🎁 Register on RunningHub & Get 1,000 RH Coins:
🎁 注册 RunningHub 领 1,000 RH 币：
BUNNY H3 High-Dynamic FL2VA / REF2VA One-Click Workflow :
https://www.runninghub.ai/zh-cn/post/2095927034992533505/?inviteCode=lk4qa7rh
RH Generation Speed / 运行效率：On RunningHub, generating a 1080P (0.9MP × 1.5) 10-second video takes ~12 minutes.
📌 Overview / 工作流简介
This workflow is specially optimized for MiniMax H3 high-dynamic video generation. It solves common issues such as noise, blur, and detail instability when applying secondary latent upscaling alongside BUNNY COMBAT LoRAs.
本工作流专为 MiniMax H3 高动态视频生成 进行深度优化，精准解决了配合 BUNNY 动作 LoRA 进行二次放大时容易出现的噪点、画面模糊和细节不稳定等痛点。
Output Capability / 画面表现：Supports direct output up to 1080P / 最高支持 1080P 直出。
12GB VRAM Optimization / 12GB 本地显存优化：
Recommended Stage 1 base resolution: 0.3MP with 1.5× Latent Upscaling.
Note: Max achievable duration depends on VRAM, aspect ratio, and the number of reference images.
本地 12GB 显存推荐设置：一采使用 0.3MP 基础分辨率 + 1.5 倍 Latent 放大。
提示：实际可生成时长会根据显存大小、画幅比例及参考图数量有所不同。
🚀 How to Use / 使用指南
Prompt & Duration / 提示词与时长：Enter your prompt and set the target video duration. / 首先填写提示词和生成时长。
Resolution & Upscale Scale / 分辨率与放大倍率：Set the Stage 1 base resolution and latent upscale scale (recommended starting value: 1.5×). / 设置一采基础分辨率与 Latent 放大倍率（推荐从 1.5 倍开始）。
Reference Images / 参考图设置：Load images directly when using FL2VA or REF2VA. For Text-to-Video (T2V), bypass or disconnect all reference image nodes. / 使用 FL2VA 或 REF2VA 时直接加载参考图；使用文生视频（T2V）时，请旁路（Bypass）或断开所有参考图节点。
Conditioning / 条件节点：Keep both Conditioning nodes set to AUTO. / 两路 Conditioning 节点保持 AUTO 默认设置即可。
LoRAs Fine-tuning / LoRA 调整：Replace or adjust the Stage 1 and Stage 2 LoRAs according to your subject. / 一采与二采的 LoRA 可根据生成题材自行更换或微调。
Stage 2 Enforcement / 二采必须开启：Stage 2 reconstructs the upscaled latent—do NOT disable it. / 二采负责重构放大后的 Latent 潜空间，切勿关闭。
🧩 Required Custom Nodes & Models / 必需依赖项
1. Custom Nodes / 必需第三方节点
MiniMax H3 Audio T8: GitHub Link
ComfyUI-KJNodes: GitHub Link
ComfyUI-Easy-Use (Optional: Used for post-generation VRAM cleanup only / 仅用于可选的生成后显存清理功能): GitHub Link
2. Required Latent Upscaler / 必需 Latent 放大模型
⚠️ Must use the exact model linked below / 必须使用指定版本：
Model Download / 模型下载：minimax_h3_latent_upscaler_3d_fp16.safetensors
Target Path / 存放路径：
ComfyUI/models/latent_upscale_models/SHA-256（节点实际强制 / runtime enforced）:
043E5A48E161610EF6C3EA974645220354D06FA618ABCA15F76D084812EB55C2加载器还会检查 322 个 tensor 的键、形状和类型，防止旧版、重打包、转换版或同名异构 checkpoint 生成坏 latent。/ The loader also validates 322 tensor keys, shapes and dtypes to reject old, repacked, converted, or structurally different same-name checkpoints before they can produce corrupted latents.
❤️ Credits & Open Source Spirit / 致谢与开源共建
Special Thanks to T8 for the underlying MiniMax H3 repair logic and two-pass sampling setup.
特别感谢 T8 老师 提供 MiniMax H3 底层修复与双采样逻辑支持。
Thanks to LBH-123-AI for training and sharing the MiniMax H3 3D latent upscaler model.
感谢 LBH-123-AI 提供 MiniMax H3 3D 潜空间放大模型。
✨ This is how an open-source community moves forward: creators build great workflows using my LoRAs, and I optimize those workflows based on my deep understanding of BUNNY LoRAs. Everyone is welcome to continue building upon this version. Every contribution helps our technology and community grow!
✨ 开源社区的魅力就在于不断的互助与进化：其他创作者基于我的 LoRA 制作出优秀的工作流，我再结合自己对 BUNNY LoRA 特性的理解对其进行针对性优化。非常欢迎大家在这个版本的基础上继续改进与创作，每一次微小的优化，都在推动技术与社区共同进步！
⚠️ Disclaimer / 注意事项
This download includes the ComfyUI workflow JSON file only. Third-party custom node source code and model weights are not included and must be installed separately.
本资源仅包含 ComfyUI 工作流文件，不包含第三方节点源码及模型权重，请按上述说明自行安装。
P.S. Hey, if I slap a VSR super-resolution node on the end to bump this to 2K, I should probably list this workflow for 800K Buzz too, right? Just checking the 2026 market rate. 😉
P.S. 顺便问一句：如果我再顺手挂个 VSR 超分节点给它升到 2K 分辨率，我是不是也该把这个工作流标价 800K Buzz 挂出来卖了？纯粹打听一下 2026 年的市场行情。😉