LEOSAM FilmGirl Ultra Realistic Big Model

🚨FilmGirl Ultra recommended live-picture parameters:

Clip skip: 1

CFG scale: 9

Sampler: Restart/DPM++ 2M Karras

Direct output image size: Approximately 500,000 pixels (640x768)

Saying Goodbye to FilmGirl Ultra and SD1.5's AI internet celebrity face:

On February 24 of last year, I completed the production of the first version of FilmGirl LoRA. This LoRA was my first high-download model, marking the beginning of my AI journey dreams. Since the release of SDXL, I have invested a lot of energy in improving the two XL models, HelloWorld and AIArt. The FilmGirl series has not been updated for 8 months.

Whether it's FilmGirl, the later Instant LoRA, or Helloworld XL that followed, I have always pursued the ultimate realism. Now a year has passed, and as an anniversary celebration, I have decided to release a model that can elevate the realism of SD1.5 to new heights. The LoRA model is no longer enough to carry out this mission. The new FilmGirl Ultra is an SD1.5 XL model.

To completely eliminate the homogenization of large models and AI face issues in SD1.5 realism, FilmGirl Ultra did not choose the basilmix, chilloutmix, and their descendants as training base models, but instead selected UCLA's latest release of SPIN-Diffusion. SPIN-Diffusion is an SD1.5 base model fine-tuned through self-play using winning images from the pickapic_v2 dataset, performing better than the original SD1.5 base model and DPO base model, while indicating far superior alignment performance compared to Chilloutmix and other extensively fine-tuned and fused base models.

The training set of FilmGirl Ultra comes from HelloWorld XL. In fact, the training set used by the first version of HelloWorld XL comes from the last version of FilmGirl LoRA. I have been meticulously accumulating and screening this training set over the past year, and now the total number of images in the training set has reached 10,000. Various labeling methods, including GPT4V natural language captions, GPT4V tag-based captions, and Blip+Clip captions, were used throughout the entire training process of FilmGirl Ultra. Concluding a year of meticulous accumulation, the entire training set amounted to about 10,000 images. To trigger the effects of 1girl, best quality, and masterpiece that are commonly used, I have appropriately added these three words to some images (however, you can still accurately trigger the effect of a little girl through the words child girl/little girl). Using multiple sets of labels is intended to maximize the probability of the training set effectively triggering the desired effect. As a tradition of FilmGirl, special attention was paid to the film style, and you can trigger this style by using film grain analog photography.

The model underwent a total of 7 stages of training, utilizing different batch sizes, optimizers, learning rates, and training set proportions in each stage to achieve its current effectiveness. If you are interested in fine-tuning SPIN-Diffusion, I recommend a total training iteration of more than 50,000 steps. In practice, I trained for about 100,000 steps using a batch size of 40-64.

The realism of FilmGirl Ultra has exceeded my expectations and is now comparable to the image quality of SDXL. Below are comparisons of this model with Realistic Vision v6 and epiCPhotoGasm. The former is currently the most downloaded 1.5 base model on a certain site, and the latter has been the most realistic 1.5 base model I have seen for a long time. I pay homage to these two excellent models and their creators.

close-up couple's portrait, African young woman and man, clear skin face, looking at camera, fashion photography, simple background
Negative prompt: watermark, anime, cartoon, open mouth

close-up couple's portrait, African little girl and boy, clear skin face, looking at camera, fashion photography, simple background,
Negative prompt: watermark, anime, cartoon, open mouth,

Thanks to the GPT4V labeling and SPIN-Diffusion base model, this model performs excellently in prompt word alignment.

Ethnicity Test

Body Type Test

Skin Color Test

Age Test

Animal Test

However, FilmGirl Ultra is not leading in all dimensions. Starting from a new point, it has abandoned the continuous optimization and polishing of the community for over a year regarding the 1.5 base models. After extensive testing and comparison by me, this base model has a higher limb error rate than mature community realistic models. Additionally, due to a lack of anime-related content in the training set, the output quality is not optimal when the prompt words include related anime tags. I recommend avoiding words like digital art, anime, and cartoon. These are currently the primary two deficiencies of FilmGirl Ultra.

FilmGirl Ultra serves as the year-end summary of my first year on the AI journey, a gift to the AI enthusiasts who have supported me. The open-source community has brought me many friends, memories, joy, and knowledge, and I also hope to contribute back to the community. I hope that the above model-making summary can be helpful to everyone, and I also welcome you to train or fuse your models based on FilmGirl Ultra. If you find this model helpful in improving your own model, please mention it in the model description. I hope that FilmGirl Ultra and SPIN-Diffusion can be better understood and used by more people.

There will be continuous updates for FilmGirl Ultra in the future. Enjoy using it!

Let's continue to progress together with AI, and hope that we can still meet here at this time next year!

Copyright Notice:

The FilmGirl Ultra series model (referred to as "this model" hereafter) is an SD1.5 XL model developed based on SPIN-Diffusion by me (referred to as "owner").

The owner authorizes individuals or institutions to use the images generated by this model for non-commercial educational or informational purposes, under the following conditions:

- Comply with relevant legal regulations and do not infringe upon the legitimate rights and interests of this model or any third party.

- When using images, attribute the image source to "Generated by LEOSAM's FilmGirl Ultra large model".

For commercial use, a commercial authorization agreement must be signed with the owner beforehand. For commercial authorization and model customization inquiries, please contact the owner through the information on their LiblibAI platform homepage.

The owner will continue to provide free updates for the FilmGirl Ultra model for individual players, as a show of support and gratitude to the open-source community contributors. Commercial user cooperation is an important driving force for the development and continuous improvement of this model. Thank you for the understanding and support of every user.

Please note that any unauthorized use may violate relevant legal regulations and may incur legal liabilities. The ultimate interpretation right of this declaration belongs to the owner and is subject to the relevant laws and regulations of the People's Republic of China.