Key Content:
/Lotus Anime Diffusion/
Today we’re releasing Anime 2.5.5, our most powerful Anime model to date. This open release includes several customizable variants that run on consumer hardware and can be used under the permissive Eschscholzia AI Community License. You can quickly download weights by visiting the Eschscholzia Model Panel and download inference code on our publishing platform.
In July, we released Lotus Anime V5.3, the latest generation of the basic Lotus Anime series, but generally, due to a lack of polish, the models did not meet user expectations, so we planned to redesign and plan the Anime series models based on the latest DiT architecture. We completed the official release of Lotus Anime NG2.0 in October, which was well received, so we focused on reducing resources and increasing aesthetic diversity in Anime 2.5.5, using most of the mature technical approaches of NG2.0, with a few changes and technical innovations.
Lotus Anime 2.5.5NG continues to reflect our commitment to providing builders and creators with widely accessible, cutting-edge, and free tools. We encourage developers to use our models for fine-tuning, Lora, applications, or other non-profit activities.
What’s being released
Lotus Anime NG 2.5.5β Due to stability and the first open beta, only one model is currently available to meet the needs of researchers, hobbyists, startups and enterprises:
Lotus Anime NG 2.5 Lite_Preview(A25P) : With 2 billion parameters and improved architecture and training methods, the model can run on most mainstream consumer-grade hardware (but is still resource-constrained) and can run most functions smoothly on the vast majority of computing units with 16 gigabytes of memory and above, striking a balance between quality and ease of customization. It is capable of producing Anime images at up to 2 megapixel resolutions, and supports infinite left-right extension (Pan functionality) and HD diffusion magnification (in testing).
Unified and diverse art forms
By tweaking the Anime inference options, you can quickly pick up multiple styles or stick to the same style while keeping your aesthetic in line with the baseline.
Development model
When developing the model, we prioritized customizability to provide a flexible foundation on which to build. To achieve this, we continue to integrate as much as possible into the transformation block, stabilizing the model training process and simplifying further fine-tuning and development.
To support this level of downstream flexibility, we have to make some trade-offs. The same hint using different seeds can produce a greater difference in output, which is intentional because it helps to preserve a broader knowledge base and different styles in the base model. However, due to the lack of specificity of the cue, which may lead to increased uncertainty in the output, the aesthetic level may vary.
It is important to note that the current 2.5 Lite model cannot generate any sensitive content, and the generation of such content after Finetune may violate our Terms of Service, which we may restrict with appropriate trade-offs.
Our commitment to safety
We believe in safe, responsible AI practices and take deliberate steps to ensure integrity starts at an early stage of development. For more information about our terms of use, you can contact [email protected] for a detailed, worldwide legal terms of use.
Cooperation with Stability
Eschscholzia strictly adheres to the Stability AI Community license and thanks Stability for its guidance and assistance in the development process. Lotus Anime NG 2.5.5 is a highly modified version of SD3.5 Medium that includes secondary distillation and structural changes. For more information, consult [email protected], Lotus is an open source product for the market, we do not make any profit from it, and this applies to non-commercial regulations.
This also means that you may need to follow a stable usage agreement to use the model properly.
