不会报 gradio.exceptions.Error: '没有人物LoRA，请先训练(There is no character LoRA, please train first) 的错误，但是却发生另一个错误： File "D:\workspace\xxxxx\extensions\facechain\facechain\merge_lora.py", line 85, in merge_lora curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up, weight_down) RuntimeError: The size of tensor a (768) must match the size of tensor b (2048) at non-singleton dimension 1 提示：Python 运行时抛出了一个异常。请检查疑难解答页面。
这又是怎么回事呢？
有没有大神可以指点迷津一下，非常感谢！
[DDDDD] ================================================================ [DDDDD] weight_up shape: torch.Size([640, 4]) [DDDDD] weight_down shape: torch.Size([4, 2048]) [DDDDD] curr_layer.weight.data shape: torch.Size([640, 768]) [DDDDD] else, weight_up shape: torch.Size([640, 4]) [DDDDD] else, weight_down shape: torch.Size([4, 2048]) [DDDDD] else, alpha: 1.0 [DDDDD] else, multiplier: 0.9 concurrent.futures.process._RemoteTraceback: """ Traceback (most recent call last): File "D:\workspace\xxxxx\python\lib\concurrent\futures\process.py", line 246, in _process_worker r = call_item.fn(*call_item.args, **call_item.kwargs) File "D:\workspace\xxxxx\extensions\facechain\facechain\inference.py", line 559, in call gen_results = main_model_inference(self.pose_model_path, self.pose_image, self.use_depth_control, File "D:\workspace\xxxxx\extensions\facechain\facechain\inference.py", line 438, in main_model_inference return main_diffusion_inference(pos_prompt, neg_prompt, input_img_dir, base_model_path, File "D:\workspace\xxxxx\extensions\facechain\facechain\inference.py", line 195, in main_diffusion_inference pipe = merge_lora(pipe, lora_human_path, multiplier_human, from_safetensor=lora_human_path.endswith('safetensors'), device='cuda') File "D:\workspace\xxxxx\extensions\facechain\facechain\merge_lora.py", line 92, in merge_lora curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up, weight_down) RuntimeError: The size of tensor a (768) must match the size of tensor b (2048) at non-singleton dimension 1
显然，报错的原因是 curr_layer.weight.data 的形状是 [640, 768]，无法与 [640, 2048] 的矩阵相加。
please try out the newest train-free, 10s inference version facechain-fact.
