- Jan 24, 2020
- 1,401
- 3,793
I tried only a bit of this, but I'm not too bothered tbh. I was only helping Nano. But thanks for the tip.
You must be registered to see the links
You must be registered to see the links
interesting to look at styles
That makes me want to make a spreadsheet for a prompt creator for SD, but it's a really monumental task.
If anyone around here wants to help along in a shared gdocs document, hit me up. Something along the lines of
You must be registered to see the links
but for SD rather than Midjourney. SD is kinda..opaque as far as prompting and prompting optimization goes and the best I've seen of such a tool is
You must be registered to see the links
that I pretty much ignored it and resorted to trial an error to the dismay of a few people around the forums.
- Sep 15, 2021
- 894
- 3,294
In reply to farore in the Show Us Your AI Skills thread:
Firstly, I tested which would be the best Model to use:
Personally, I think the one that looks most like her (not necessarily which looks the best, we're going for accuracy here) is with DeliberateRealistic_v10.
I'd already run some tests to determine the best CFG, ended up with 10.
This is the result:
I then copied the .pt file and renamed it to WQZSW.pt and changed the prompt accordingly.
There was no difference between that and using ZOE_VOSS3.pt.
I think the female-orientated results you got with your ZOE_VOSS5 test was probably just that you were using a model that specialises in women* and it just pulled something up, trying in some way to use the (to it) random string of characters you'd entered (I noticed one tshirt with approximate text on it). The underscore prevents SD splitting it into two tokens as it would if I'd just used a space.
*Aren't we all?!
I managed to finish the training on ZOE_VOSS3 (best results at 18000 steps!) and ran some tests.
Firstly, I tested which would be the best Model to use:
Personally, I think the one that looks most like her (not necessarily which looks the best, we're going for accuracy here) is with DeliberateRealistic_v10.
I'd already run some tests to determine the best CFG, ended up with 10.
This is the result:
I then copied the .pt file and renamed it to WQZSW.pt and changed the prompt accordingly.
There was no difference between that and using ZOE_VOSS3.pt.
I think the female-orientated results you got with your ZOE_VOSS5 test was probably just that you were using a model that specialises in women* and it just pulled something up, trying in some way to use the (to it) random string of characters you'd entered (I noticed one tshirt with approximate text on it). The underscore prevents SD splitting it into two tokens as it would if I'd just used a space.
*Aren't we all?!
- Apr 11, 2022
- 75
- 74
In reply to farore in the Show Us Your AI Skills thread:
I managed to finish the training on ZOE_VOSS3 (best results at 18000 steps!) and ran some tests.
Firstly, I tested which would be the best Model to use:
Personally, I think the one that looks most like her (not necessarily which looks the best, we're going for accuracy here) is with DeliberateRealistic_v10.
I'd already run some tests to determine the best CFG, ended up with 10.
This is the result:
I then copied the .pt file and renamed it to WQZSW.pt and changed the prompt accordingly.
There was no difference between that and using ZOE_VOSS3.pt.
I think the female-orientated results you got with your ZOE_VOSS5 test was probably just that you were using a model that specialises in women* and it just pulled something up, trying in some way to use the (to it) random string of characters you'd entered (I noticed one tshirt with approximate text on it). The underscore prevents SD splitting it into two tokens as it would if I'd just used a space.
*Aren't we all?!
why are these guys using strange/unambigous default triggers?
You must be registered to see the links
j0rd7nj0n3s
You must be registered to see the links
koh_sashagrey
You must be registered to see the links
j3nna0rt3ga
just check this guy:
You must be registered to see the links
- Oct 29, 2017
- 812
- 2,167
Because those persons are already in the dataset of the 1.5 model of stable diffusion.
Even if you don't use those loras you can go and write Sasha Grey in your prompt and you'll get a lookalike.
Those loras are just more specific to those characters.
edit: forgot to add that the dataset may include 'outdated' versions of those persons, like in the case of jenna ortega, you write her name and get almost childlish version of her, which is not what people wants to see, they mostly want to see the version of her that looks like wednesday.
- Apr 11, 2022
- 75
- 74
well, these are textual inversions...
so anyway zoe_voss3 is not a good default, because people use other checkpoints. I seldom use SD at all
and maybe in future zoe voss is going to be added to SD
- Oct 29, 2017
- 812
- 2,167
Textual inversions and Loras serve practically the same purpose.well, these are textual inversions...
so anyway zoe_voss3 is not a good default, because people use other checkpoints. I seldom use SD at all
and maybe in future zoe voss is going to be added to SD
Most checkpoints are based on SD 1.5 so that's what I meant.
- Jan 24, 2020
- 1,401
- 3,793
There are prompt generators already and you also have scripts with prompt functions such as "prompt matrix".
To get the prompt generator go to extensions/available and press "load from", the correct adress is written by default.
Then find "Prompt Generator" and press "install". You obviously need to restart SD.
img2img has an "Interrogate" function that scans the image and then ads a prompt automatically.
Simply press "Interrogate Clip" if it's a realistic photo etc or "Interrogate DeepBooru" if it is anime.
Then wait until the prompt is "Interrogated".
Last edited:
I had no idea what those Interrogate buttons did. I tried one from the extensions but was really bad, didn't try others after it.
- Oct 29, 2017
- 812
- 2,167
They don't give you any magic information, but it doesn't hurt to have some approximation words to get started.
- Jun 6, 2020
- 44
- 630
you set --theme dark in your COMMANDLINE_ARGS in the webui-user.bat
set COMMANDLINE_ARGS= --theme dark
- Sep 15, 2021
- 894
- 3,294
Had you not done the bare "zoe_voss5" test I would have said that there's no ambiguity and "Zoe_Voss3" is fine as a prompt.well, these are textual inversions...
so anyway zoe_voss3 is not a good default, because people use other checkpoints. I seldom use SD at all
and maybe in future zoe voss is going to be added to SD
At the moment it seems to work because it's a TI and therefore, if SD has an exact match for the TI filename string and has no other possible value for that token it uses the TI 100%. In the event that Zoe Voss is added into the training data for a model it may pull something from that as well / instead. Not an issue for SD1.5 itself due to the NSFW nature of most of her pictures, but some specifically trained checkpoints may be affected.
I think I'll be using unambiguous strings in future, just to be 100% sure e.g. "z0ev0ss3". Can't be arsed changing the filenames on what I've already uploaded to Civitai though, especially given it's hypothetical at the moment as no effect has yet been found.
- Jan 24, 2020
- 1,401
- 3,793
Interesting read about why AI Image Generators has such difficulty getting hands correct.
You must be registered to see the links
- Mar 11, 2023
- 46
- 132
Hi all, amazing thread. I just got here but I've been playing with sd for a few weeks. Trying out Nano and MrFox's recommendations on the neverendingdreamNED_bakedVAE model:
Works perfectly fine for me if I use a 512x512 resolution, but I get some pretty strange facial deformations when I try with 448x640 hires by 1.75x. The NON-hires, non-upscale (native 448x640) via txt2img is fine, but even if I remove the LORAs, my faces are always deformed by the txt2img+SD upscale at a non 1:1 aspect ratio. I've tried with a lower (0.45) denoise, and it's still messed up.
This seems to be a problem specifically with the hires fix.
Just using img2img upscale alone is inferior, and removes all the skin textures, so that's no use...
TLDR:
512x512 hires fix upscales via txt2img looks great, works fine.
Non 1:1 ratio hires fix upscale via txt2img ... sucks.
Does anyone have any suggestions? I wonder if it's an Apple Silicon issue? I've googled and reddited to no avail.
beautiful, masterpiece, best quality, hiqcgbody, anime, 1girl, sam yang, halter cut dress, mascara, flirty, blunt bangs, portrait shot, looking at viewer, glossy lipstick, nightclub, intricate details,<lora:hipoly3DModelLora_v10:0.5> <lora:samdoesartsSamYang_normal:0.5>
Negative prompt: (painting by bad-artist:0.9), easynegative, (worst quality, low quality:1.4), watermark, text, error, blurry, jpeg artifacts, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, artist name, bad anatomy, deformed iris.
Size: 448x640, Seed: 3686603218, Steps: 28, Sampler: DPM++ SDE Karras, CFG scale: 7, Clip skip: 2, Model hash: d8722b4a4d, Hires steps: 20, Hires upscale: 1.75, Hires upscaler: Latent (bicubic antialiased), Denoising strength: 0.55
Works perfectly fine for me if I use a 512x512 resolution, but I get some pretty strange facial deformations when I try with 448x640 hires by 1.75x. The NON-hires, non-upscale (native 448x640) via txt2img is fine, but even if I remove the LORAs, my faces are always deformed by the txt2img+SD upscale at a non 1:1 aspect ratio. I've tried with a lower (0.45) denoise, and it's still messed up.
This seems to be a problem specifically with the hires fix.
Just using img2img upscale alone is inferior, and removes all the skin textures, so that's no use...
TLDR:
512x512 hires fix upscales via txt2img looks great, works fine.
Non 1:1 ratio hires fix upscale via txt2img ... sucks.
Does anyone have any suggestions? I wonder if it's an Apple Silicon issue? I've googled and reddited to no avail.
Last edited:
- Sep 15, 2021
- 894
- 3,294
That is an odd one!
Have you tried a different upscaler? ESRGANx4 is the one I normally use. Try a couple, see if that isolates it.
Also, I'd generally recommend not taking a starting resolution below 512, and then always working in multiples of 64 pixels. Closest to your aspect ratio would be 512x704. Upscale by 1.6 or so should give about the same finished size
- Jan 24, 2020
- 1,401
- 3,793
I don't know what could cause this. I'm currently scanning through
to see if anyone has something like this and if there's an answer.
You could post your issues there and someone there might be able to help if no one here has an idea.
You must be registered to see the links
to see if anyone has something like this and if there's an answer.
You could post your issues there and someone there might be able to help if no one here has an idea.
- Mar 11, 2023
- 46
- 132
Yeah, it is kind of strange - I'm just copying exactly what Nano and MrFox were doing - that's why I was so surprised at the squashed facial features. Those are their settings I'm using (which originally came from the civitai post). I'm just trying to replicate at this point before further experimentation.
- Jan 24, 2020
- 1,401
- 3,793
It is possible it's an apple specific bug and if so I don't know if there is any solutions. Maybe there is a forum or thread somewhere for apple users that could provide info and/or solutions. I did a quick google search but didn't find anything yet.