oppie85

Additional information; it seems like information is being cached in the ClientContext - I've tried working around this issue by automatically splitting the template into unique templates for each page...

I've found limited success in "diluting" the new token by making the prompt more vague - for exmple "a painting of \*" results in pretty much the same image as...

Here's what you can try to verify that textual inversion can create a convincing likeness; First of all train at 256x256 pixels with larger batch sizes; depending on your GPU...

I'd say 2 photos is actually not enough for training a likeness; I use around 10-20 pictures for my experiments. For the 256x256 method it works best to mix in...

You also have to update **num_vectors_per_token** in **v1-inference.yaml** to the same value you trained with. With 50 vectors per token, extreme overfitting is to be expected; I'm currently trying to...

@1blackbar - looks great! Can you share what method you used to achieve this?

> * `embedding_reg_weight` will bring your prompts closer to the `init_word`, but should be used sparringly. Results can vary, but arbitrary low values (like 0.00001) will show you exactly how...

I feel like I could come up with better strategies if I knew more about how the training worked. My current understanding is as follows: Before training, we create a...

> The CLIP encoder has an issue where specific words can dominate the output vector (see e.g. [No token left behind](https://arxiv.org/abs/2204.04908) or the ipod textual attack against CLIP) which seems...

I feel like I'm close to something; I've been able to train an embedding that does succesful style transfer without having to overwhelm the prompt and it seems to ignore...