It feels like the model has overfitted the training dataset...

#2
by SelimAI - opened

Hello,
I'm very interested in your work about SVGs and how you can improve the output provided by LLMs. However, I've encountered significant brittleness when testing OmniSVG1.1_8B, even with recommended prompts.

For example, here is a recommended prompt that works as expected :
image
but here is the same prompt, only for a different color :
image
same thing here, the adherence to the prompt isn't great either :
image

Here is another recommended prompt :
image

but changing colour once again gives some unexpected result.
image

For information, I'm using OmniSVG1.1_8B on default settings.

Am I doing something wrong here ?

Thanks in advance.

Sign up or log in to comment