Forum Replies Created
-
Posts
-
I am reading the paper and the codes. It applies a blur to the latent to the elements selected by the cross attention process, which are supposed to be spatial elements that are important.
It is the latent image that the final image steers away from. I plan to write an article along with a usage guide.
Hi!
-
LoRA is better than TI, in the sense that it can change the model more.
-
Typically instances with T4 or V100 GPU. You can stick with 1 GPU for inference. AWS is not the most cost effective. You can consider Google Colab if you are ok with the interactive interface. They offer the same GPUs at a much lower rate.
Hi! You are setting denoising strength. Are you using img2img? These settings should work on txt2img.
Some other settings I can think of:
- Have you enabled the controlnet?
- Try the SAI canny lora (select through controlnet)
If it still doesn’t work, would be helpful to upload a screenshot of settings.
The easiest way is to copy from another node.
Here’s an example node file on comfyui.
https://github.com/comfyanonymous/ComfyUI/blob/master/custom_nodes/example_node.py.example
Miya, sorry for the late response.
I have updated your profile pic. If you have another one you like, please upload it here so I can update!
I can suggest an approach that requires some coding…
In the clip interrogator, you can define a list of keywords to check against an image.
https://github.com/pharmapsychotic/clip-interrogator/tree/main/clip_interrogator/data
You can replace them with the set of keywords you want.
Use clip interrogator to automatically select the keywords for an image, and write it to the image metadata.
-