tom white

Well - I meant something like this 😉  If interested, I can try to clean this up a bit and submit it as a pull request as a pipeline...

> looks great! came looking for the same Thanks - my version is a quick hack, but I cleaned it up a bit and pushed it to a [with_looping branch...

Thanks Matt for this first version. My intuition is that this change would make the canvas 101 pixels wide, but by allowing x=101 and not x=0 as in the original...

Here's a useful unit test for this change. ``` // symmetric canvas test Paper 20 Pen 100 Line 0 0 0 100 Line 50 0 50 100 Line 100 0...

I also tried llava-1.5 and got the same error. Following suggestions online I added ```python model.enable_input_require_grads() ``` after loading the model, which resolved this issue. However, the visual attack code...

Thanks Xiangyu - yes I can confirm that the `liuhaotian/llava-llama-2-13b-chat-lightning-preview` checkpoint you suggest works well with the codebase as-is. I also have made progress in adapting the code to work...

Clue? - for me right now this is happening only when I increase spread. output seems sensible if i ignore it. 🤔

Thanks Leland - encouraged that this idea is not DOA let me motivate this a bit more providing a sample use case, proposed API, and baseline implementation. One use case...

Great to hear you thinking along similar lines. No, I don't have any better implementation - just was trying to communicate that use case more concretely. There is another use...

IIUC: [this recent paper](https://arxiv.org/abs/2406.09413) from @yossigandelsman and others claims that the weight space of diffusion models also has interpretable latent spaces and so perhaps these could also be tested for...