Closed 311-code closed 1 week ago
i will look more into this when i have more time. thank you for this detailed dive and work you put in! i added it to the node!
No problem, I think I made some progress with SD3 now at least but not really certain. Feel free to use any of the code I posted here if you find some of it useful. https://github.com/cubiq/prompt_injection/issues/12#issuecomment-2183170554
Forked here, the svd and sd3 injection I did get working and will post the code for that there soon. It was very difficult and still have doubts I'm doing sd3 right. https://github.com/brentjohnston/Magic-Prompt-Injection-SDXL-SD15
Disclaimer: This does not work yet.
Posting this here because this repo helped me a ton (and the other fork).
I actually got clip conditioning working to some extent for injecting svd a little with this repo's ideas and learning about clip text embedding and injecting weight and bias layers with images and text. I updated the svd_img2vid_conditioning (will create repo)
Here I've included the various SVD model probing results I got back, where I put comment dummy data to scan for hidden inputs and outputs in the svd model and just exploring this.
Updated the prompt_injection.py with Attn2 Prompt Injection Node. This is all wrong after review, but maybe can use as a framework. This week I'm using the pipeline_stable_video_diffusion.py for how to actually do the embeddings right and will update this in the future.
Here are a bunch of SVD probe results for hidden input/outputs I meantioned above. I was mostly interesting in CLIPTextTransformer, CLIPVisionTransformer Edit: I did find a way to get clip conditioning working now with it to some extent, and additional guidance image inputs, will make a repo soon: