V2 UPDATE - WIP - Need help with testing & making things better

-

Hey,

Finaly some update. I went through all of Blame! and kept 170 HD scans that i further cleaned with Ps to make this v2. Phew.

I consider this V2 a WIP, because right now i'm not able to spend much more time on this and SD in general.

It can create some satisfying results, but i know there are some errors and that it's possible to make it better.

That's where i require your help. I'm sharing all my dataset, pics, captions and training infos. If you want to go through this, have time to test, write down your results, make better captions, maybe re-train, alone or with others, feel free to comment and share ideas.

✌️peace nihei bros✌️

Here are some of the main keywords :

killy, dhomochevsky, cibo (normal), cibo (level9), iko (boy), iko (ghost), pcell

sanakan (safeguard), sanakan (normal), GBE, GBE gun, big GBE gun

aiming gun at viewer, firing gun

in color --i added the few colored panels in the dataset--

debris, dynamic, battle scene, fight, action, group, scars

human, fisher, boy, pilot, fisher helmet, fisher armor

safeguard, safeguard (puppet), safeguard (high level), on four legs, standing, safeguard sign on forehead

admin, admin sign on forehead, hood

bioelec, bioelec guard, bioelec carrier, bioelec police

small builder, giant builder

silicon life, full mask, half mask, hair, no hair, weapon

spaceship, spaceship interior, cockpit

scenery, vast scenery, city, tubes, cables, smoke, clouds, trees, nature, decrepit, strong perspective, horizon, deep horizon, clear background, black sky, empty, stairs, bridge, from above, clear sky

heavy contrasts, monochrome, no gradient, --in negative prompt:gradient

-

-

- Thanks a lot to ViHtor, Blame Connoisseur and Captionning Teacher who made this better blame lora possible

-

-

Tsutomu Nihei - Blame! LORA v1

Trained on kohya-LoRA-dreambooth with Anythingv3-1 model as base, using 40 pics from Blame! Vol.1.

Some words that works well :

vast, gigantic, decrepit scenery, covered in large tubes and cables, buried under cables, dim light, heavy black and white, giant-gross-organic, monsters-creatures, cyborg girl portrait etc. …

I can add that organic works very well with buildings and city scenery.

Negative i used:

easynegative, lowres, ((bad hands)), text, missing finger, extra digits, fewer digits, ((mutated hands and fingers)), (poorly drawn face), ((deformed face)), extra face, (double head), (extra head), ((extra feet)), logo, (cropped), worst quality, low quality, normal quality, jpeg, humpbacked, ((jpeg artifacts))

I think i had anime.vae.pt for these screenshots. I also used epiNoiseoffset_v2 Lora too, weight 1.3.

"60% of the time, it works everytime". This was done very fast and could easily be done way better, but results are satisfying so i'm posting it here. Will maybe update one day with more training on faces.

Enjoy and please share your results !