My models
Support
Place bid in the Auctions section for availability in the generator
Pony model
At this point I feel like there's not much reason to keep using the SD 1.5 models since the pony models have better understanding of pretty much everything. And since I can only have 2 models for on site generation I'll see if this one gets added instead of the SD1.5 v1.3.
Prompt
score_9, score_8_up, score_7_up, (anywhere in your prompt)
Negative Prompt
score_4, score_3, score_2, score_1, (or empty)
Sizes / Resolution
A baseline of 1024 x 1024 should give good results
896* 1152 for ~3:4 aspect ratio
Everything bellow is for SD1.5 verions
Lore / Backstory
I just wanted a model that could generate high quality art and hentai and that actually followed the prompt. Couldn't find one so I made it myself. Hope you like it and share your generations with everyone!
I hope this model doesn't become just another pointless merge that no one will ever use.
Using the model
The best advice I can give you is to experiment for yourself and see what works for you. So you are encouraged to completely ignore this and use the prompts/settings you like and find ways to get the best results out of this and any model.
Like Danbooru's Rule of thumb "Tag what you see, not what you know", don't use words that should not be visible in the final image.
IE: Don't use words for shoes if you are making a portrait or upper body image
Prompt
Use danbooru tags or sentence-based prompts.
You can use any generic filler prompts like: masterpiece, best quality, high quality, highres, etc..., or none.
Extension for tag autocompletion tagcomplete
Prompt Order for good composition
Tags at the start have more influence than the ones at the bottom.
The 7 blocks template (Work in progress):
participants, sex acts, view angle
posture, positions and secondary actions
secondary actions,
head orientation (looking back, looking at viewer, etc...),
body
breasts size, stature (petite, tall female, etc...),
age (mature woman, old woman, etc...),
body fat (skinny, curvy, plump, fat, obese, narrow waist, etc...),
hair
clothing or lack of clothes
face expressions, cum location and eye color
scenery and others
furniture and objects that should be visible,
styles
character (probably only if it was released / popular before early 2023)
(hatsune miku, ganyu \(genshin impact\), etc...)
You don't need to fill in every category, it's just for reference.
Negative prompt
You can just use EasyNegative or some of the generic prompts:
worst quality, low quality, bad quality, normal quality, bad anatomy, monochrome,
And whatever you don’t want to be on the image
Some people use the standard (worst quality, low quality: 1.4) + negative embeddings with good results too.
If you use too much emphasis, like (worst quality:2.0) you'll start getting burned images.
Samplers
You can use any sampler. Some samplers converge at a defined image, others keep changing with each step. You can find better explanations here and here.
I'd say the top choices are:
DPM++ SDE Karras (slow)
DPM++ 2M Karras (fast)
Euler a (fast)
DPM adaptive (slow)
DPM++ 2S a Karras (slow)
DDIM (fast)
Steps
25 - 50
CFG
7: standard
3–15: working range
Lower cfgs don’t follow the prompt as much. Higher cfgs will make “burned” images
The extension sd-dynamic-thresholding can play around with different CFGs at different steps
Sizes
1:1 [512, 512], [768, 768]
2:3 [512, 768]
3:2 [768, 512]
You can use Hires fix to get a bigger picture with more details
Upscaling
I'm getting good results with Hires fix enabled.
Upscaler: Latent nearest-exact
Hires steps: 20 - 50
Denoising Strength: 0.5 - 0.55
Upscale by: 1.5 - 1.875
Denoising at 0.5 seems the most stable for preventing duplicate body parts and body horrors
Latent upscalers tend to create more detail while upscaling, the rest of the upscalers don't add too much to the base image with the same settings.
VAE
There are many VAEs out there. These are the 3 most popular.
kl-f8-anime2 for vibrant colors and anime face/eyes
orangemix.vae for muted colors and anime face/eyes
vae-ft-mse-840000-ema-pruned for vibrant colors and real face/eyes
Baked VAE for each version.
v1.0 has vae-ft-mse-840000-ema-pruned.VAE baked in.
v1.1 has vae-ft-mse-840000-ema-pruned.VAE baked in.
v1.2 has vae-ft-mse-840000-ema-pruned.VAE baked in.
v1.3 has kl-f8-anime2 VAE baked in.
The [no VAE] variants don't have any VAE baked in.
Permissions
Just use the model and share your generations here or wherever you want
I’d appreciate if you link to this model page when sharing your generations 😊
I made a patreon if you want to support me https://www.patreon.com/TianChimp
Please share your generations!
Tag considerations
I've noticed some tags having too much influence over the entire image resulting in a loss of style and quality, but sometimes they are necessary for your target composition. This is probably only noticeable on <75 token prompts.
You can try setting lower influence with parentheses IE 0.5 - 0.9 Needs more testing but I'll keep a list of potential suspects.
fellatio (tends to put hand around the penis)
cum (variations)
multiple boys
FAQ - Frequently asked questions
My images look desaturated, with dim colors
Your software is probably missing/not using a VAE. The main model v1.1 doesn't include one, so you can use any VAE you want. The model v1.1 [VAE] comes with the standard VAE baked in, so it uses that VAE and ignores the selected one.
Different interfaces have different ways of selecting the VAE. In Automatic1111 you can set a VAE selector next to the model selector by going to Settings > User Interface > Quicksettings list, and adding "sd_vae".(Apply Settings and Restart UI)
That'll result in your interface looking like this:
Your vae selector will look for files located in your VAE folder, like "SD-1.6\models\VAE", so you have to download the VAE files you want and place them there.
Description
VAE free 1.3 version
FAQ
Comments (17)
I can't get it to generate anything that isn't kind of horrifying. It really likes putting multiple people into the image and then merging them together, and when it isn't doing that it's turning limbs backwards or sticking feet where they don't belong.
SD1.5 model (or any stable diffusion model) cannot generate image much larger than the training set. Which means image larger than 512 by 512 tend to be horrifying because the model do not know what should be filled in such space. Some suggestions for your complaint:
1. Try to generate smaller images, 768 by 512 or 512 by 768, as suggested by poster. If you really want to generate bigger image, try to upscale or use SDXL models.
2. Use better (and slower) samplers with high steps count may help. Try DPM++ SDE Karras with 50 steps first, then adjust for faster generation.
3. Use more negative prompts, you can copy the longest one in the gallery. It really works sometimes.
Please use my negative prompt. It work with all model.
((3d, cartoon, anime, sketches, closeup, close up, close-up, too close, bad eye, tanning, tanlines, muscular, bangs, from back, milf, old, back view, backside)), (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome, black & white, b&w, black and white)), ((grayscale)), bad anatomy, out of view, cut off, ugly, deformed, mutated, EasyNegative, paintings, sketches, lowres, normal quality, skin spots, acnes, skin blemishes, age spot, glans,extra fingers,fewer fingers, (ugly eyes, deformed iris, deformed pupils, fused lips and teeth:1.2), (un-detailed skin, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.2), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, 2girls, 2women, woman with penis
@pattigdesign bro thats just way too much negativity :DD hah (my negative prompts are just 1-2 keywords and then neg embeding for example: piercing, [[[embedding:bad_prompt_version2-neg]]]
Even SFW follow the prompts well :D That said, 1.3 doesn't follow the prompts as well as 1.2.
I am looking forward to the SDXL. We think it is the best model and we would appreciate it.
It is very sensitive to prompt.
sex moment
It's always mutated when i use hires. fix.
Would enjoy sdxl or sdxl lightning models. Great work Tian!
does any chance for PONY? QAQ
My friend, this model is great, thank you! It follows prompt better than any other checkpoint I have encountered, and is generally very high quality. However, I have a suggestion for 1.4 version. Could you make default girls thighs a little less thick? This model tends to generate them significantly bigger than usual, and forces me to use some loras for correcting this.
This is the model I have ever encountered with the highest success rate in generating multiplayer images, and it's so good
I was commenting on pornography, not diversity, damn Baidu translation software
This is the best anime model (with lora add_detail ). Better than any Pony-based model, and even better than Flux. It would be great if it had higher resolution and clarity.
Hello guys, what i have to do when after insalling and adding model to InvokeAI, i have often accuring error "expection". and im not sure what to do :c