# Nick Wilde — MiniMax H3 Ref2V Character LoRA /英文说明
Trained on images of Nick Wilde from Zootopia, this LoRA helps improve his visual identity consistency in reference-to-video generation (Ref2V). It is trained specifically for Nick, rather than the film’s overall visual style or its other characters.
## Recommended settings
- Rank: 16. I recommend the v0.2 checkpoint at 2000 steps, which I use most often and have found to work well. Checkpoints at 500, 1000, and 3000 steps are also available.
- LoRA strength: I typically use 0.8–0.9 in my workflow; this is a useful starting range.
- Base model: Trained using MiniMax H3 Ref2VA INT8 ConvRot. The corresponding full-model file on Hugging Face is minimax_h3_ref2va_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_ref2va_int8_convrot.safetensors). BF16 can also be tried, but my results were less satisfactory than with INT8 ConvRot.
## Reference image and prompting
Please use the supplied Nick character reference image as a reference input. In my workflow, the LoRA alone is insufficient for consistent likeness: the image is also needed, and the prompt should explicitly identify it as the source of the character’s appearance.
My usual workflow is to render a scene in 3D software with a posed mannequin, then ask H3 to replace the mannequin with the fox from the reference image while preserving the scene, composition, and camera angle. This has worked well for me.
The training trigger is Nick_Wilde, but I have not needed to include it when using the character reference image. Write your video prompt normally and clearly explain the role of each reference image; no additional prompt-format changes are needed specifically for this LoRA.
## Training and limitations
The dataset contains 60 training images and 8 separate test images, with captions and matching character references. Most images show Nick wearing a pink patterned shirt, alongside some police uniforms and formal outfits. I recommend starting with the pink shirt. Outfit changes have not been specifically tested and are not guaranteed.
Training used still images, with no dedicated motion training; the main goal is character likeness. Feet are barely represented in the training set, so those details may be unreliable. Correct four-finger hands are also inconsistent. Generating multiple candidates can help you select a better result.
The source images have limited resolution. In my experience, setting the generation resolution too high can sometimes produce less satisfactory results. Start with a moderate resolution and test.
(translated by AI)
# 尼克·王尔德 — MiniMax H3 Ref2V 角色 LoRA /Chinese
本 LoRA 使用《疯狂动物城》中尼克的角色图片训练，用于增强参考图生成视频（Ref2V）时的形象一致性。它仅针对尼克这一角色，不是通用的电影风格 LoRA，也不适用于电影中的其他角色。
## 推荐设置
- Rank：16。推荐 v0.2 的 2000 步版本，这是我最常用、效果较满意的版本；另提供 500、1000、3000 步版本。
- LoRA 强度：我的工作流中通常使用 0.8–0.9，可作为起点。
- 底模：使用 MiniMax H3 Ref2VA INT8 ConvRot 训练。Hugging Face 上对应的完整模型文件为 minimax_h3_ref2va_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_ref2va_int8_convrot.safetensors) BF16版本也可以尝试，但效果不如 INT8 ConvRot。
## 参考图与提示词
请将随附的尼克角色参考图作为参考输入。在我的工作流中，仅靠 LoRA 不足以稳定还原角色，还需要这张参考图，并在提示词中明确说明它负责提供角色外观。
我的常用方式是：先在 3D 软件中渲染带有摆姿势人偶的场景，再要求 H3 将人偶替换为参考图中的狐狸，同时保留场景、构图和机位。这种用途下，获得的效果比较理想。
训练触发词为 Nick_Wilde，但在输入角色参考图后，实际使用时不必专门加入它。按照平常方式编写视频提示词，并说明各张参考图的分工即可，无须因加载 LoRA 而额外修改提示词格式。
## 训练与局限
数据集包含 60 张训练图片及 8 张独立测试图片，配有文字标注和对应角色参考图。服装以粉红色花纹衬衫为主，也包含警服、礼服等，建议优先尝试粉红色衬衫。换装能力尚未专门测试，不作保证。
训练素材为静态图片，未针对动作进行专门训练，主要目标是角色形象还原。脚部在训练集中几乎没有覆盖，相关细节可能不稳定；四指手部也不能稳定保证。可以通过多次生成挑选较好的结果。
训练原图分辨率有限。按我的使用经验，过高的生成分辨率有时反而不理想，建议先从适中的分辨率测试。