wgty

wgty, 10 months ago

I had fun trying to do this “in camera” without any editing after generation. I ended up using segmentation but after playing around with a lot of different options for what colors to use (trying stone, book, window, painting), I gave up and just made it black and white and still used segmentation.

https://lemmy.world/pictrs/image/b0d9b641-9eda-4ab8-93b4-33f3b056afef.png

https://lemmy.world/pictrs/image/664ed9df-54cd-4b7b-8daa-30a05d2e969f.png

https://lemmy.world/pictrs/image/b2e09966-23c9-4639-a159-efecacd0569e.png

wgty, 10 months ago

Segmentation controlnet with an image of the cover template with different parts being different colors as the input. The segmentation controlnet makes Stablediffusion treats different colors as different things, so it keeps each part distinct. If you use colors that aren’t in the actual color reference chart, it just sorta improvises, so I used black and white. On one I used the color that corresponds to painting,picture for the area inside the “frame” where the actual art is. Since I was doing it all in one generation with one prompt I just sorta had to accept what it gave me for the colors and textures of the text and borders.

https://lemmy.world/pictrs/image/bc7c7d96-34fc-4f99-af6c-386f49b21cf3.png

wgty, 10 months ago

It definitely depends on the model. But if you want better results you’re going to want to use adetailer.

I threw your prompt into dreamshaper 6.31 (because it’s the latest version I had on my server). Here’s with no adetailer: https://lemmy.world/pictrs/image/8ffc3648-f11a-4091-8d21-12abec72a7d1.png

Here’s with adetailer but using the exact same prompt as the main image (leaving both fields blank but turning adetailer on):

https://lemmy.world/pictrs/image/ace24244-d2a1-4b0e-a0a0-80408795c88b.png

Here’s with adetailer but putting only the expression part in the positive prompt. Negative prompt still the same as the main image.
(confused concerned upset questioning (funny fun amused (cringe (raised eyebrow) shy blush ashamed bruh face), <lora:add_detail:1.2>

https://lemmy.world/pictrs/image/b5f2be4d-73bd-4bb6-9898-5fc626d6c051.png

Now, probably a good idea to have redhead woman in there or something, but my point is how rapidly you can improve things with very minimal effort. If you’re willing to mess around with adetailer prompts and denoising strength, you can do a lot with faces.

wgty, 10 months ago

Anachronistic photographs.

wgty, 10 months ago

Are you using any hand-specific negative embeddings? adetailer also can do a decent job of fixing hands with the hand segmentation.

wgty, 10 months ago

Reminds me of the bastion cinematic. To the point where I’m guessing either a still from that, or fanart of that cinematic was in the training set.

wgty, 10 months ago (edited 10 months ago)

Couple suggestions:

1. Try using A-Zovya RPG Artist Tools
2. Use the adetailer extension to improve faces. Just turning it on for faces will have a huge improvement. Make sure to turn off the normal “restore faces”.
3. When you’re happy with a final image, re-generate it with upscaling.
4. Try using negative embeddings like fastnegative, deepnegative, BadDream, EasyNegative, and maybe negativehand. Don’t have to use all of them, and some work better for some checkpoints and styles than others. Usually I use 2 or 3 together.
5. Use quality tags in the positive and negative prompt.

EDIT:
My quick attempt at making something with my own advice: i.imgur.com/xjhc2tA.png

(masterpiece, best quality, high quality, highres, ultra-detailed), elf rogue, stealth, fantasy, (dagger (weapon):1), black cloak, ((masculine)), (long black hair), feather collar Negative prompt: (ng_deepnegative_v1_75t, FastNegativeV2, verybadimagenegative_v1.1-6400), (worst quality, low quality, normal quality), (woman, female, feminine), (archer, bow) Steps: 35, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 607032892, Size: 512x768, Model hash: 8c4042921a, Model: aZovyaRPGArtistTools_v3VAE, Denoising strength: 0.35, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate/erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 23.6.2, Hires upscale: 2, Hires steps: 14, Hires upscaler: R-ESRGAN 4x+, Version: v1.3.2

The vampire vibe in your original image could definitely be brought back, but I assume it was more of an unintended side-effect of your prompt than intentional.

wgty, 10 months ago

I just noticed it doesn’t put the literal prompt in the png info. The “dagger (weapon)” needs to be entered as “dagger (weapon)” so that the parenthesis around weapon get sent to CLIP

wgty, 10 months ago (edited 10 months ago)

These are very large (3000x2000), so here’s an album link: imgur.com/a/LOZcR39

Love at first sight.
https://i.imgur.com/4aveMX2.png

Obligatory gas masks and fire.
https://i.imgur.com/rjgIcTR.jpg

What the fuck does anything on this vehicle do?
https://i.imgur.com/4NM7rSA.jpg

wgty, 10 months ago

Extremely dank. What checkpoint?

wgty, 10 months ago

(masterpiece, best quality), landscape, impressionism, river flowing through wildflowers, swirling marbled clouds, multicolored ink, mossy stone bridge, lora:add_detail:1Negative prompt: (worst quality, low quality), house, building, castle, BadDream, railing, power lines, utility pole, symmetry Steps: 35, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1127511940, Size: 768x512, Model hash: 86fca15160, Model: toonyou_beta5Unstable, Denoising strength: 0.35, Clip skip: 2, Hires upscale: 1.5, Hires steps: 14, Hires upscaler: R-ESRGAN 4x+, Lora hashes: “add_detail: 7c6bad76eb54”, Version: v1.3.2

wgty, 10 months ago

Here, have some 4x clipdrop upscaling:
i.ibb.co/…/00224-1127511940-clipdrop-enhance.png