SG_161222

Hi everyone! Today I will tell you about how the Verus Vision model was trained.

Now, I will show the parameters I used for training LoRAs on CivitAI.

Hello everyone!

Today I released RealFlux 1.0b Schnell on Hugging Face. The model is available in two variants: Compact (a model containing VAE, Clip-L, and T5XXL) and Transformer (for those who don't need to download the full model and already have VAE, Clip-L, and T5XXL). This version of the model has some limitations (listed on the model's page).

Additionally, a model based on Flux Dev, which can generate NSFW content and has fewer anatomical issues than the Schnell model, is currently in testing. Work on the models is done almost every day (sometimes small breaks for rest are needed).

Plans for the Schnell model include adding multiple styles, adding basic NSFW content generation, improving anatomy, and refining existing concepts.

I hope you enjoy the first version of the Schnell model. Thank you all and have a great day!

Hello everyone! Currently, I am also training a Flux model in addition to RealVisXL V5.0.

The first training session has already been conducted, which revealed some errors I made in describing my dataset. These errors were corrected manually, and now the second training session is underway based on the corrected dataset. The dataset is not very large at the moment, but it should be sufficient for the first Beta version.

Once OneTrainer supports Flux model training, I will continue to improve my RealFlux model by adding new concepts and enhancing existing ones.

Below are images from the first test version.

Thank you all for your support, and have a great day!

 
Hello everyone!

As of the time of this post, the model has already gone through 340,000 training steps (progress is constantly updated at https://huggingface.co/SG161222/RealVisXL_V5.0 and https://civitai.com/models/139562/realvisxl-v40 in the model description).
I'll say right away that you shouldn't expect a huge leap forward from this version of the model. It still has some issues with hands, which I will try to fix using Masked Training (I'll report on the results in the next post).

Currently, I am finishing the first main part of the dataset focused on photo. Besides this part, there are others focused on different aspects of the model. I wanted to release the model earlier than it will be now, but it will take a bit longer as I want to achieve improvements not only in photography. As soon as the model is very close to its final form, I will announce the release date.

Thank you all for your support and have a great day!

I've been closely following the developments of SD3 (2B), and after much consideration, I've decided not to invest my time and resources into fine-tune this model.

Here are my reasons:

1. Unclear Licensing: The licensing terms of SD3 (2B) are ambiguous and create uncertainty about how the model can be used and shared. This lack of clarity poses significant risks for anyone looking to build and deploy applications based on SD3 (2B).

2. Censorship Issues: The creators of SD3 (2B) have implemented certain techniques to censor content. While well-intentioned, these techniques have negatively impacted the model's ability to accurately render human anatomy, making it resemble the issues seen in SD 2.0/2.1.

3. Waiting for a Better Alternative: I believe that the field of AI is rapidly evolving, and it's only a matter of time before we see a more robust and reliable alternative to SD3 (2B). I'm choosing to wait for a solution that addresses the current shortcomings of SD3 (2B).

For now, I will continue exploring other options and remain hopeful for a future model that meets the high standards we all expect.

Thank you for your understanding and support.

Hello everyone! Recently, I released the Hyper version of the RV5.1 model, which surpasses the regular version in realism and aims for the photorealism of RealVisXL. I believe I can do even more and better in the V6.0 B2 Hyper version, which will be trained on the same sizes as B1. Currently, the dataset for training is being collected, and as soon as there is news, I will share it with you. Have a great day!

I'm completely re-training Realistic Vision V6.0 model using a new trainer called OneTrainer, which has shown great results so far. My goal is to achieve more realistic images as well as more flexibility of the model. I'm using a dataset of high-quality images, mostly photos, and some high-quality images generated with RealVisXL.

The dataset I'm using is generic, which means it can be used for SDXL and Stable Diffusion 3 training.

Stay tuned for updates, and thank you for your support!