Sign In

The WonderMix

1.4k
13k
141
Updated: Mar 9, 2024
styleanimesexyfemalewomangirls
Download (2.14 GB)
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
1,602
Reviews
(149)
Published
Mar 9, 2024
Base Model
SD 1.5
Hash
AutoV2
2E5A2E46E0
default creator card background decoration
First Birthday Badge
Lorim's Avatar
Lorim
License:
CreativeML Open RAIL-MAddendum

From V13

For the models and LoRa used in the merge, i will create public collection and link to them in the version details.

Just tell me if it doesn't work, i will change back to the old link listing.

From V7

Details of the model in the Version Details

For prompt exemple, look at he images prompt 'for negative pormpt either use the one in the images or the classic (worst quality, low quality), lowres )

All the images were generated using fp32 version

V7

This Version is more on the anime side but, you'll still get some nice 2.5D output.

Also, there is a bit less details than the V6, but it was a choice, i wanted something more "clean".

You'll find the details fo the merge in the About this version part.

For the prompt and setting, look at the samples, you'll find them. I will just add that this time i used R-ESRGAN X4+ Anime6B as upscaler.

No HiRes Fix, Addetailers etc... were used.

For the Negative prompt, either use the one you'll find in the samples, or the classic (worst quality, low quality), lowres if you want the base style of the model (beware that you'll get some hands problems) or some TI/Negative Embeddings (the combo FastNegativeV2 + verybadimagenegative_v1.3 work quite well.... But you'll get stiffer facial expression)

Have fun :) (yup i did shorter this time.... The model details is way too long to add too much more ^^; )

V6

Big change on this one.

  1. Only 4 models were merged

  2. The base models have changed : one came back and the 3 other was never used before (See the ressource for more info)

  3. i used Super Merger with cosineA instead of the classical merger.

  4. A LoRa was merged into the model. this LoRa was a merge i did with several Style Lora each whith quite a low ratio which didn't altered much the original style but just enhanced it.

  5. No TI/negative Embeddings was used for the generation. In fact i didn't tested the model with them so, i don't know how it will behave.

  6. fp16 Pruned version only (The Full fp32 was way too big <<)

  7. Still Blessed2 as baked VAE

This version have some specifics.

  1. The closer you get the more detailled it become. To a point where it's almost too detailed ^^;

  2. Ofc the farther you get, less detailed it become.... Well not totally in fact. The body, clothes and background stay well detailed.... It's the face that got quite the loss of details and the hands have tendency to become messed up.

  3. Those loss of details start to appear from cowboy shot. but not always. it depend of the angle view you've chosen.

  4. Be aware that because of the use of 2 specifics models, if you don't specify the body feature of the girl, you have a good probability to get young looking one.

  5. the use of only two tags can make a simple prompt like "1girl, standing, short dress," into a sexy shot. Thoses tags are : sexually suggestive and nsfw. (he two together can create some freaking hot scene).

  6. This version is highly H Content compliable.

  7. It's quite flexible when it come to LoRa

  8. The model also work well with style tags (realistic, anime etc..) see the exemple post here : https://civitai.com/posts/532602 (First image being the base without style tags)

So, you shouldn't get too much hands, messed up body or extra limbs problem.

If you encounter too much hands problems, i suggest you to use EnvyBetterHands LoCon at 1. It doesn't alter the style much (a bit that's all) and work very well with it

I used some older prompt for the previews.... Because i was to lazy (and short on idea) to remake all of them XD.

Generation Settings Used

  • Sampling steps : 40. 40 this is where the model generate an image very close to the prompt. Lower settings (especially 20) have tendency to add things and to deviate, sometime a lot, from the prompt.

  • Sampler : DPM++ 2M karras (DPM++ SDE karass at 20 to 30 work well too)

  • CFG : 7

  • Clip Slip : 2

I didn't try Highres fix and Face restoration (or another add-on like Addetailer for exemple).

Prompt

Positive : (masterpiece, best quality, absurdres, high details, perfect)

Negative : (worst quality, low quality), normal quality, lowres, watermark, signature, greyscale, monochrome, jpeg artifacts, blurry, out of frame, aliasing, imperfect, ugly, gross proportions, disfigured, dehydrated, malformed, mutilated, mutation, mutated, misplaced, error, poorly drawn, what, body horror, anatomical nonsense, engineering nonsens, wrong hand, wrong feet, wrong fringers, wrong digits, extra arms, extra legs, extra hands, extra feet, deformed hands, defromed feet, deformed fingers, bad anatomy, bad proportions, bad perspective, bad drawing, bad lighting, bad details, bad shadow, bad outline, bad highlights, bad colors, bad hair, bad face, bad body, bad skin, bad muscles, bad joints, bad background, bad foreground, bad clothes, thick eyebrows, misplaced genitals, bad genitals, wrong hair color, wrong hairstyle, wrong hair length, wrong eyes color, wrong breasts size, wrong clothes, wrong clothes color, bad makeup, bad jewelry, bad accessories, bad ornament, bad object, bad vehicle, bad building, bad vegetation, wrong colors, extra clothes, imperfect hair, imperfect face, imperfect body, imperfect clothes, imperfect hands, imperfect fingers, fused fingers, fused limbs, fused clothes, malformed hands, malformed fingers, entangled fingers, bad reflection, bad refraction,

This is the base prompt i used for generating all the images.

Enjoy :)

V5 (Read before using)

The V5 is very different from the other version because it don't use the same base at all, and most of the model used for the merge are different (80% i would say)

The base for this version was : YabaL_Mix True2.5D

Second : no fp16 version. There is a significant and very visible loss of quality and details when converting to fp16. Sorry about that.

But the most importante part is that : the 15 first sample images of the showcase were generated WITHOUT TI/Negative Embeddings, LoRa, Hires Fix and inpainting (well i never used inpainting anyway :p).

So what you see is what you get.

But if you want to use TI/Negative Embeddings feel free (i will ^^), this version is quite flexible and the style isn't too much altered (but still is)..

Of course not using TI/Negative Embeddings as is downside. from cowboy shot you start loosing details (especially on the face). But i will give you a fix for this in the prompt section.

Tips for the V5

Same settigs as always

  • Sampling steps : 20

  • Sampler : DPM++ 2M karras

  • CFG : 7

  • Clip Slip : 2

Didn't test Hires fix or face restoration. So, see if you got good result with them ^^

Prompt

Here the important part.

Positive : (masterpiece, best quality, absurdres)

It's enough in most of the case, but if you need a little bit more details, i suggested you to use ultra detailed, intricate details, and any detailed you need (face, eyes etc.... All except hand and feet. Seriously don't). It's not perfect, it add token to the prompt, but if you want the out of the box style, it's the best way.

Negative : (worst quality, low quality), lowres, jpeg artifacts, aliasing, error, bad anatomy, bad proportions, bad perspective, bad hands, bad feet, wrong feet, wrong hand, fewer digits, extra digits, fused fingers, fused limb, extra arms, extra legs, missing limb, poorly drawn, poorly drawn feet, deformed, what, blurry, anatomical nonsense, body horror, bad arm, bad face, bad neck, bad reflection, bad vulva, misplaced genitals, watermark, text, username, signature, monochrome, greyscale, duplicate, poorly drawn joints, bad joints, poorly drawn genitals, bad image, poor lighting, ugly, out of frame

Yeah, i know it's a long one, but it work very well, not perfect but working.

Enjoy and have fun! :)

V4

This version is a very big improvement from the botched V3.

For the V3, i started from the V2 and tried to fine tune it.... But the result was disapointing.

For the V4 i came back to the beginning : the V1. And after a lot of trials and error, i finally got a very nice result.

I mean, badhandv4 work quite well on this version. It's the most stable version of my model.

Tips for the V4

  • Sampling steps : 20

  • Sampler : DPM++ 2M karras

  • CFG : 7

  • Clip Slip : 2

Highres fix doesn't add much.

Tag like anime, realistic, illustration, pastel colors etc.... work, not by a lot, but there is still some change.

At the start of the prompt i advise for the following : (masterpiece), best quality, highres, ultra-detailed.

For Negative prompt : badhandv4, EasyNegative, verybadimagenegative_v1.3, monochrome, jpeg artifacts, text, watermark, signature, (worst quality), (low quality), (normal quality), lowres, greyscale, multiple views, comic, silent comic, extra digits, fewer digits,

This time, i really hope you will enjoy :)

V2 vs V3

So, after further testing on V3 here the conclusion ;

  • Use the V3 if you want to generate portrait, close up or focused work. there will be more details and it will be more accurate.

  • use V3 for male character and when generating more than two characters.

  • use V3 for background generation.

  • From cowboy shot to further use V2.

The V2 is more balanced for close to afar generation, but have a lot of hands drawing problems.

The V3 is more close and focused work oriented, but have a lot less hands drawing problems.

About the V3

What the V3 add?

Well, body are more detailled, a bit more NSFW content, a bit more cosplay content.... well a bit more of everything. and less bad drawing (fused limbs extre limbs etc... are very rare now).

you can generate output with several characters, but i advise you to not go above 4, after it's becomming horrible <.<.

up tu 3 characters for portrait,. Above you need to generate in landscape.

You can also generate men.

Tags like anime, realistic, pastels colors or illustration have very little impact on characters, but work very well on background. See the tips below.

The only things i can't do anything about are the hands. i tried several fix and embeddings but still have case of missing fingers or messed up hands.

So here a tips : the more token you use for generation, more are the risks of getting bad hands.... So when you are above 150 token, try to hide them.

Baked VAe is still Blessed2, but you can use other for getting different output. (here the difference between 6 VAE)

Tips for using the V3 :

  1. Don't use a lot of Negative Embeddings. EasyNegativeV2 (link) is enough. You can use more if you want, but you are prone to greatly change the model and to lessen the quality of the output.

  2. the following negative prompt work well for fixing the lack of detail on face from afar : EasyNegativeV2, monochrome, jpeg artifacts, text, watermark, signature, (worst quality), (low quality), (normal quality), lowres, greyscale, multiple views, comic, silent comic, blurry, (poorly draw hands, poorly draw fingers, extra digits, fewer digits:1.5), extra limbs, bad proportions, poorly draw face, poorly draw eyes, poorly draw teeth,

  3. And it even greatly lessen the hands problems.

  4. add the following at the beginning of your prompt : (masterpiece, best quality, highres, ultra-detailed, detailed background, detailed eyes, beautifull face:1.2). this will upgrade the overall quality.

  5. the drawback of this methord is : the model will generate more anime-ish drawings.

  6. The Photorealistic tag make a more 2.5D render, not by much, but the difference is still here. I didn't try other similar tags, but i assume they will add some changes too.

  7. here the Prompt Exemple Post ; https://civitai.com/posts/331426

Here come the V2!

So what's the difference?

  • A baked VAE : Blessed2 VAE

  • More Stable output (less variation within a seed, less odd number of limbs, less fused limbs, less badly draw hands)

  • But more importanly, the way of prompting is very different from the v1 ; No more anime, realistic or photorealistic tags. the output will look almost the same and only the image composition will sightly change.

  • No need to use highres fix anymore, except, if you want more variation in a seed.

  • Now you get really good result with a low number of sampling steps! (20 is really enough with DPM++ 2M karas)

  • well there is a con : nsfw tag needed for activate mature content.

Now a special note on the VAE : Blessed2 VAE produce the best output for me, now some other VAE produce also great things.

If you want a more anime-ish output, kl-f8-anime2 and vae-ft-mse-840000-ema-pruned do the job well.

pastel-waifu-diffusion.vae is quite nice too if you want more pastel colors themed output.

I hope you will enjoye this V2 :)

Edit : as requested, i added a fp16 version ^^

i present you : The WonderMIX!

Why this name? because it's a wonder howi got this result by only merging 4 models on a whim.

About this model.

This model is really good to make semi-real character, you could even say 2.5D.

While having body very close to realistic drawing and very detailed style, they keep the roundish form of the anime style face.

Also the body proportion are very good : small breasts are small breasts, medium breasts are medium breasts and large breasts are large breasts.

So if you want to produce a small tsundere petanko or a tall, bulky warrior with huge breasts, you can!

it can also produce some really good backgrounds.

For illustrate those difference i generated 3 series :

  1. One with a JK (High School Girl) using anime tag and kl-f8-anime2 VAE

  2. One with an OL(Office Lady) using realistic tag and orangeMIX VAE

  3. One with a Cyberpunk Bounty Hunter using photorealistic tag and kl-f8-anime2 VAE

All the images were generated without LoRa.

Settings used for generating images

  • Sampler : DPM++ 2M Karras

  • Sampling steps : 40

  • CFG : 7.5

  • Resolution : 512x728

  • Hires Fix :

    • No Upscale (i have an old 4Go vram GPU)

    • Hires Steps : same as Sampling Steps

    • Denoising strength : 0.3

  • Upscale

    • Resize : x4

    • Upscaler : ESRGAN_4x

Prompt tips for this model.

in all prompt i used the following : (masterpiece, highres, best quality, ultra detailed, detailed background:1.2) with either anime, realistic or photorealistic before masterpiece.

For the negative well..... the classic i think :

About NSFW content.

It's capable, no top notch, but with a little patience you can get naughty content. It will mostly be vanilla content.

Now, i would advise you to use LoRa for more precise or special content.

You will find some samples in the gallery below (No LoRa involved).

Conclusion.

Hope you will like this model and have fun with it :)

Credits.

Thanks for the creator (train and merge) of the 4 models used for this merge