The model was created for personal for editing anything photo, because I love all sorts of shiny realistic stones. The mindless generation of 100500 images of who knows who and why does not excite me at all, because I made the model exclusively for use in IMG2IMG, to simulate the material marble, porcelain interspersed with gold, but it turned out so good that I decided to share it.
I made this model last year, when there were no Textual Inversions, so this is 2GB and not 40kb. The recipe, how I did it, is unfortunately lost from the information of only 12 images used and 10,000 steps, that was a happy time of experimentation, because I can’t repeat this anymore.
Yes, and I barely found the recipe for creating Prompt, it's good that everything inside the pictures is saved.
Token to activate: goldenglass, although it is possible without it, there will be a picture without golden blotches, which usually spread over hair, mustaches, beards, eyebrows and eyelashes.
Prompt: [redshift style] goldenglass a marble statue, highly detailed, smooth, focus, black granite background, film, studio lighting
Negative prompt: cartoon, 3d, disfigured, bad art, deformed, poorly drawn, extra limbs, close up, b&w, weird colors, blurry:0.25
Sampler: DPM++ SDE Karras, DPM++ SDE, DDIM, Euler a
CFG scale: 12-15
Size: 1920x1280, 768x1152, 768x960 (image resolution affects the final result no less than SEED, CFG or Denoise STR)
Steps: 25-40
Denoising strength: 0.25-0.38
Attention!!!
Denoising strength variations can both change the entire model, turning it into marble, and turning only a person into marble, without affecting clothing.
P.S. I would be very grateful if someone could point out similar models that are aimed specifically at IMG2IMG and do not change the recognition of the face, only changing the material or styling of the object.
Description
FAQ
Comments (17)
That's really cool! Say I wanted to teach it new materials, and participate in its development for fun and science, how would I get it to stick? For example, a cactus woman made out of cactus... Or a superhero made out of bread like Superbreadman2? Say I'd use my superbreadman2 embedding on this, would it catch on? I'm impatient to try. I hope it's aware of the existence of bread!
as I wrote above, to my deepest regret, the technology is lost ... I can’t repeat this anymore, although I tried a lot in Textual Inversion and Hypernetwork and in Dreambooth, the only thing that didn’t get to Lora at all .. But I think there are chances not enough.
@djarf What do you mean by "the technology was lost"?
Nevermind, I found it out.
I guess you could reproduce the experiment if you can remember the source materials you used and try to follow a method that resembles the initial one. It would be a good experiment for you to try to redo it from scratch but document everything this time so you can reproduce the experiment and go further with your attempt to create new materials! Good luck with that!
@alltheirbasesnowbelongtous these trainings were carried out even when there was an old dreambooth, by the way, perhaps that's why I couldn't repeat it on the new one, just yesterday the man posted the old version! I highly recommend downloading and doing model training in it https://www.youtube.com/watch?v=usgqmQ0Mq7g this is absolutely accurate, it will significantly improve the result
https://mega.nz/file/iRNXEJaQ#5bXv2TFIUp6B_Pzp2Eu2ic8aSjeg_psFXhQNZv0DDIE
what is the difference between SafeTensor and PeackleTensor?
Safetensor is a newer file format, safer and faster than pickle. Though only a few tools like Automatic111 support safetensor. Most others use only pickle.
@pedxco Fucking dude, you're damn right! You can't even imagine how much you helped me, it doesn't say anywhere that Pickletensor is ckpt !)
@villiance this is a big drawback for the A1111, it just adds everything en masse, but nowhere describing what has changed. I learned about this difference from the video of one blogger
>P.S. I would be very grateful if someone could point out similar models that are aimed specifically at IMG2IMG and do not change the recognition of the face, only changing the material or styling of the object.
Also interested in this
ChromaV5 is amazing for that purpose
@alltheirbasesnowbelongtous for img2img, the face does not break, basically only the "SD 1.5 Inpainting" model with the highlighting of its awareness of 75%, but it does not give much result either, or I did not understand how to use =)
@djarf You mean, with your model? I'll give it a try and tell you. Otherwise there's a thread on the discord that explains how to create an inpainting model. Basically, you use the 3-models method, you put in your model, the 1.5 inpaint model, then the 1.5 pruned model and you get your own inpainting model so you can publish under this model page as an additional asset. https://discord.com/channels/1037799583784370196/1066431771069140992
@alltheirbasesnowbelongtous I tried this method, only it is not applicable to any model, it simply does not merge with something.
@djarf Maybe it's your PC that isn't powerful enough? What kind of a graphics card do you have?
@alltheirbasesnowbelongtous 4090, 48gb RAM, Rhyzen 5 3600, Do you still have doubts?
@djarf That's more than enough to merge checkpoints together! IDK why it fails on your end, but on my end, I had zero issue making a merge between your model and MaxwellTheCat model. There are several new merging tools that just came out, one is called "Merge Board" https://github.com/bbc-mc/sdweb-merge-board then there's "Supermerge" https://github.com/hako-mikan/sd-webui-supermerger and finally, there's "merge-block-weighted-gui" https://github.com/bbc-mc/sdweb-merge-block-weighted-gui
I haven't tried any yet. Tho normally you should have been able to merge using even the base A1111 merging tool too!
@alltheirbasesnowbelongtous interesting, I'll take a look)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.