I'm currently learning to use Stable Diffusion using three "easy" distros: ComfyUI, Easy Diffusion and NMKD.
Each has its pro and cons, and each require to somehow tweak the textual prompt. I'm using Civitai's huge database of datasets and images as training ground. However, most of the time I'm not able to recreate the images posted there. I noticed that when I copy/paste the textual prompts used in there like this one:
Ultra-realistic 8k CG, masterpiece selfie,best quality,((very detailed face)),(photorealistic:1.4),absurdres,RAW photo,(film grain:1.2), professional portrait, shy, embarrassed, blushing red cheeks, beautiful Czech with Latina origins, sitting in bus seat, background parking lot, age 18, sweaty body, contrasting light, alluring look, white sheer button down shirt, jean shorts, shirtlift, revealing a cup breasts <lora:shirtliftv1:0.8>, flat chest, black bob hair, having orgasm, ((eyes closed)), orgasming, cumming, arched back, <lora:hand-down-clothes:0.9> hand down clothes, (hand down shorts:1.3)
Negative prompt: bra, duplicate people, twins, (EasyNegative:1.2),(worst quality:2),(low quality:2),(normal quality:1.6),(ng_deepnegative_v1_75t),deformed nipples,monochrome,(watermark:1.4),artist name,(character name:1.2),text,bad anatomy,sketch,duplicate,(negative_hand-neg:1.2),bad_pictures,FastNegativeV2, bad hands, asian, Asian-Less-Neg, bad eyes
Steps: 20, Size: 512x768, Seed: 4231704986, Model: Degenerate_deliberateV1, Version: v1.6.0, Sampler: DPM++ 2M Karras, "breasts: 9466f683b3da, CFG scale: 7, Model hash: 4eb333dceb, "shirtliftv1: 5ce36857badc, easynegative: c74b4e810b03, Asian-Less-Neg: 22d2f003e76f", hand-down-clothes: e7856254ccb4"
The last part is what is giving me more problems:
How am I supposed to use those extra information like: "breasts: 9466f683b3da, or Model hash: 4eb333dceb, "shirtliftv1: 5ce36857badc, easynegative: c74b4e810b03, Asian-Less-Neg: 22d2f003e76f", hand-down-clothes: e7856254ccb4" ?
Does it force the algorithm to pickup a specific set of data to make the subject repeatable? Because I usually obtain results that are quite distant from the pictures posted on Civitai, even if I download the exact same data (checkpoints, loras, etc...). I know that different UIs also introduce some variations in how they pass the data influencing the end result, but most of the time the results I obtain are very different!
Each has its pro and cons, and each require to somehow tweak the textual prompt. I'm using Civitai's huge database of datasets and images as training ground. However, most of the time I'm not able to recreate the images posted there. I noticed that when I copy/paste the textual prompts used in there like this one:
Ultra-realistic 8k CG, masterpiece selfie,best quality,((very detailed face)),(photorealistic:1.4),absurdres,RAW photo,(film grain:1.2), professional portrait, shy, embarrassed, blushing red cheeks, beautiful Czech with Latina origins, sitting in bus seat, background parking lot, age 18, sweaty body, contrasting light, alluring look, white sheer button down shirt, jean shorts, shirtlift, revealing a cup breasts <lora:shirtliftv1:0.8>, flat chest, black bob hair, having orgasm, ((eyes closed)), orgasming, cumming, arched back, <lora:hand-down-clothes:0.9> hand down clothes, (hand down shorts:1.3)
Negative prompt: bra, duplicate people, twins, (EasyNegative:1.2),(worst quality:2),(low quality:2),(normal quality:1.6),(ng_deepnegative_v1_75t),deformed nipples,monochrome,(watermark:1.4),artist name,(character name:1.2),text,bad anatomy,sketch,duplicate,(negative_hand-neg:1.2),bad_pictures,FastNegativeV2, bad hands, asian, Asian-Less-Neg, bad eyes
Steps: 20, Size: 512x768, Seed: 4231704986, Model: Degenerate_deliberateV1, Version: v1.6.0, Sampler: DPM++ 2M Karras, "breasts: 9466f683b3da, CFG scale: 7, Model hash: 4eb333dceb, "shirtliftv1: 5ce36857badc, easynegative: c74b4e810b03, Asian-Less-Neg: 22d2f003e76f", hand-down-clothes: e7856254ccb4"
The last part is what is giving me more problems:
How am I supposed to use those extra information like: "breasts: 9466f683b3da, or Model hash: 4eb333dceb, "shirtliftv1: 5ce36857badc, easynegative: c74b4e810b03, Asian-Less-Neg: 22d2f003e76f", hand-down-clothes: e7856254ccb4" ?
Does it force the algorithm to pickup a specific set of data to make the subject repeatable? Because I usually obtain results that are quite distant from the pictures posted on Civitai, even if I download the exact same data (checkpoints, loras, etc...). I know that different UIs also introduce some variations in how they pass the data influencing the end result, but most of the time the results I obtain are very different!