Anonymous
8/20/2025, 10:35:51 PM
No.11359983
>>11359649
A nice idea though unsure of how it'd interact when having things like guro and blood heavy in the negatives. Should at least try it and see, thanks.
>>11359869
Yeah just learn your models uses, the tools and the proper way to prompt. Loras aren't needed as often now, especially for illustrious models the text parser can be pretty exact with even low sample danbooru tags, but if you try writing freeform it'll still tokenise the individual words as though they're tags. Generally good for things like "holding [specific object]", "sitting/leaning/standing on [thing]", or even "[character] cosplaying as [character]" but don't try for anything more elaborate. Tokens can also be confused within danbooru tags, like "chopper_(motorcycle)" understands the concept but is still going to see it as chopper from one piece and might sneak that in somewhere, same for something like "pear-shaped_figure", that's where negatives help. Get used to using things like from side/above/below, foreshortening/perspective, full body/cowboy shot/upper body/portrait etc, can do a lot to make a 1girl solo image more interesting to look at.
Controlnet is the most useful tool there is for txt2img, I use depth constantly to control the image composition, especially useful alongside regional prompting. Reference is really useful for concepts the model lacks grasp on. Always use "prompt is more important".
For inpainting get a feel for functional ratios of mask size, denoise, resolution, and only masked padding pixels, alongside adjusting the prompt right. I usually make a 1014x1024 aspect ratio equivalent then upscale by 1.4 and use mask padding between 40-120 depending on how much of the rest of the image the section needs to align with or see to understand the subject. Denoise at 0.4 can do more than you'd think and I rarely go as high as 0.6. Larger mask generally means less detail and difference also.
A nice idea though unsure of how it'd interact when having things like guro and blood heavy in the negatives. Should at least try it and see, thanks.
>>11359869
Yeah just learn your models uses, the tools and the proper way to prompt. Loras aren't needed as often now, especially for illustrious models the text parser can be pretty exact with even low sample danbooru tags, but if you try writing freeform it'll still tokenise the individual words as though they're tags. Generally good for things like "holding [specific object]", "sitting/leaning/standing on [thing]", or even "[character] cosplaying as [character]" but don't try for anything more elaborate. Tokens can also be confused within danbooru tags, like "chopper_(motorcycle)" understands the concept but is still going to see it as chopper from one piece and might sneak that in somewhere, same for something like "pear-shaped_figure", that's where negatives help. Get used to using things like from side/above/below, foreshortening/perspective, full body/cowboy shot/upper body/portrait etc, can do a lot to make a 1girl solo image more interesting to look at.
Controlnet is the most useful tool there is for txt2img, I use depth constantly to control the image composition, especially useful alongside regional prompting. Reference is really useful for concepts the model lacks grasp on. Always use "prompt is more important".
For inpainting get a feel for functional ratios of mask size, denoise, resolution, and only masked padding pixels, alongside adjusting the prompt right. I usually make a 1014x1024 aspect ratio equivalent then upscale by 1.4 and use mask padding between 40-120 depending on how much of the rest of the image the section needs to align with or see to understand the subject. Denoise at 0.4 can do more than you'd think and I rarely go as high as 0.6. Larger mask generally means less detail and difference also.