Sign In

Smooth Rotation Slider LoRA (Left/Right)

2

Updated: Jun 21, 2026

conceptsliderfacing

Download

1 variant available

SafeTensor

16.63 MB

Verified:

Type

LoRA

Stats

194

0

Reviews

Published

Jun 21, 2026

Base Model

Anima

Hash

AutoV2
CDC8D0C057

Trigger Words

consistent light
default creator card background decoration
Bronze Anima Badge
FourBunny's Avatar

FourBunny

License:

Anima

The Anima Model is licensed by CircleStone Labs LLC. 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

更新 v2.0 / Update v2.0

【主要改进 / Key Improvements】

✅ 使用了10+种不同画风的训练素材,大幅缓解了画风污染问题

Trained with materials spanning 10+ distinct art styles, significantly reducing the previously reported style-pollution issue.

✅ 可用权重范围扩大至 -5 到 5

Usable weight range expanded to -5 to 5.

✅ 优化了训练参数(采样策略、锚点设置),整体稳定性进一步提升

Refined training parameters (sampling strategy, anchor settings) for improved overall stability.

在提示词中加入 side view 或 profile view,即可在 -1 到 1 的权重区间快速稳定获得90度侧脸方向,方向控制更加精准听话。
Improved View Control: Simply add side view or profile view in the prompt. You can now quickly and stably achieve precise 90-degree side/profile angles within -1 to 1 weight range

感谢大家的反馈,帮助这个模型不断改进!

Thanks to everyone for your feedback — it's helping this model keep improving!


V1.0:

This version has been retrained using a completely new prompt template. the slider can now directly control left and right rotation without requiring additional angle prompts.

本版本使用全新的提示词模板重新训练

现在无需额外添加视角提示词,即可直接控制人物向左或向右转动。

For most situations, a weight range of -2 to +2 is recommended.

大多数情况下,推荐使用 -2 到 +2 的权重范围。

This range provides smooth rotation while maintaining image quality and character consistency.

该范围能够在保持画面质量和人物一致性的同时,实现平滑转向。

For stronger side views, you can extend the range to approximately -4 to +3.

如果需要更强的侧面效果,可以将范围扩大至 -4 到 +3 左右。

You can simply use the slider with your normal prompts.

你可以直接配合普通提示词使用本滑块。

Addingfront_viewcan serve as a neutral baseline for more predictable rotation control.

添加front_view可以作为一个中性基准,使转向控制更加稳定和可预测。

Please avoid using weights above +3, as image quality may become unstable or blurry.

请尽量不要超过 +3,否则画面可能出现模糊或不稳定现象。

Angle Prompts / 视角提示词

Traditional angle prompts are no longer required for the slider to function.

传统视角提示词已经不再是本滑块生效的必要条件。

However, you can still combine it with:

不过你仍然可以搭配以下提示词使用:

  • three quarter view

  • side view

  • profile view

These prompts can make the rotation effect more sensitive while preserving image quality.

这些提示词能够让转向效果更加明显,同时不会对画面质量造成负面影响。

Every character LoRA and base model is different — go test and find your number 🎲

每个角色LoRA和底模的最佳权重都不一样,自己多测多玩 🎲


🧠 Author’s Notes / 一些碎碎念

After failing more than ten different left/right view slider experiments, I learned one important lesson:

在经历了十多个失败的左右视角滑块实验后,我总结出了一个非常重要的经验:

Prompt design is often more important than training parameters.

提示词设计的重要性,很多时候甚至高于训练参数本身。

Earlier on, I had been focusing on focus only on angle keywords such as:

刚开始我一直把重点放在视角关键词上,例如:left view/facing left/side view

However, these words are often too abstract for the model.

但对于模型来说,这些词往往过于抽象。

The anima model does not truly understand "left" or "right" the way humans do.

ANIMA模型并不会像人类一样真正理解“左”或“右”的概念。

Instead, it learns from what is visually present in the image.

它真正学习的是图像中实际出现的视觉特征。

Because of this, controlling only angle-related keywords may produce weak or inconsistent results.

因此,仅仅控制视角关键词,往往会导致训练效果不稳定,或者控制力不足。

A better approach is to describe what should actually be visible in the image.

更好的做法是直接描述你希望在画面中看到什么。

For example,Instead of:left view Try describing:left cheek clearly visible/left ear visible/right side of face partially hidden/left shoulder closer to camera etc;

These are concrete visual features that the model can directly learn from.

In my experience, a slider becomes much more reliable when prompts describe visible structure rather than abstract concepts.

例如,不要只写:left view而是描述:左脸颊清晰可见/左耳可见/右侧脸部部分遮挡/左肩更靠近镜头;这些都是模型能够直接学习的具体视觉特征。

根据我的经验,当提示词描述的是“可见结构”而不是“抽象概念”时,滑块的稳定性会明显提高。

Don't only tell the model what the image means. Tell the model what the image looks like.

不要只告诉模型这张图代表什么,而要告诉模型这张图看起来是什么样子。


v0.1:Usage

Depending on the angle you want,You must add one of the following angle prompts, otherwise the slider effect will be very weak:

  • three quarter view (≈45°)

  • side view

  • profile view (90°)

  • You can also experiment with other angle-related prompts, they might give interesting results.

Recommended Weights:

  • 0 weight: Mostly faces left (default)

  • Positive weight (+0.8 ~ +1.8): Strengthens left facing

  • Negative weight (-2.5 ~ -3.5): Controls right facing — Recommended to go straight to -3


v0.1使用方法

根据你想要的角度(这个版本没有能力丝滑调整角度)必须在提示词中加入以下角度词之一,否则滑块效果会很差:

  • three quarter view (约45度)

  • side view (侧面)

  • profile view (90度正侧面)

  • 其他角度提示词也可以自行尝试,可能会有奇效。

权重建议:

  • 0权重:大概率朝左

  • 正权重(+0.8 ~ +1.8):强化朝左

  • 负权重(-2.5 ~ -3.5):控制朝右,推荐直接拉到 -3 一步到位