- This topic has 104 replies, 7 voices, and was last updated 2 years, 1 month ago by
.
-
Posts
-
@frank22 Are you using A1111 or SD Forge for the interface to SD? It had been on my radar for a while to upgrade to the SD Forge, finally got around to doing it. Much improved in every way. Speed is significantly quicker, even on renders much quicker. And a ton of new options and things to tinker with.
Hi, I'm still using a1111 (1.5). I'd really like to try DiffEx but I'm scared I might stuff up something if I have to install a new version of python. Same with Forge if it installs as a platform on top of the Stable Diffusion WebUI.
My install of SD Forge was a breeze. I did have to manually add 1 line to the web-user.bat file to launch it, though, and it required me to look up a cryptic error message to figure out the issue. Otherwise it was well worth it.
You can try this though, copy your entire stable-diffusion folder and make a backup of it with a different name, like backup-SD. Then do the install on the real folder. If something goes wrong, delete the folder if its all screwed and rename the backup folder back to the real name.
A VERY worthwhile watch...
https://www.wired.com/video/watch/tech-support-prompt-engineer-answers-ai-questions-from-twitterAs he said, with prompting it's a good idea to give examples, particularly with styles, ie. in the style of a 1950s comic book.
Even though my Superman lora was crap, I think the prompt I got [sigh... stole] from here helped a great deal. I don't think I could write such a detailed prompt myself. I did add extra details but not sure it helped.
@frank22 Forge has been sorta discountinued now. The project was taken by another guy and rebranded reforge. It uses the same version of python and uses less vram.
I am back on A1111 since it works great for me, though.
As for the complex prompt, it is weird that it works in 1.5, it is more in the style of sdxl.
I think I'm so wary because when I got my new PC last September I experienced huge problems getting a1111 set up and running. It would work one day but the next nada. Now it's stable I'm afraid to change anything.
I found a site that extracts meta data from Loras. Not only the training data but also the captions used. Using an actual training caption can give more specific and accurate results when using the lora.
Do you mean by taking one of the training captions that were used during training, and use that caption as a prompt to generate a new image? I guess it could help, but not if you are looking for output of something different than the caption is saying.
I think it would be useful to make a lora adhere more to your prompt than it is for repeatablilty.
Repeatability is definitely achievable in SD. I use SD mainly for post-processing stuff. Very rarely do I use it for just prompting and spitting out an image. But for the things that I do use it for, repeatability is not a problem, there's a lot of options in there to get exactly what you need. Theres so many ways to skin a cat in there.
"Do you mean by taking one of the training captions that were used during training...."
Yeh gts6, from the view of using a training caption, or specific words/phrases from it, as a prompt, similar to how trigger words work.
It would seem to make sense considering how captioning and trigger words affect Lora behaviour. I guess I mean that using specific words/phrases from captions will generate images more reliably than otherwise.
A prompt that includes the phrase "a girl riding a bike in a park" won't work as well as "a girl in a park riding a pushbike" if the latter is the caption description.I guess it could come in handy if you are unsure of what the trigger word is. When I trained anything, I always made sure I added a triggerword to the prompt, that usually sufficed. But if using one that someone else trained, that could definitely come in handy for sure, to figure out a triggerword.
This is just a single HTML file that runs locally on your machine, no need for internet access, though it has additional formatting and civitai features if you do.
It separates all Lora metadata into sections, including the TAGs/trigger-words, training data, resolutions, etc.
Again, this runs locally, it doesn't need internet access. Drag/drop any safetensors Lora into it and it'll extract the metadata.
I copy-paste the data I want into a 'Lora-name'.txt file for every Lora I use.
https://github.com/Xypher7/lora-metadata-viewer/archive/refs/heads/main.zip
- You must be logged in to reply to this topic.
