With draw things, I have been using high resolution fix to basically set the scale of the items that are rendered to the final resolution. Maybe put "shallow depth of field" in the negative? personally 27k images created since the start of may with draw things. btw do you read the discord?
Hi, my compliments for the nice explanation about draw things. you are really the only one who makes good tutorials about it in English! I have been using draw things for more than half a year now, and it's nice to see how the app is developing. I have a MacBook M2 32GB so luckily it can run quite well. But I'm recently running into a problem that I hope you can help me with. I use inpainting a lot, but now I notice that if I inpaint certain parts and then I zoom the part of inpainting until it falls within the frame. That other parts that I had previously inpainted become lower quality again... have you also had this? and what am I doing wrong? I use the 1280x1280 image size and then a square, original image size setting 1024x1024, target image size 1,280x1,280, negative 960x960 and high resolution fix 768x768 with 6%...upscaler real esrgan x4+ Hopefully you understand what I'm saying and can you help me? 🙏🏼
I have noticed that issue as well! It almost seems like the infinite canvas data gets compressed as more generations are layered on top of one another. I also think there is an issue with the viewport grid snapping sometimes disengaging which will cause generations to be pasted onto the canvas incorrectly.
Good stuff. My problem is how do I increase easiness to access to darker skin tone than bronze... it's there but the randomness don't care enough 😢 Still waiting for my computer to be fixed though 😢
the more 'photoreal' the less melanin.... Ugh! LoRA trained on my dusky character - illustration style she is Black, photography she is some metallic champagne color that only happens in Vogue and luxury automobiles.
I guess we still have to use the old trick of separating the blue channel (less scattering), isolate the skin, then apply the blue as luminosity 🙄which is faster than trying to fit the luminosity curve, but without true references (try finding albedo map of really dark skin) I can't say it's as rich as real dark skin @@CutsceneArtist
@@CutsceneArtist good catch! I wonder how it fares with the hair. Currently I'm without a computer, still missing a hard drive. But I have been using dezgo, a site with some model. Juggernaut xl is quite good with darker skin, skin blemish and natural body shapes, sometimes a bit too much, you'll get gnarly rendering, ALTHOUGH it doesn't seems to know about stretch mark... Though it's a bit too smart, sometimes he slap cloth shapes with skin texture, which give unnatural fold lol. It also has same face syndrome that many model has.
@@CutsceneArtist re: base on example and comment, it doesn't seem to help with the darker skin tone that much. It's still much copper brown. Dezgo allows for Lora insertion, I'll try when I have some times. But I have already some techniques and prompt to access darker skin, and image to image allows me to probe a model affinity (generally lack of) for darker skin. Afro hair typically fairs better once you introduce the keyword: natural hair. Also the model generalizing hair has been great to explore style where that hair don't exist, anime afro hair using natural hair has been a real revelation, given its seldom existing and unlikely to exist in the training set, I would know, I spent years scouring internet just for that. But the way the model do it I can even extract a drawing workflow, you can see how the model figure out the logic from straight hair to 4c hair. I also use image to image with low strength to study and show how wrong modeling techniques are about that type of hair. Ai model also get very inventive when pushed applying afro texture in different style, seeing afro hair rendered with plausible techniques artifact is really cool, for example typing horizon zero dawn or botw as keyword give you aliasing accurate model lol I have a folder entirely dedicated to game history that never was, if game were not casually racist by omissions lol. Also the model get inventive with hairstyle, not shy applying style that aren't seen. Using keyword in vanilla stable diffusion like hiro isono give striking image, even though it's not in the style of the real artist at all lol.
