search

dvruette / sd-webui-fabric

MIT License
402 stars 23 forks source link

[ISSUE] - Images brought in feedback, seem to remain in cache somehow. Even after clearing. Which forces a restart #8

Closed PurpleBlueAloeVera closed 1 year ago

PurpleBlueAloeVera commented 1 year ago

Even after clearing the images. When generating, it still slowed the whole generation process down as if the images were still in feedback. So they must be stuck in cache or something. mmh

zethfoxster commented 1 year ago

image i can confirm this too

dvruette commented 1 year ago

I'm having trouble reproducing this issue. Could you provide some steps to reproduce? Are you using any CLI options? And how much slow down are we talking, in seconds?

PurpleBlueAloeVera commented 1 year ago

No CLI, and you simply load images in, put em in likes, and dislikes. do some generating. Then if you "clear" 'em all. And do generations again, in the console it'll do the whole process as if the images were still there. And slow down the generation time, etc. + the ultimate proof is at the end the images do LOOK like they were influenced by the previous images that seem stuck in cache now. :(

dvruette commented 1 year ago

Hmm, that is weird indeed. Can you try updating to the new version and see if it still happens? It should now tell you exactly how many/which feedback images were used when you select an image (where it usually shows generation params like seed, cfg scale, etc.)

catboxanon commented 1 year ago

Just adding that I cannot reproduce this either. Generation speed goes back to normal if I tick Disable feedback and VRAM usage also returns to normal.

zethfoxster commented 1 year ago

theres more than one issue here...i personally don't have the massive slow down everytime...but in the picture i posted, this is the effects of imgs stuck in the cache, I disabled and enable ,both ways, the lora that creates that character and i get this, the character in rainbow lighting.... this is after "clearing" the influence of what was previously in there remains.

dvruette commented 1 year ago

So it happens after you disable LoRA? That might have something to do with it, it’s possible that the old U-Net forward pass doesn’t get restored correctly, which would have the observed effect.

PurpleBlueAloeVera commented 1 year ago

Happened to me even without any LoRAs. Could it be embeddings ?? I use my own trained embeddings all the time...

github-actions[bot] commented 1 year ago

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.