The moment I try to load it into the 9B model, it rejects it with an error message saying "size doesn't match."
"When I load it as a 9B model, I get the error 'mat1 and mat2 shapes cannot be multiplied (512x7680 and 12288x4096).' It works fine with a 4B model, so did you accidentally upload a file intended for 4B?"
The LoRA weights have shape [128, 4096] which matches Klein 9B's hidden dimension — so it's correctly a 9B LoRA. If it's loading on 4B without erroring, it's silently skipping the mismatched weights and not actually applying the LoRA at all. Please refer to the quickstart code in the README for the correct setup.
Thank you for the clarification! I understand that the primary weights are [128, 4096].
However, the issue seems to be specifically isolated to the bs_vton_comfyui.safetensors file.
When we use this specific file in ComfyUI with the Klein 9B base model, the system console spits out multiple shape mismatch errors for 3072 (not 4096) before completely crashing. Here is the exact log:
ERROR lora diffusion_model.double_blocks.3.txt_attn.qkv.weight shape '[9216, 3072]' is invalid for input of size 50331648
RuntimeError: mat1 and mat2 shapes cannot be multiplied (512x7680 and 12288x4096)
The diffusers version might be perfectly fine, but is it possible that the conversion/export script used specifically for the ComfyUI version accidentally packed the 4B (3072) weights instead?
Could you please inspect the tensor shapes inside bs_vton_comfyui.safetensors itself? We would really appreciate it!
Thanks for the detailed report! Just checked the ComfyUI file directly — the QKV weights are [128, 4096] which is correct for Klein 9B. The error you're seeing ([9216, 3072]) suggests the base model being loaded in your workflow might be a wrong checkpoint which would cause exactly this mismatch.
I just added a ComfyUI quickstart section to the README with the full setup instructions including the full workflow — give it a look and let me know if it resolves it!
Thank you so much for the quick reply and for checking the files directly!
You were completely right—it turns out I had the wrong base model checkpoint loaded in my ComfyUI workflow without realizing it. I just downloaded your new ComfyUI quickstart workflow from the README, and it works perfectly with the 9B model now!
I apologize for the false alarm and any confusion it caused. Thank you for going out of your way to add the ComfyUI instructions, it will definitely help a lot of people.
This is an amazing LoRA and I am really excited to use it. Thank you again for your hard work and great support!
Thanks!
Really glad you got it working!
Hope you enjoy the LoRA.