SilvermoonMix-Anima-Evolved

中文说明

An Anima fine-tune by silvermoong, using the "Evolved" training strategy with a focus on stable hands, a neutral art style that works with artist tags, and multiple characters.

Original release and sample images on Civitai

Anima 2.9B v2.3

Update ComfyUI or Forge Neo to the latest version and restart before loading either 2.9B model. Older versions may not load or run the model correctly.

Version	Download
v2.3 2.9B	Standard checkpoint
v2.3 2.9B INT8 + ConvRot	Standard INT8 checkpoint
v2.3 2.9B Turbo	Turbo checkpoint

The standard and Turbo BF16 checkpoints are 5.84 GB each. The standard model also has a 3.16 GB mixed INT8 + ConvRot version, quantized from its BF16 weights (45.96% smaller). All variants use the same separate text encoder and VAE listed below. SHA-256 checksums are in releases.json.

Choose BF16 for the original precision, or INT8 + ConvRot for a smaller download and potentially lower VRAM use and faster inference on supported hardware. Quantization can change image details and is not lossless; performance depends on hardware and backend. Use a current ComfyUI or Forge Neo with native INT8 + ConvRot support. This INT8 release is the standard model, not Turbo.

ComfyUI LoRA compatibility: Legacy Anima 2B LoRAs may need block remapping when used with the 2.9B versions. Community options include:

• ComfyUI-Anima-Remap
• ComfyUI-Anima-2.9B-loraPatch
• ComfyUI-Anima-LoRA-ControlNet-Patch

Please refer to each project's documentation and choose the solution that suits your workflow.

Recommended settings for standard and Turbo versions, including 2.9B, are listed below. See the Civitai release page for sample images.

Earlier 2B Downloads

These are BF16 diffusion-model checkpoints. The text encoder and VAE are separate downloads. Both checkpoints are byte-identical to their corresponding Civitai releases.

Version	Checkpoint	Civitai release
v2.3	SilvermoonMix-Anima1-Evolved-V23.safetensors	v2.3
v2.3.0 Turbo	SilvermoonMix-Anima1-Evolved-Turbo-V2.3.0.safetensors	v2.3.0 Turbo

SHA-256 checksums are in releases.json.

Recommended Settings

These settings follow the author's current Civitai model description.

Setting	Standard	Turbo
Sampler	Euler a	ER SDE
Steps	20-28	8-12
CFG	4 (2B) / 3 (2.9B)	1

• Start without a negative prompt. The Civitai sample images were generated without one.
• Artist names must be prefixed with @, for example @big chungus. Without this prefix, their effect is very weak.

Required Components

Both versions use the same components from the official Anima repository:

• qwen_3_06b_base.safetensors
• qwen_image_vae.safetensors

Their SHA-256 hashes match the corresponding text encoder and VAE on the Civitai releases.

ComfyUI

Use an up-to-date ComfyUI with native Anima support:

• Put the selected checkpoint in ComfyUI/models/diffusion_models/.
• Put the text encoder in ComfyUI/models/text_encoders/.
• Put the VAE in ComfyUI/models/vae/.
• Follow the official Anima workflow and setup instructions, selecting the SilvermoonMix checkpoint.

Forge Neo

The Civitai examples were generated with Forge Neo, which supports Anima natively. Place the checkpoint in models/Stable-diffusion/ and select the Anima text encoder and VAE. Right-click a sample image on Civitai and open it in a new tab to download the original file with its generation metadata.

Limitations

The model is intended for anime and illustration. Improved hand stability and multi-character generation are goals, not guarantees; anatomy and interactions can still need correction. Follow the upstream Anima limitations, including its limitations in photorealism and long-form text rendering.

License and Attribution

Use is subject to the CircleStone Labs Non-Commercial License v1.2, including its stated exceptions. Generated images may be used commercially under its terms. Rights to use the Anima model and its derivatives are granted directly by CircleStone Labs under that license.

SilvermoonMix-Anima-Evolved is a modification of Anima by silvermoong. It is not an official CircleStone Labs product and does not imply their endorsement, approval, or validation.

The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

Built on NVIDIA Cosmos. Anima is a derivative of Cosmos-Predict2-2B-Text2Image; the NVIDIA Open Model License Agreement applies insofar as it governs derivative models. See also NOTICE.

中文说明

SilvermoonMix-Anima-Evolved 是 silvermoong 基于 Anima、采用「Evolved」训练策略制作的微调模型，面向动漫与插画创作，主要着重于：

• 更稳定的手部表现。
• 较中性的基础画风，便于搭配不同画师标签。
• 多角色场景的表现。

Civitai 原始发布页与示例图片

新版 Anima 2.9B v2.3

加载 2.9B 普通版或 Turbo 版之前，请将 ComfyUI 或 Forge Neo 更新至最新版本并重启。 旧版本可能无法正确加载或运行模型。

• 下载 v2.3 2.9B 普通版
• 下载 v2.3 2.9B 普通版 INT8 + ConvRot
• 下载 v2.3 2.9B Turbo
• SHA-256 文件校验值

普通版和 Turbo 的 BF16 权重均约 5.84 GB。普通版现另提供由其 BF16 权重量化而来的混合精度 INT8 + ConvRot 版本，约 3.16 GB，文件缩小 45.96%。所有版本沿用下方列出的独立文本编码器和 VAE。

BF16 保留原始精度；INT8 + ConvRot 的下载文件更小，在支持的硬件上可能减少显存占用并加快推理。量化可能改变图像细节，并非无损转换；实际收益取决于硬件和后端。请使用原生支持 INT8 + ConvRot 的较新 ComfyUI 或 Forge Neo。本次 INT8 发布对应普通版，不是 Turbo。

ComfyUI LoRA 兼容性：在 2.9B 版本上使用旧 Anima 2B LoRA 时，可能需要层号重映射。可参考以下社区方案：

• ComfyUI-Anima-Remap
• ComfyUI-Anima-2.9B-loraPatch
• ComfyUI-Anima-LoRA-ControlNet-Patch

请自行查阅各项目说明，选择适合自己工作流的方案。

普通版和 Turbo（含 2.9B）的推荐参数见下方；示例图片请参阅 Civitai 发布页。

旧版 2B 版本与下载

本仓库提供普通版 v2.3 和低步数版 v2.3.0 Turbo。两者均为 BF16 扩散模型权重，与对应的 Civitai 发布文件完全一致；文本编码器和 VAE 需要单独下载。

• 下载 v2.3 普通版
• 下载 v2.3.0 Turbo
• SHA-256 文件校验值

推荐参数

以下参数与作者当前的 Civitai 模型简介一致：

参数	普通版	Turbo
采样器	Euler a	ER SDE
步数	20-28	8-12
CFG	4 (2B) / 3 (2.9B)	1

建议先不使用负面提示词，Civitai 示例图片也未使用负面提示词。**必须在画师（或艺术家）名字前加上 @**，例如 @big chungus；否则效果会非常微弱。

所需组件与使用方法

普通版和 Turbo 共用以下官方组件，其文件哈希与 Civitai 配套组件一致：

• 文本编码器 qwen_3_06b_base.safetensors
• VAE qwen_image_vae.safetensors

ComfyUI：使用原生支持 Anima 的较新版本，将模型放入 ComfyUI/models/diffusion_models/，文本编码器放入 ComfyUI/models/text_encoders/，VAE 放入 ComfyUI/models/vae/。参照 Anima 官方工作流与安装说明，将扩散模型切换为 SilvermoonMix。

Forge Neo：模型放入 models/Stable-diffusion/，并选择 Anima 对应的文本编码器和 VAE。Civitai 示例图片由原生支持 Anima 的 Forge Neo 生成。右键并在新标签页中打开 Civitai 示例图片，即可下载带有生成参数的原图。

局限与许可

手部稳定性和多角色表现是优化目标，不代表每张图都能正确生成；人体结构及角色互动仍可能需要调整。本模型主要面向动漫与插画，照片写实和较长文本渲染等限制请参阅 Anima 官方说明。

使用本模型须遵守 CircleStone Labs Non-Commercial License v1.2 及其中明确规定的例外。许可证允许按其条款将生成图片用于商业用途，但不能据此认为模型权重及所有部署方式均可自由商用。相关使用权由 CircleStone Labs 依据该许可证直接授予。

本模型由 silvermoong 修改制作，并非 CircleStone Labs 的官方产品，也不代表其认可、批准或验证。模型基于 NVIDIA Cosmos，亦须遵守适用于衍生模型的 NVIDIA Open Model License Agreement。完整声明见上方英文许可部分及 NOTICE。