AnimeDen版本V3 (ID: 4331) 综合资源合集综合资源合集

This is a mix of realistic and anime models to get better characters and scenes. Individually the models had good results but was hoping to get anime like results and them be more lifelike. I typically use it with the AnythingV3 vae, but it can work without too.

The merges are with safetensors float16:
EDIT: On further digging it looks like I will need to remake yet another version without elldreth, maybe then I can get a commercial version. I keep finding some of the ones I had use dreamlike or seekart, so need to remove those and merge again... -_- hopefully V4 will be commercial.

V3: sadly Not commercial due to any Elldreth models -- 32 model merge
V2: Not commercial due to https://civitai.com/models/1956/kaelvinsmix -- 4 model merge
V1: Should be Commercial -- 4 model merge
The process for the merges is listed in their respective info tabs.

Example image prompts have the same settings and negative prompt. The only thing that changes is the keyword positive prompt. The [prompt]:number is what I use to help control how strong it is. I usually use about 0.5 because going down to 0.25 is a bit too washed out, and 0.75 is sometimes too literal. For the negatives though I have it at 150.

I used an xyplot to generate all of the images. Here is the info from the first picture that can be used to fill in the base settings:

[purple leather, proper proportions, copy paste, vibrant colors, priority order, desirable assets, rtx textures, beautiful work, superb craftsmanship, masterfully aligned, extremely intricate, detailed features, defined edges, intense focus, sdr hdri, pic vid, hyper realistic, high quality, max resolution, epic composition, cinematic lighting, full shot]:0.5
Negative prompt: [loli child, poorly drawn, manga panel, manga page, comic panel, comic page, abiotic anatomy, abiotic hands, jpeg artifacts, boxart logo, watermark overlay, flat coloring, digital painting, dutch angle, random lines, bad proportions, thumbnail icon, data corruption, corrupt data, no detail, limited color, standard definition, warped perspective, copyright text, title font, author signature, split area, cut off, undefined shapes, blurry shapes, hitchcock zoom, data error, texture error, missing texture, missing depth, low quality, no color]:150.0
Steps: 64, Sampler: Euler a, CFG scale: 16, Seed: 20343595335598760, Size: 768x512, Model hash: de2f2560, Model: AnimeDenV3, ENSD: 9876543210

The x is prompt S/R:

purple leather,female naked model posing exposed at viewer,male naked model posing exposed at viewer,macro closeups biotic wedding photo of humans interlocking holding hands,cthulhu godzilla monster,crowd running from a real life mech suit mecha in japan streets,photo of a white raven perched on rock,ancient scroll diagram character sheet,book cover horror lush island sequoia jungle edge mountain in the distance,ix nz az,side boob big breasts blue hair topless nude fox girl fox ears fox tail kitsune blue fur blue tail 1girl,anthropomorphic vampire succubus,buff toned male elf hero champion holding sword,muscular black female amazon tribe warriors,3d figurine cute display anime adult,3d disney pixar cute animal companion puppy fun movie,cel shading vector illustration 2d flat colors anime action op ed,pyroclastic unreal engine render,animal planet national geographic sea turtle lit underwater caustics octane

The y is checkpointname:

AnimeDenV3,AnimeDenV2,AnimeDenV1

描述:

V3: sadly Not commercial due to any Elldreth models -- 32 model merge

This was a merge of 32 different models. I used the merge board addon and used various recipes to do it. The base pattern of how I merged them was thus. This was the first recipe I had: (Note including all of the originals, and every step of the recipes will take 577 GB, if you were to try and do it yourself)

if you want to see the full recipe list:

https://pastebin.com/n1EXG0rR

for simpler viewing this is one of the pairs:
A-ws05 = C + D, 0.5, fp16, safetensors

B-ad05 = C + D + __O1__, 0.5, fp16, safetensors

C-ad05 = D + C + __O1__, 0.5, fp16, safetensors

D-ws05 = __O2__ + __O3__, 0.5, fp16, safetensors

it will get a half weighted sum

then subtract that from an add difference

then swap and do another add difference

then do a half weighted sum of those differenced models

Using each pair with those steps from 32 to 16, 8, 4, 2, and finally 1

I might have to try a different arrangement to get better results but this is what the order was for V3