AnimaYume

I. Introduction

AnimaYume is a text-to-image model fine-tuned from Anima, a high-quality anime-style image generation model developed by CircleStone Labs. It builds upon Cosmos 2, a model developed by NVIDIA’s research team.

II. Information

For version 0.1:

• This model is a preview version fine-tuned from the Anima base model using a custom dataset. Training was conducted across multiple resolutions ranging from 768 to 1280 pixels, with a primary focus on 1024 pixels. The goal of this release was to improve stability and minimize unwanted artifacts when generating high-resolution images.

For version 0.2:

• This model is a continuation of AnimeYume v0.1. In this version, I improved the quality of my dataset and used several techniques to prevent oversaturation and low-quality outputs. Based on my testing, I found that prompt coherence is better than in v0.1, while the model remains very stable when generating images at a resolution of 1536 pixels.
• Note: I am still waiting for the final version of Anima and testing several methods to make my training process faster. I know the license might make the model less popular, but I mainly care about whether the model is good or not. I am aware that many other models use more permissive licenses, but I am too lazy to spend a huge amount of money training a model completely from scratch.

For version 0.25:

• This version was trained on Anima Preview 2. Due to several issues with the base model, such as overfitting, black and white borders, inconsistent quality, and problems with artist tags, I decided to focus primarily on improving the model's knowledge, reducing these issues, and making it as stable as possible.
• Note: In this version, I did not attempt to improve the model's style. I tried doing so, but it caused the model to forget some of its existing knowledge. The training process is similar to v0.2, but the dataset was adjusted to better address the issues present in Anima Preview 2.

For version 0.3:

• This version was trained using Anima Preview 2. It is an experiment with a new training method for the model. You can consider it another branch of AnimeYume v0.25, developed in parallel. However, this version uses new techniques and a larger dataset compared to v0.25.
• Note: In this version, I experimented with a new training approach, so the model is slightly different from v0.25

For version 0.4:

• This version was trained on Anima Preview 3 using a custom dataset. In this release, I improved prompt understanding and artist style reproduction. Based on my testing, some artist styles match my expectations, although I have not tested everything in detail because I am currently quite busy :<. Additionally, I fixed several issues from Anima Preview 3 that were also present in Preview 2.
• Note: I have only tested the model with simple test cases rather than conducting a comprehensive evaluation, so if you encounter any issues, feel free to let me know. I also used a larger AI computing cluster to speed up the training process :D.

For version 0.5:

• This version was trained on Anima Base v1.0 using my custom dataset, consisting of a mixture of a small e621 dataset and Danbooru. In this release, I added many new characters and improved the existing ones. I also enhanced support for various artist styles, allowing the model to generate results that are much closer to the original styles. In addition, the model now understands some concepts and knowledge from e621, although the support is still limited.
• Notes: I have only tested the model with a few simple test cases so far, so if you encounter any issues, feel free to let me know. This release can be considered a demo version showcasing my new training method, which focuses on preserving existing knowledge while adding new knowledge at the same time. The release also came sooner because I was finally able to use all the resources I had available :D.

For version 1.0:

• This version was fine-tuned on Anima Base v1.0 using a variety of datasets, including Danbooru, e621, Gelbooru, and Konachan. The training approach differs from the original model in several ways. Most notably, I did not include quality score tags in the training data.

• I also experimented with multiple captioning styles, ranging from traditional tag-based annotations to different forms of natural language descriptions, similar to the approach I used for Netayume Lumina.

• Note:

  • This release has two versions, each trained using different methods. For the v1.0 Demo, I experimented with a mixed training approach, but it was difficult to control. The v1.0 Final is different from the v1.0 Demo, so please do not compare them directly, even though they were trained on the same dataset. I created both versions to test different ideas for training diffusion models. Since Anima is relatively small, it gives me the flexibility to experiment with various training methods and see what works best.
  • Unlike Netayume, I did not use my full dataset for training. The complete dataset contains around 25 million images, and if I were to use the entire dataset, I would train a model from scratch rather than fine-tune an existing one. Additionally, I have been quite busy recently, so I have not been able to test this model as extensively as I would have liked. Moreover, this model does not have any default style :L.
  • V1.0 Final has a watermark embedded in the model. However, this does not affect the results of generated images. Moreover, v1.0 currently supports chain-of-thought prompting, as demonstrated in my example images.

• If you encounter any issues or have any feedback, please feel free to share them with me.

For version 1.5:

• This version was fine-tuned on Anima Base v1.0 using a variety of datasets, including Danbooru, e621, Gelbooru, and Konachan. The dataset used to train this version contains 5 million images (all data collected up to August 6, 2026), with nearly 25 million captions across different caption types, including tags, natural language descriptions, chain-of-thought, and others, corresponding to those 5 million images. This version was trained for 12 epochs and was completed on September 5, 2026, so I have not had enough time to conduct extensive testing yet. This is currently a base version without any aesthetic tuning.
• This version can understand many concepts from e621, Gelbooru, and Konachan. In addition, a large number of characters have been added to this version, including characters from almost all anime released during Summer 2026. The model can generate their general appearance, but the results may not always be accurate because there is not enough training data for these characters.
• For a more comprehensive comparison, I have included comparisons between this version and Anima v1.0, Anima 2.9B, and Anima Aesthetic 1.1 in the gallery.
• For the best results, I recommend using more complex prompts with this version.
• Note: V1.5 contains a watermark embedded in the model, just like V1.0. This does not affect the results of generated images. Moreover, this version may break all LoRAs trained on Anima v1.0, as well as LoRAs trained on merged or variant models based on Anima v1.0.

III. Model Components & Training Details

• Text Encoder: Pre-trained Qwen-3-0.6b
• Variational Autoencoder: Pre-trained Qwen Image VAE
• Image Backbone: Fine-tune Anima Image Backbone

IV. Suggestion

Recommended Settings

• CFG: 4–7

• Sampling Steps: 25-40

• Sampler: Euler a (with scheduler: normal)

V. Acknowledgments

• narugo1992 – for the invaluable Danbooru dataset
• CircleStone Labs - for creating the a wonderful model!