>>1073290
Here's an long post detailing my methodology. It's good to have a starter prompt ready to go with a handful of tags that you know improve the output then just add to it. I've made an effort to show what I'm using & how I used it.
Prompt: best anatomy, detailed eyes, large pupils, detailed pupils, thick eyelashes, shiny skin, detailed background, chiaroscuro lighting, incredible lighting, detailed shadows, vintage dining room, glossy wooden table, dark green wall, teapot, carrot on plate, steaming tea, 1girl, teen, skinny, medium breasts, pale skin, green eyes, very long hair, straight hair, swept bangs, long sidelocks, black hair, bunny girl, (anthro:0.6), freckles, freckles on shoulders, freckles on breasts, human nose, small nose, long black rabbit ears, strapless white dress, frilly dress, baggy clothing, bare shoulders, half-closed eyes, closed smile, soft blush, one eyebrow raised, looking at viewer, sitting at table, one hand lifting tea cup, drinking tea, elegant, beautiful, (traditional media, oil painting, classic media),
Negative prompt: IL_Neg, body paint, watermark, heart pupils, (small head), (2girls, group, multiple_girls), depth of field, out of focus, blurry, cuffs, (double ears, human ears:1.5), (hairband, white rabbit ears), tea on plate,
When making bunny girls I often struggle to remove human ears on this model. I had (anthro:0.6) in prompt to see if that helped & (double ears, human ears:1.5) in negative. I believe the nose is more red because anthro is in prompt although it's at a lower weight.
Steps: 35, Sampler: Euler a, Schedule type: Align Your Steps, CFG scale: 4, Seed: 1550528052, Size: 1024x1024, Model hash: 05369ba2dd, Model: Better_Days_Illustrious_XL_V01, VAE hash: 235745af8d, VAE: sdxl-vae-fp16-fix.safetensors, Denoising strength: 0.4, Clip skip: 2, Hires schedule type: Align Your Steps, Hires upscale: 2, Hires steps: 15, Hires upscaler: ANIME_4x_IllustrationJaNai_V1_135k,
I use hires upscaler at a denoise of 0.3 - 0.4 typically. With a simpler style it can potentially go higher, with more realism it generally has to go lower. I don't always use IL_Neg embedding in negative prompt, but it
can help anatomy issues.
Checkpoint:
https://civitai.com/models/623891?modelVersionId=1143094
VAE:
https://civitai.com/models/140686
Upscaler:
https://openmodeldb.info/models/4x-IllustrationJaNai-V1-ESRGAN
IL_Neg embedding:
https://civitai.com/models/993446/illustrious-embeddings
Embedding link 404s. Tried alternative, but civit.ai is 404ing that too.
I'll just upload the embedding I used:
https://d.uguu.se/jwJDBglB.safetensors
Note: Embedding didn't work for me unless I included --disable-safe-unpickle in launch settings.
Inpainting:
Masked content: original, Mask blur: 15, Inpaint area: Only masked, Soft inpainting enabled, 35 steps, CFG: 4, Seed: random.
<Face - 0.48 denoising strength
<Cup+Steam - 0.45 denoising strength
Avoided inpainting hands since it kept fucking them up, Sometimes it decides to be a bitch about that. Face & steam benefited from inpainting. Other areas didn't benefit much so I didn't bother. Didn't want to make this one look like it took a stupid amount of effort.
Photoshop:
>Filter - Blur - Surface Blur: Radius 3 Threshold 4
Loses detail, but unfortunately necessary. Inpainted areas have noticeably different grain that must be cleared first. Not needed in all cases. Could inpaint all the thick grain away or else prompt it away. But in this case, I have to wipe the grain in PS & reintroduce it to make it evenly grained.
>Duplicate layer
>Filter - Sharpen 40% - Sharpen filter on the duplicate layer, set layer to 40% Opacity, merge with original layer
>Action: "grain" 1 - slight at 40% Fill, merge with the layer below
This re-introduces grain evenly across the image. Dithers gradients, hides artifacts, film grain is useful.
<!! Kept this current layer for later. Continued on duplicate layer.
>Duplicate layer
>Action: Vibrant glow 15% Opacity
>Action: Sunglow 2 15% Fill
>Image - Adjustments - Color Balance: Shadows -3 -2 +3, Midtones +4 -1 +7, Highlights -4 +6 +3 (color tweaking)
>Image - Adjustments - Vibrance: -15 +5 (makes skintones paler)
>Curves: input output, 0 3, 82 77, 181 179, 255 250 (brightened areas that became too dark)
<!! All these changes are merged with the layer from earlier at 60% Fill
>Duplicated layer, named new layer FX (Action only applies to layer named FX)
>Action: CA Diagonal 1PX
>Action: Fade Edges (only affects layer named "Final effect", I fade edges of this layer to avoid lines on the edges)
>Action: Hallftone Overlay 1
I sometimes increase the scale of grain or halftone if they appear "too fine" for the resolution. Vibrant glow & Sunglow 2 are from "Kubota Artistic Pak I" found online. Other actions I either created or don't remember where I found 'em. Could've done more Photoshop, but the section is too long already. I'm using Photoshop CC 2019. Heard actions can be fickle about versions so my actions may not work for you. Photoshop Actions:
https://f.uguu.se/tvSHcswC.7z (the Styles are also necessary for some actions)
The simpler answer is use a model that easily generates good outputs with low effort & work from there. If you get something like the base NoobAI XL model you'll be tard-wrangling it for ages to get anything good. BetterDays by contrast "just werks" with a decent default style so you have a good starting point to work with. Artist tags, style tags & LoRAs can take you the rest of the way.