ComfyUI YoYo H3 TST
Training-free Temporal State Transport (TST) temporal attention correction for MiniMax H3 video generation in ComfyUI.
ComfyUI-YoYo-H3-TST
Training-free temporal attention homeostasis correction (YoYo H3 Temporal State Transport (TST)) for MiniMax H3 video generation in ComfyUI.
Node type: YoYo_H3TemporalStateTransport · Menu category: YoYo/MiniMax H3
Overview
Based on the paper Temporal State Transport in Video Generation: Diagnosing and Correcting Spectral Imbalance (Luyao Tang et al.) and its official reference repository lytang63/temporal-state-transport (MIT License), this plugin provides an independent ComfyUI implementation adapted for MiniMax H3's packed attention layout:
- Per-head Spectral Tension diagnosis: Pools the video segment's post-RoPE
q/kper frame to build the frame-level transport operator $A$, and computes Spectral Tension $T = H_{\text{row}} - H_{\text{vN}}$ (row entropy minus von Neumann entropy of the normalized Gram matrix).- $T > 0$: over-mixed transport (physically implausible motion / structural collapse)
- $T < 0$: fragmented transport (flickering details / identity or outfit drift)
- Homeostatic query temperature: Scales only the video-row queries by $\gamma = \exp(T \cdot \tau_{\text{eff}})$ to sharpen over-mixed heads and soften fragmented ones, while leaving balanced heads nearly untouched.
- Cosine schedules: $\tau_{\text{eff}}$ is scheduled with cosine layer and step weights (stronger in deeper layers and earlier denoising steps).
- Non-intrusive integration: Chains cleanly with any existing
optimized_attention_override(such as Sol-Attn) rather than overwriting it, and leaves text, keyframe/reference, and audio segments untouched.
Installation
Via Comfy Registry / ComfyUI-Manager
comfy node install comfyui-yoyo-h3-tst
Or search for ComfyUI YoYo H3 TST in ComfyUI-Manager.
Manual
Place the ComfyUI-YoYo-H3-TST folder into ComfyUI/custom_nodes/ and restart ComfyUI.
中文说明
一个节点:YoYo H3 Temporal State Transport (TST)
(节点类型 YoYo_H3TemporalStateTransport,菜单位置:YoYo/MiniMax H3)
原理与功能
基于论文 Temporal State Transport in Video Generation: Diagnosing and Correcting Spectral Imbalance 及其官方开源仓库 lytang63/temporal-state-transport(MIT License),按公开数学公式针对 MiniMax H3 打包注意力结构独立实现的时间注意力稳态校正节点,专治 H3 视频中的细节闪烁/变形、人物与服装身份漂移、运动不合理等问题:
- 在每个注意力头上,将视频段的
q/k按帧池化,构造帧级传输算子 $A$ - 通过谱张力 $T = H_{\text{row}} - H_{\text{vN}}$(行熵 − Gram 矩阵特征值的冯诺依曼熵)逐头诊断:
- $T > 0$:过混合(物理崩坏方向)
- $T < 0$:过碎片(细节/身份漂移方向)
- 以逐头 query 温度 $\gamma = \exp(T \cdot \tau_{\text{eff}})$ 仅缩放视频行的
q:过混合的头锐化、过碎片的头柔化,已平衡的头几乎不动 - $\tau_{\text{eff}}$ 采用两条余弦调度:深层网络更强、早期采样步更强
安装与用法
- 通过 ComfyUI-Manager 安装,或把
ComfyUI-YoYo-H3-TST文件夹放进ComfyUI/custom_nodes/并重启 ComfyUI - 将节点串进模型链任意位置(
MODEL进 →MODEL出),无需更换采样器
参数与特性:
tau:校正强度,默认0.2(论文设定,范围0.0–1.0);0.5通常偏强;tau = 0关闭校正但保留诊断日志log_diagnostics:是否在控制台打印逐次前向的[YoYo TST]诊断信息- 与已有的
optimized_attention_override(如 Sol-Attn)链式组合而非覆盖:本节点先做q缩放,再交给先前的 override - 仅处理打包序列末尾的视频行;文本、关键帧/参考、音频行原样通过
- 非 MiniMax H3 模型会记录警告并原样返回;若序列布局不匹配则静默跳过,不会中断采样
验收方法
开启 log_diagnostics 后运行一次生成,控制台每次模型 forward 会输出一行:
[YoYo TST] frames=37 step=0/5 heads corrected=47 mean|T|=... mean|gamma-1|=...
- 看不到这行日志通常表示未生效(先检查节点是否接在实际使用的模型链上)
- 固定同 seed、同提示词,对比
tau = 0与tau = 0.2的闪烁与身份漂移改善效果 - 建议先在全尺寸 H3 工作流上验证,再考虑 Flash 等变体流程(分解注意力布局尚未验证)
References / 参考文献
- Paper: Temporal State Transport in Video Generation: Diagnosing and Correcting Spectral Imbalance
- Official Implementation: lytang63/temporal-state-transport (MIT License)
@article{tang2026temporal,
title={Temporal State Transport in Video Generation: Diagnosing and Correcting Spectral Imbalance},
author={Tang, Luyao and Luo, Bingjun and Yi, Dong and Guo, Jialin and Xi, Haoning and Chen, Cheng and Yu, Yizhou and Chen, Chaoqi},
journal={arXiv:2609.08505},
year={2026}
}