temp / FlowCache /FlowCache4MAGI-1-dev6-adaptive /README_MOTIONDETAIL.md
Cccccz's picture
Add files using upload-large-folder tool
153dd42 verified
|
Raw
History Blame Contribute Delete
2.34 kB

MotionDetailCache on MAGI-1 (dev4-detail)

基于 FlowCache4MAGI-1-dev3-motion 的扩展分支,在 MotionCache 的 motion 度量 之外,增加 spatial detail 度量,Phase 2 active token 由两者共同决定。

与 dev3 (MotionCache) 的差异

维度 dev3 MotionCache dev4 MotionDetailCache
Phase 2 权重 仅 motion 权重 W_m motion W_m + detail W_d 融合
Detail proxy latent 局部 k×k 空间方差
适用场景 运动区域优先计算 静止纹理/边缘/语义细节也优先保留

Detail 度量

对每个 latent 像素 (t, h, w)

  1. 通道聚合幅度:mag = mean(|x|, dim=C)[N, T, H, W]
  2. 局部空间方差:在 detail_window_size×detail_window_size 窗口内计算 Var(mag)
  3. Soft-mapping:帧内 min-max 归一化 → W_d ∈ [detail_alpha, 1]

高方差区域对应边缘、纹理、物体边界等 heterogeneous 邻域。

Motion + Detail 融合

W_combined = combine(W_motion, W_detail)
A[p] += W_combined[p] · Δ_chunk
active[p] = A[p] > τ

weight_combine_mode(默认 max):

模式 公式 特点
max max(W_m, W_d) 运动或细节任一显著即可加速累积(推荐)
product W_m × W_d 两者同时高才快速激活,更激进 reuse
blend (1-λ)W_m + λW_d 线性权衡,λ=detail_lambda

默认超参

继承 dev3 的 motion 参数,并新增:

参数 默认值 说明
detail_alpha 0.5 detail 权重下限
detail_window_size 3 局部方差窗口(奇数)
detail_lambda 0.5 blend 模式下的 detail 权重
weight_combine_mode max 融合方式

快速运行

cd FlowCache4MAGI-1-dev4-detail
conda activate magi
export CUDA_VISIBLE_DEVICES=0
export MASTER_ADDR=localhost
bash scripts/single_run/motiondetail_t2v.sh

代码结构

inference/pipeline/cache/
├── motioncache.py          # dev3 基类 MotionWiseCache
└── motiondetailcache.py    # dev4 MotionDetailCache

yaml_config/single_run/motiondetail_config.yaml
scripts/single_run/motiondetail_t2v.sh

其余 sparse forward、KV cache、integrate 逻辑与 dev3 完全一致。