banner_image ×
discard-icon
vip-icon

ONE FOR ALL «Plus Ultra»

v 3 DPO+Pony
v. Final
v. Final AB (VAE)
v. Final AB (no VAE)
v 2.0 DPO+VAE
v1.0 (VAE)
最終更新:2024-09-24 18:07:17
avatar
創作回数
1.8K
お気に入り
29
ダウンロード
255
モデルソース:
タイプ:
Checkpoint
基本モデル:
SDXL 1.0
トリガーワード:
許可範囲:
創作許可範囲
オンライン画像生成
統合
ダウンロード
商用利用の許可範囲
生成された画像は販売または商業目的での使用
モデルの転売やモデル統合後の販売
Model Parameters:
評価:
5

About version 3

First of all, the "Plus Ultra" model is an SDXL 1.0 model with DPO. However, given the popularity of Pony models, I took the step of adapting the checkpoint to understand Pony tags and reproduce similar results. However, please note that this is not a Pony model in essence, so some prompts may not generate the intended result. Also, compatibility with LoRAs Pony tends to be perfect, so feel free to use LoRAs for Pony.

Version 3 was made with twice as much training as version 2, reinforcing the DPO and significantly increasing the quality of the generations. However, I would like to point out that the model may need LoRA for more specific NSFW content, especially sexual acts. This is because I preferred to pay attention to other factors when training the base model.

Finally, this is the last version of the ONE FOR ALL series in the XL segment. As explained in version 2, the training and results of the XL models are a real nightmare and end up requiring very high energy consumption and the series as a whole does not generate any financial return, so there is no point in continuing.

About DPO (Direct Preference Optimization).

DPO fine-tune on datasets scored as highly visually appealing by an aesthetics classifier, to bias the model to visually appealing generations, while maintaining distributional guarantees and improving generic text-image alignment in addition to visual appeal. The training is via gradient descent, automatically validate checkpoints with the 500 unique prompts of the Pick-a-Pic validation set: measuring median PickScore reward of generated images. DPO scheme implicitly learns the differences in between images by taking an expectation over the inner term of the prompt.

So, DPO pay more attention on your prompt and try to generate a better image from it.

I'm very grateful to the authors of the Checkpoints and LoRAs used in the ONE FOR ALL series and I recommend everyone to visit them and appreciate their work:

  • PurpleSmartAI - Thank you for the hard work creating the Pony Diffusion.

And to all those who in some way collaborated or encouraged this work, I affectionately offer this simple Special Thanks:

Special Thanks to those creators:

Special Thanks to my amazing followers:

AND YOU WHO IS AMAZING TO READ SO FAR HERE!! WOW!! 😱🤩💓

YOU ARE REALLY AMAZING!! THANK YOU VERY MUCH!! 🥰🥰🥰

もっと見る
コメント:0件
29
25
0