Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
Does anyone have any prompt advice for making SDXL better at proportions and perspective? For example in my profile picture the sunglasses are visibly off, with the further away lens being significantly larger than the closer one. SD1.4 tricks for making better glasses like putting "[girlswithglasses:0.7]" at the end of the prompt don't seem to have any effect on XL.
 
snov said:
Does anyone have any prompt advice for making SDXL better at proportions and perspective? For example in my profile picture the sunglasses are visibly off, with the further away lens being significantly larger than the closer one. SD1.4 tricks for making better glasses like putting "[girlswithglasses:0.7]" at the end of the prompt don't seem to have any effect on XL.
Click to expand...
Can you post the original .png with the attached generation metadata? It's way easier to just take a quick look at that than to play 20 questions.
 
inception_state said:
Can you post the original .png with the attached generation metadata? It's way easier to just take a quick look at that than to play 20 questions.
Click to expand...
tmpluggjcui.png
Exiftools can't find any generation metadata, is there some other tool I should be using? I'd love to be able to dissect what prompts other people were using when I find their creations.
Prompt is (ray-ban aviator's sunglasses), Cyberpunk lofi art, facing forwards, futuristic cityscape backdrop, moody neon lighting, retro-futuristic vibes, detailed character design, pixel art elements, cyberpunk aesthetic, 4K resolution, inspired by Syd Mead's Blade Runner concept art and Ariana Grande cover art
I'm doing 50 steps of SDXL, switch to refiner at 0.8, seed for this one was 1650234854, and I'm using a 4090.
 
  • Like
Reactions: Vecr, SIMIΔN and inception_state
snov said:
View attachment 5850007
Exiftools can't find any generation metadata, is there some other tool I should be using? I'd love to be able to dissect what prompts other people were using when I find their creations.
Prompt is (ray-ban aviator's sunglasses), Cyberpunk lofi art, facing forwards, futuristic cityscape backdrop, moody neon lighting, retro-futuristic vibes, detailed character design, pixel art elements, cyberpunk aesthetic, 4K resolution, inspired by Syd Mead's Blade Runner concept art and Ariana Grande cover art
I'm doing 50 steps of SDXL, switch to refiner at 0.8, seed for this one was 1650234854, and I'm using a 4090.
Click to expand...

I've gotten the metadata from loading it into stable diffusion. But it depends on if the metadata is pruned or not (eg. 4chan prunes it). In the early days, may still be the case, people used catbox to share images so that metadata was preserved and you could see the full prompt and seed.
 
  • Like
Reactions: Vecr, Susanna and inception_state
snov said:
View attachment 5850007
Exiftools can't find any generation metadata, is there some other tool I should be using? I'd love to be able to dissect what prompts other people were using when I find their creations.
Prompt is (ray-ban aviator's sunglasses), Cyberpunk lofi art, facing forwards, futuristic cityscape backdrop, moody neon lighting, retro-futuristic vibes, detailed character design, pixel art elements, cyberpunk aesthetic, 4K resolution, inspired by Syd Mead's Blade Runner concept art and Ariana Grande cover art
I'm doing 50 steps of SDXL, switch to refiner at 0.8, seed for this one was 1650234854, and I'm using a 4090.
Click to expand...
If you drop an image into the PNG info tab in A1111, it should show the generation metadata. As @Lemmingwiser said though, some sites like 4chan scrub it or automatically downsize the image for efficiency in a way that removes metadata. The "send to txt2image" button is incredibly convenient though if there is metadata, you can just drag and drop the image instead of messing around with a bunch of settings.

I can't seem to generate the exact image you did, we must have a difference in setup somewhere. It seems to help to add "depth of field", "perspective", "bokeh", etc. Describing the angle of the photo can help as well, eg "three-quarters portrait shot". Sunglasses do tend to be a bit wonky, but the reflection effects can be neat.

Basically your prompt with minor tweaks:
24997-1650234855-korean woman wearing a leather jacket and (ray-ban aviator's sunglasses), Cyb...png24998-1650234856-korean woman wearing a leather jacket and (ray-ban aviator's sunglasses), Cyb...png25006-3141334434-three-quarters portrait shot of a korean woman wearing a leather jacket and (...png
Me messing around with different stuff:
24865-1206266693-a close-up three-quarter portrait of a beautiful korean woman with long dark ...png24864-1206266692-a close-up three-quarter portrait of a beautiful korean woman with long dark ...png24817-1788761836-The image portrays a close-up of an asian woman with long dark hair wearing r...png

Edit: can confirm, KF doesn't scrub generation metadata, so it must have happened previously. The metadata seems fine on the images I posted above.
 
Last edited:
  • Feels
Reactions: Vecr and Susanna
inception_state said:
I can't seem to generate the exact image you did, we must have a difference in setup somewhere. It seems to help to add "depth of field", "perspective", "bokeh", etc. Describing the angle of the photo can help as well, eg "three-quarters portrait shot". Sunglasses do tend to be a bit wonky, but the reflection effects can be neat.
Click to expand...
I could have sworn "¾ portrait" was in there somewhere, I must have accidentally'd it when I was experimenting with the prompt. I quite like the ones you made! First one second row in particular, shame it messed up on making them actual aviators, though it's close enough you could photoshop them into shape easily. I can't reproduce any of yours either, so there's definitely something different between our setups. Oh well. Mine still makes wonky sunglasses, but I guess I can just use liquify to fix that.
 
Obviously, check your metadata thoroughly before sharing. ComfyUI can definitely leak paths and stuff which could reveal real-world information. For instance /home/david/ would tell you what my real name is.

Annoyingly, some of the defaults in Comfy also don't save enough information. I like giving random checkpoints wildcard prompts and without a couple extra nodes the exact checkpoint and parsed wildcard wasn't being saved. So I had to augment with a Save node which let me save that data into an additional Exif field.
 
  • Like
Reactions: Vecr, Mr.Miyagi and inception_state
Been out of the ML scene for a bit now and want to get back into the swing of things. What are the best models for photorealism currently?
 
Dishonourable Harlot said:
Been out of the ML scene for a bit now and want to get back into the swing of things. What are the best models for photorealism currently?
Click to expand...

epicRealism and the detail tweaker lora do a very good job for most things regarding photorealism but sometimes the people it generates wind up with really bizarre clothing. You could probably fix it with another lora but I don't bother since it does a great job at scenery/backgrounds/buildings for wallpaper.

ComfyUI question: Anyone have a good workflow for testing the same prompt vs a bunch of checkpoints, preferably with some compositing to combine and annotate the results?

eta: I think I have a new favorite lora.
 
Last edited:
  • Informative
Reactions: Vecr and Dishonourable Harlot
ducktales4gameboy said:
ComfyUI question: Anyone have a good workflow for testing the same prompt vs a bunch of checkpoints, preferably with some compositing to combine and annotate the results?
Click to expand...
I'm traveling so I don't have access to my workflows and scripts. But as I recall I made a text file with all the checkpoints and used the line in file node(WAS I think) to feed a checkpoint loader. Also the text name of the checkpoint I use a special save routine to save it to a different Exif field. The default exif won't save the checkpoint name in the default workflow when you load them indirectly like this. Alternatively you can do similar to use filenames or directories to sort them.

Then what I did is wrote a script that just makes an HTML page with each(or multiple) images from each checkpoint based on the filename or Exif tag and annotations.

Same script I can do things like wildcards and random LORAs and see how those turn out. The custom exif field I also have saving the derived prompt after the wildcards are parsed.

Today I have a REST web service that returns the next checkpoint name this way I can share the workflow among multiple instances of Comfy and not get duplicate checkpoints.
 
  • Informative
Reactions: ducktales4gameboy
ducktales4gameboy said:
epicRealism and the detail tweaker lora do a very good job for most things regarding photorealism but sometimes the people it generates wind up with really bizarre clothing. You could probably fix it with another lora but I don't bother since it does a great job at scenery/backgrounds/buildings for wallpaper.

ComfyUI question: Anyone have a good workflow for testing the same prompt vs a bunch of checkpoints, preferably with some compositing to combine and annotate the results?
Click to expand...
With automatic111 you could use that replace function with a grid.
 
I've used the auto grid before, this particular interest in doing it in Comfy is due to having access to a shared 4090 setup that's only Comfy and has something like 850 checkpoints installed.

Also worth bringing up because it's unfathomably based: Civitai reworked all their content filters and the default browsing filters block every model tagged 'anime' or 'furry'.
 
  • Like
Reactions: Overly Serious, macrodegenerate, woowie queen and 1 other person
I know this probably isn't the best place to ask this, but I'm training a Lora again and I'm worried about tokens not matching.

Do underscores matter in stable diffusion or nai prompts? For example: is red_dress the same as red dress in a prompt. I'm worried that red_dress is becoming a trigger word as opposed to just a normal tag.
 
ducktales4gameboy said:
Also worth bringing up because it's unfathomably based: Civitai reworked all their content filters and the default browsing filters block every model tagged 'anime' or 'furry'.
Click to expand...
I was actually able to browse through the site at work the other day during some down time. Love not seeing anime and furry shit, big booba still gets through sometimes unfortunately.
 
You must log in or register to reply here.
Back