ONE FOR ALL «Furry»版本v. 3 DPO+VAE (ID: 343698)

If, you like the ONE FOR ALL series, please, leave a ?? and post your masterpiece. Follow me on Instagram. This model is available for generations on Yodayo.

BTW, take a look on this article: Five Nights at Freddy's | Civitai

The concept behind this model is to allow YOU to have the widest range of features with a single model. Today it's not 100% possible to have a "Swiss army knife" when it comes to SD 1.5, however with the well-made and structured merge of well-trained models, it is possible to offer a broad spectrum of features within a category.

ONE FOR ALL models should work with both simple and complex prompts. Moreover, it interacts well with all types of LoRAs, extending to a wide and varied range of possibilities. My intention is to offer a ONE FOR ALL solution, so that you can save time and disk space.

☕ I LOVE COFEE!! Want to buy me a cup?? ☕

You can choose between 4 models in this «FURRY» series:

And you can make these styles:

This model is specialized in:

Furry; Fursona; Anthro; Realistic; 2D & 2.5D.

It's a simple, all-purpose furry model with a great pedigree.

But, it's a ONE FOR ALL model, so, YOU CAN MORE!!

ONE PROMPT:

one woman, evening, purple n white dress, river, purple lily, smiling

ALL POSSIBILITIES

(2.5D) || (anime) || (realistic) || (furry art) || (cartoon) || (3D) || (chibi) , "insert prompt here"

⟬ despite the diversity with which the model was created, keep in mind the specialization and limitations of the model and that the results may vary according to the prompt, img size, sampling etc. ⟭

If you use a prompt that is not specific to Furry/Anthro, I recommend putting it in the first line:

(Ultra-HD-details), (Furry Art : Anthro : "specie")

Replace "specie" with the species you want, for example:

(Ultra-HD-details), (Furry Art : Anthro : Fox)

«NOTE ON THE DPO VERSION (v. 3 and Final)»

This version 3 is the result of a lot of training and testing. I take pride in saying that this is the definitive version in the SD 1.5 model. It's worth mentioning that despite using a proprietary training model as a base, this version incorporates other trained models. Therefore, it classifies as a Merge rather than a Fully Trained model.

In this version, a revolutionary training system called DPO - Diffusion-DPO has been included. It is adapted from the recently developed Direct Preference Optimization (DPO), a simpler alternative to RLHF that directly optimizes a policy based on human preferences under a classification objective.

What does it do? The trained DPO models have been observed to produce higher quality images than their untuned counterparts, with a significant emphasis on the model's adherence to your prompt. This model can bring a better prompt adherence than other fine-tuned Stable Diffusion models.

Watch your Prompt!

Due to the use of DPO, version 3 is much more prompt-sensitive, since version 1 it was already possible to use linguistic prompts, however version 3 takes this to another level thanks to the integration with DPO. Therefore, old prompts can result in completely different generations.

PLEASE NOTE that the DPO system was NOT designed with the FURRY/ANTHRO style in mind. One of the reasons why this checkpoint gave me so much trouble was to try to find a balance and train the DPO for the necessary style, please be careful when using conflicting words such as (Anime and Photo-realism), avoid prompts such as:

masterpiece, perfect something, absures, etc.

Watch your Clip Skip!

By default I recommend using Clip Skip 2, however, depending on your prompt and what you want to generate, it may be better to use Clip Skip 1.

Generally speaking:

Clip Skip: 02
- It's suitable for all generations, but tends to offer an aesthetic closer to anime, regardless of whether it's realistic or not;
Clip Skip: 01
- It tends to produce more anthro-realistic images, detaching itself a little from the anime aesthetic.

Recommendations (v. 1 ~ 3):

As a rule, it is not necessary to use Hires Fix or ADetailer, but:
- Hires Fix may be necessary in long distance or crowded scenes;
- ADetailer may be required depending on the prompt or result obtained.
For negative, I strongly recommend:
- Quality Embeddings For Furries.
VAE: Vae-ft-mse-840000-ema-pruned
The «Furry» model has been trained to generate furry without the need for a specific trigger, however, in some cases the prompt doesn't help. If you are encountering difficulties, try using: PK FurryMix - PK-FurryMix-128 v1
LoRA (although I avoid to use LoRA in the model showcase images, I often use these fantastic LoRAS, maybe you'll like them):

Generation Recommendations/Tips:

Sampling method:
- This version works with any sampling method, my preferred ones are:
  - DPM++ 3M SDE Exponential;
    - DPM++ 2M SDE Karras.
Sampling steps:
- 30+ (it depends on the size and sampling method, for DPM++ 3M SDE Exponential, I recommend 70 steps, for DPM++ 2M SDE Karras it is 30+, for Euler it is 30 ~ 40 and for LCM 30).
Width/Height:
- Normally I use: 1280 x 800 or 800 x 1280, (BE CAREFUL WITH YOUR LoRA, some LoRAs doesn't work with resolutions higher than 512), but, I've created the images on this post using 1600 x 400 without upscaling or Hires. Fix.
CFG Scale:
- I use 8 as default.
Clip skip:
- 1 or 2
  - (someone once used a skip 9 and made a bad review for my model, why?? ? Use 1 if you think 2 isn't giving you a realistic touch, but 2 is the default recommendation in any case and should work in all scenarios.)

WebUi-user-bat parammeters (You probably DON'T NEED IT):

@echo off

set PYTHON=

set GIT=

set VENV_DIR=

set COMMANDLINE_ARGS=--autolaunch --no-half-vae --theme dark --api --xformers --no-half --allow-code --opt-sdp-attention --opt-split-attention --skip-torch-cuda-test --disable-nan-check --force-enable-xformers --disable-safe-unpickle --update-check --update-all-extensions --cors-allow-origins http://localhost:5173 --upcast-sampling --precision full --no-half

set CUDA_VISIBLE_DEVICES=0

git pull

call webui.bat

Anything beyond that depends on your taste and the results you want. ?

You can use my generations for reference (click here), also, you for sure don't need:

(masterpiece, 8K, HD, Absurd Res, bla bla bla, insert noise here,)

Thank You!! ???

This model is only possible because there are people with incredible skills and effort. My intention at no time is to belittle any other model, on the contrary, ONE FOR ALL is an ode to these incredible people, without whom this model would not exist. Thank you very much!

If you happen to like the ONE FOR ALL series, please leave a good review and post your creations, I'll love to get to know the amazing universe you'll create! ???

If you want to thank me, like every Brazilian I love coffee, another way to encourage or thank me is to follow me on Instagram. ???

If you have any difficulties or questions, or just want to have a chat, you can contact me on Instagram or here on Civitai.

Last statements:

Please use this model for fun or work, but respect the legal regulations in force in your country and international treaties, especially regarding impersonation and pedophilia. I declare myself exempt from any violations resulting from the use of any model from the ONE FOR ALL series or its base models, being the creator of the content solely responsible.

This checkpoint (ONE FOR ALL «Furry») is not available for use outside Civitai on external generators (except Yodayo). If you find this checkpoint running on any generation site or available for download outside Civitai, please contact me.

If you are responsible for an online generation site and would like to use any model from the ONE FOR ALL series, please contact me.

Finally, all images generated using any ONE FOR ALL model belong to their creators (user), as well as the use-fruit and responsibilities for the content created. This means, but is not limited to, the right to sell the images generated, for example.

The use of this checkpoint is conditional on acceptance of these terms.

(SP, BR, 01/01/2024)