ILLUSTRIOUS V 1.0 - V 3.5 DPO+VAE

one-for-all-anime-w-illustrious-v-1-0-v-3-5-dpo-vae

Text to Image Community ModelFree for Premium Users LLMs.txt

Thank you very much for using the ONE FOR ALL series!!

Version 3.0 presented some inconsistencies when generating images at higher resolutions such as 800 x 1280 @ 70 Steps, a lot of noise started to appear, as well as an excess of flare and glare pollution, which was very much outside the intention of the model, so I was forced to discard version 3 and upgrade to version 3.5. As it was more a question of adjustments, I thought it best to rename it to version 3.5.

This version 3.5 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.

More information: 2311.12908.pdf (arxiv.org )

In addition to the innovative DPO system, this version also features:

Sharper eyes using proportional scaling with symmetry-trained images;
Ethnic training. I was always tired of seeing the same Asian face, so with that in mind, I included non-Asian images in the training. But there was still a need for greater diversity and a more cosmopolitan sense of inclusion. This led me to insert over a thousand images with a more extensive range of ethnicities;
There was a dynamic increase in details, so it's important to be careful when using LoRAs that enhance details. In the case of saturation, I recommend using LoRAs with negative values;
In this version, I included dozens of quadratic meta-model AI diagrams. This training will provide a greater sense of Ray Tracing for those who like images with brightness and flares. You'll hardly find a similar model; MidJourney, watch out, lol;
The entire training was done with high-definition images, at least 16 bits, and a resolution of 720p and above. This ensures a greater sense of authentic lighting and intricate textures;
During the training, there was an enhancement of the color performance of the screen and an increase in saturation compared to previous versions. Brighter colors will be easily obtained;
Hands! It's still challenging to work with hands. However, during training, (EnvyBetterHands LoCon) was used with smaller weights within the training parameters. This should help achieve better overall results;
Last but not least, training with NSFW was conducted, aiming for greater variety and quality in generations.

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: