nodes/Extraltodeus-ComfyUI-AutomaticCFG
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Update README.md

Browse Source
This commit is contained in:
Extraltodeus
2024-05-14 21:28:39 +02:00
committed by GitHub
parent 2e9558b70a
commit 5bd842517e
1 changed files with 8 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+8 -1
View File
@@ -2,7 +2,14 @@ My own version "from scratch" of a self-rescaling CFG / anti-burn. It ain't much
## Last update (14.05.24):
- added node: **preset loader**. Can do what the other can and MUCH MORE! (note: mostly tested with SDXL, the "Red riding latent" preset does not work with SD1.5)). The presets are .json files and can contain a string which will go through eval(). Always check what is inside before running it.
- added node: "**Excellent attention**" developped by myself and based on this [astonishingly easy to understand research paper!](https://github.com/Extraltodeus/temp/blob/main/ihave.jpg) But in short This node allows to disable the input layer 8 on self and cross attention. But also to apply a custom modification on cross attention middle layer 0. I have only tested with SDXL. Also for some reason the Juggernaut model does not play too well with it. You can find a grid example of this node's settings in the "grids_example" folder.
- added node: "**Excellent attention**" developped by myself and based on this [astonishingly easy to understand research paper!](https://github.com/Extraltodeus/temp/blob/main/ihave.jpg) But in short:
- This node allows to disable the input layer 8 on self and cross attention.
- But also to apply a custom modification on cross attention middle layer 0.
- While the modification is definitely not very ressource costy, the light patch uses less VRAM.
- The multiplier influences the cross attention and may actually reinforce prompt-following. Like for real.
- I have only tested it with SDXL.
- You can find a grid example of this node's settings in the "grids_example" folder.
- For some reason the Juggernaut model does not work with it and I have no idea why.
- Attention modifier presets! Try them with the preset loader!
- I discovered that disabling the input layer 8 on the cross or self attention tends to make more random images. This layer seems to be the main responsible for the overall composition while not being alone at the task. Disabling it for the self attention for generating the positive denoised is risky regarding the image coherence but tends to create more interesting images more driven by your prompt. For the cross attention it creates more coherent images with a bit less prompt following. The results are most interesting. Doing it in the negative seems to be an improvement. The only rule is to keep at least one of the attentions for the positive. You will find the possibility to do experiment about that through the presets loader and the "excellent attention" node.