Anima Model’s Dual Licensing! Open Source Technology from Alibaba and NVIDIA’s Ambition

Introduction

Hello, this is Easygoing.

This time, I’d like to explore the structure and licensing of the new anime-specialized model called Anima.

What is the Anima Model?

The Anima model is a text-to-image AI model specialized in anime-style illustrations, released by CircleStone Labs.

gantt
    title Image Generative AI Roadmap
    dateFormat YYYY-MM-DD
    tickInterval 12month
    axisFormat %Y

    section Black Forest Labs
        Flux.1 : done, 2024-08-01, 2026-05-01

    section Alibaba
        Wan2.1 : 2025-02-25, 2026-05-01
        Wan2.2 : 2025-08-07, 2026-05-01
        Qwen-Image : 2025-08-04, 2026-05-01

    section NVIDIA
        Cosmos-Predict1 : done, 2025-01-06, 2026-05-01
        Cosmos-Predict2 : 2025-04-30, 2026-05-01
        Cosmos-Predict2.5 : 2025-10-06, 2026-05-01

    section CircleStone Labs
        Anima Preview 1 : 2026-01-26, 2026-05-01
        Anima Preview 2 : 2026-03-11, 2026-05-01
        Anima Preview 3-base : 2026-04-08, 2026-05-01

The Anima model was first released as Preview1 on January 26, 2026, and as of May 2026, it has progressed to Preview3-base.

The Roots of Anima: Alibaba’s Wan

The Anima model is composed of the following components:

• Text Encoder: Qwen3 lightweight version (Alibaba)
• Transformer: Cosmos-Predict2-2B (NVIDIA)
• VAE: Qwen-Image_vae (Alibaba)

The Anima model uses Alibaba’s lightweight Qwen3 as its text encoder, NVIDIA’s Cosmos-Predict2-2B as its transformer, and Alibaba’s Qwen-Image_vae as its VAE.

Anima Uses a Dual Licensing Structure

While the text encoder and VAE are released under the open-source Apache-2.0 license, the transformer part adopts a dual licensing structure that combines NVIDIA’s and CircleStone Labs’ respective proprietary licenses.

To make the licensing of the Anima model easier to understand, let’s organize the path that led to its creation.

Organizing the Timeline by Focusing on the VAE

When looking at the origins of image generation AI models, focusing on the VAE (Variational Auto Encoder) makes things much clearer.

The VAE is a specialized model that generates the latent space for efficient AI computation. If you change the VAE, you generally need to restart training from scratch.

Stable Diffusion and Flux as the Foundation

Open-source image generation AI began with Stability AI’s Stable Diffusion.

gantt
    title VAE Roadmap
    dateFormat YYYY-MM-DD
    tickInterval 12month
	axisFormat %Y

    section Stability AI
        Stable Diffusion 1 : done, 2022-08-22, 2026-04-18
        Stable Diffusion XL 0.9 : done, 2023-06-22, 2026-04-18
        
    section Black Forest Labs
        Flux.1 : 2024-08-01, 2026-04-18
        Flux.2 : 2025-11-25, 2026-04-18
		

Later, former Stability AI staff founded Black Forest Labs and developed Flux.1.

The Flux.1_vae not only dramatically improved performance compared to previous versions but was also released under the Apache-2.0 license, allowing free development and commercial use.

As a result, it has been adopted by many image generation AI models.

VAE Performance Comparison

Image Generation AI Model License Overview

Alibaba Develops Wan_vae

China’s Alibaba then developed Wan_vae, which added temporal understanding based on Flux.1_vae. In February 2025, they released the video generation model Wan-2.1 that incorporated this VAE.

gantt
    title Image Generative AI Roadmap
    dateFormat YYYY-MM-DD
    tickInterval 12month
    axisFormat %Y

    section Black Forest Labs
        Flux.1 : done, 2024-08-01, 2026-05-01

    section Alibaba
        Wan2.1 : 2025-02-25, 2026-05-01
        Wan2.2 : 2025-08-07, 2026-05-01
        Qwen-Image : 2025-08-04, 2026-05-01

    section NVIDIA
        Cosmos-Predict1 : done, 2025-01-06, 2026-05-01
        Cosmos-Predict2 : 2025-04-30, 2026-05-01
        Cosmos-Predict2.5 : 2025-10-06, 2026-05-01

    section CircleStone Labs
        Anima Preview 1 : 2026-01-26, 2026-05-01
        Anima Preview 2 : 2026-03-11, 2026-05-01
        Anima Preview 3-base : 2026-04-08, 2026-05-01

Because Wan-2.1 was released under the Apache-2.0 license, which allows free development and commercial use, it became the most widely adopted model for local video generation AI, along with its successor Wan-2.2.

Furthermore, Alibaba adjusted the Wan_vae structure for still images, creating Qwen-Image_vae, which is also widely used in their image generation model Qwen-Image.

NVIDIA’s Cosmos-Predict2 Is Based on Wan

Now let’s turn our attention to NVIDIA.

NVIDIA originally had a physics simulation AI model called Cosmos-Predict1, but it was not specialized for image or video generation and significantly lagged behind competitors in both performance and efficiency.

gantt
    title Image Generative AI Roadmap
    dateFormat YYYY-MM-DD
    tickInterval 12month
    axisFormat %Y

    section Black Forest Labs
        Flux.1 : done, 2024-08-01, 2026-05-01

    section Alibaba
        Wan2.1 : 2025-02-25, 2026-05-01
        Wan2.2 : 2025-08-07, 2026-05-01
        Qwen-Image : 2025-08-04, 2026-05-01

    section NVIDIA
        Cosmos-Predict1 : done, 2025-01-06, 2026-05-01
        Cosmos-Predict2 : 2025-04-30, 2026-05-01
        Cosmos-Predict2.5 : 2025-10-06, 2026-05-01

    section CircleStone Labs
        Anima Preview 1 : 2026-01-26, 2026-05-01
        Anima Preview 2 : 2026-03-11, 2026-05-01
        Anima Preview 3-base : 2026-04-08, 2026-05-01

Cosmos-Predict1 Model Composition

Cosmos-Predict2 Model Composition

In April 2025, NVIDIA developed Cosmos-Predict2 by adopting Wan_vae and retraining the model from scratch. In essence, it can be described as the NVIDIA version of Wan.

Of course, it wasn’t built entirely from scratch—some datasets and parts of the architecture were likely carried over from Cosmos-Predict1.

Cosmos-Predict2 Comes with Strict Licensing

While Alibaba’s Wan uses the permissive Apache-2.0 license, Cosmos-Predict2 switched to NVIDIA’s proprietary NVIDIA Open Model License.

Key Features of the NVIDIA Open Model License:

• NVIDIA does not claim rights to the outputs.
• Developers of derivative models can claim rights to their own additions, provided they acknowledge NVIDIA’s ownership of the base components.
• If a third party makes a claim against NVIDIA related to a derivative model, the derivative model developer must bear the costs.
• If the NVIDIA Open Model License is revised, the new license applies retroactively to existing models.

The license explicitly states that NVIDIA retains ownership of the parts it developed, even in derivative models.

If legal claims arise due to a derivative model, the developer is obligated to cover NVIDIA’s expenses in NVIDIA’s place.

Furthermore, even existing models must comply with any future revisions to the license. Failure to do so may require the model to be taken down.

Cosmos-Predict2 can be described as a very strict license (essentially terms of use) that forces derivative developers to act as NVIDIA’s shield.

Anima Was Fine-Tuned on Millions of Anime Illustrations

Let’s look again at Anima’s structure.

The Anima model replaces the text encoder in NVIDIA’s Cosmos-Predict2-2B with the lightweight Qwen3-0.6B-Base.

The transformer uses the Cosmos-Predict2-2B model as-is but has been further fine-tuned on millions of anime illustrations.

The VAE has been changed to Qwen-Image_vae, the still-image version of Wan_vae. Since Qwen-Image_vae maintains compatibility with Wan_vae, structural consistency with Cosmos-Predict2-4B is preserved.

Anima Uses a Dual License from NVIDIA and CircleStone Labs

The text encoder and VAE of the Anima model are under Alibaba’s Apache-2.0 license.

However, the transformer part inherits the NVIDIA Open Model License and adds CircleStone Labs’ own restrictions under the CircleStone Labs Non-Commercial License v1.0.

CircleStone Labs Non-Commercial License v1.0:

• Commercial use of the model itself is prohibited.
• Commercial use of the outputs is allowed.
• Using the outputs to train other models is prohibited.

This license is similar to Black Forest Labs’ FLUX [dev] Non-Commercial License. It was likely modeled after it.

Both prohibit commercial use of the model while allowing commercial use of the generated outputs. However, even with FLUX [dev], opinions are divided on whether outputs can truly be used commercially (many interpret it negatively).

Therefore, at this point, it is safest to avoid commercial use of the Anima model.

Why Choose Cosmos-Predict2 Instead of Wan?

Here’s where my questions deepen.

As explained earlier, Anima is subject to dual licensing restrictions. If CircleStone Labs had chosen Alibaba’s Wan model—which offers comparable performance—instead of NVIDIA’s Cosmos-Predict2, they would not have been bound by NVIDIA’s restrictive license.

Why would CircleStone Labs choose Cosmos-Predict2 as the base model, even if it meant taking on the obligation to indemnify NVIDIA against damages and accepting the risk of the model being taken down?

Anime Illustration Has a Strong Local/Offline Orientation

Anime-style illustrations are popular not only in East Asia but around the world.

Among image generation AI fields, anime has particularly high demand for NSFW content, making it a domain that strongly prefers local/offline usage rather than online services.

Because of its unique characteristics, active technical development continues, resulting in high-performance community models such as Illustrious-XL and Animagine-XL that far exceed what one would expect from the original base models.

ComfyUI Dominates Next-Generation Models

Currently, the most widely used tools for anime illustration are derivative models based on Stable Diffusion XL, with ownership held by Stability AI and the respective derivative developers.

In contrast, for next-generation models such as Flux and Z-Image, ComfyUI has established a near-monopoly as the go-to generation tool.

Why ComfyUI is Dominant

If NVIDIA can popularize an anime model under its license to replace SDXL through ComfyUI, it would allow NVIDIA to exert overwhelming dominance even in the local development environment, which is normally difficult to control.

Anima Is Strongly Backed by Comfy Org

The development of the Anima model has received strong support from Comfy Org.

In January 2026, Comfy Org announced that it provided $1 million in funding to CircleStone Labs as part of its open-source model development support program.

Around the same time, on January 24, 2026, Comfy Org announced a 30% price reduction for Comfy Cloud, citing improved access to GPU compute power at lower costs.

For cloud AI service providers including Comfy Cloud, GPU rental costs directly impact their business. The allocation of GPU resources naturally involves the influence of NVIDIA, which supplies over 90% of industrial-grade GPUs worldwide.

Although there is no official document linking the Anima model directly to NVIDIA, I suspect that the decision to use Cosmos-Predict2 was not made by CircleStone Labs alone, but was likely predetermined from the beginning.

Summary: Open Source and NVIDIA’s Ambition

• Anima is an image generation AI whose roots trace back to Wan2.1.
• It is subject to NVIDIA’s strict licensing restrictions.
• It is safest to avoid commercial use for now.

In this article, I examined the licensing of the Anima model.

When the excellent open-source video generation AI called Wan was released, NVIDIA quickly abandoned its previous assets, adopted the new technology, leveraged its overwhelming computing power to rapidly create a model of equal or greater performance, and then deployed it along with restrictive licensing to dominate the ecosystem. This move truly demonstrates the dignity of a champion.

For now, I have no plans to develop using the high-risk Anima model. However, given the immense power of NVIDIA and Comfy Org, I may not be able to say the same in six months.

I will continue to observe closely to understand what “open source” truly means to NVIDIA and Comfy Org.

Thank you for reading until the end!