pixelvideo / video_pixeldit_claude_implementation_guide.md
linxin02's picture
Upload video_pixeldit_claude_implementation_guide.md
bcf0c00 verified
|
Raw
History Blame Contribute Delete
42 kB

Claude Code 工程任务书:将 PixelDiT 扩展为统一的 Video Pixel Diffusion

本文档用于直接交给 Claude Code 执行。目标是在现有 PixelDiT 文生图代码基础上,实现一个不依赖视频 VAE、直接在 RGB pixel space 中训练和采样的视频扩散模型,并同时支持 Text-to-Video(T2V)与 Image-to-Video(I2V)。

项目暂定名:VideoPixelDiT / V-PixelDiT

核心原则:尽可能继承现有 PixelDiT 的图像生成能力;避免完整时空 self-attention;将低频语义运动和高频像素细节的时间建模分层处理。


0. Claude 的工作方式与硬性要求

0.1 开始编码前必须先做仓库审计

先阅读并总结当前仓库,不要直接改代码。至少检查:

  • PixelDiT 主模型类;
  • patch-level DiT / MM-DiT block;
  • pixel-level PiT block;
  • timestep、RoPE、AdaLN、attention 实现;
  • flow matching scheduler;
  • REPA loss;
  • 训练入口、checkpoint loader、EMA、CFG、采样器;
  • dataset 和 multi-aspect-ratio pipeline;
  • 当前 tensor layout;
  • 当前 distributed training 方式。

如果代码基于官方 PixelDiT,重点对应:

pixdit_core/modules.py
pixdit_core/pixeldit_c2i.py
pixdit_core/pixeldit_t2i.py
t2i/train.py
t2i/inference.py
t2i/diffusion/
t2i/configs/

官方 PixDiT_T2I 的关键逻辑是:

  1. 将 RGB 图像按 patch_size × patch_sizeunfold
  2. patch-level MM-DiT 建模全局语义;
  3. 每个 patch 内保留所有 pixel token;
  4. PiT 将一个 patch 内的 pixel token 压缩成一个 attention token,跨空间 patch 做 attention;
  5. 再展开回 patch 内全部 pixel token并输出 RGB flow。

如果当前内部项目的类名和目录不同,先在新文档 video/REPO_MAPPING.md 中写出映射表,再实现。

0.2 不允许破坏原有 image PixelDiT

必须满足:

  • 原 T2I 训练和推理入口仍可运行;
  • 新功能放在独立模块中,优先复用而不是复制大段代码;
  • T=1 且所有 temporal gate 为零时,视频模型输出必须与原图像模型一致;
  • 加载图像 checkpoint 时,missing keys 只能来自新加的 temporal、frame/fps、known-frame 模块;
  • 不要改变原 checkpoint 中已有参数的名字,除非提供完整转换脚本和测试。

0.3 不要自动启动大规模训练

Claude 需要:

  • 完成代码;
  • 完成 unit test;
  • 完成 synthetic moving-shapes smoke test;
  • 完成一个小数据集 overfit test;
  • 提供大规模训练命令和配置。

不要未经明确命令直接下载 TB 级数据或启动多卡长训练。

0.4 每一个阶段都要有验收结果

每完成一个 milestone,输出:

1. 修改了哪些文件
2. 关键 tensor shape
3. 已运行的测试
4. 测试结果
5. 尚未完成或存在风险的部分

1. 项目目标

1.1 最终模型能力

统一模型应支持:

Text-to-Video

输入:

text prompt
video length T
height H
width W
fps
random seed

输出:

[B, T, 3, H, W]

Image-to-Video

输入:

first frame
optional text prompt
video length T
fps
motion strength

第一帧在整个采样过程中保持固定,模型生成后续帧。

保留 Image Generation

T=1 时,模型退化成原始 PixelDiT 文生图模型。


1.2 论文/研究主张

该项目不能只描述为“PixelDiT 加 temporal attention”。目标论点应是:

Pixel-space video diffusion 同时面对低频时空语义、物体运动和巨量高频像素噪声。统一的 full spatiotemporal token modeling 成本过高且优化困难。V-PixelDiT 将时间建模分解为:

  1. patch-level global temporal modeling:负责语义、物体身份、相机和主体运动;
  2. pixel-level compressed temporal modeling:负责纹理、边缘、小物体和局部细节的跨帧稳定性。

核心结构名可暂定:

Hierarchical Spatiotemporal Pixel Diffusion
Patch Temporal DiT + Compressed Temporal PiT

1.3 第一版非目标

第一版不要同时实现:

  • 长视频自回归;
  • 音频条件;
  • 3D causal VAE;
  • motion control / trajectory control;
  • camera pose control;
  • flow-guided deformable attention;
  • one-step distillation;
  • GAN loss;
  • 复杂物理 loss。

这些可以预留接口,但 MVP 先完成稳定的 8–16 帧 RGB flow matching。


2. 总体架构

输入视频:

x: [B, T, C, H, W]
t: [B]                 # diffusion / flow timestep,整个 clip 共用一个 t
text: [B, L_txt, D_txt]
frame_mask: [B, T]     # 1 表示有效帧,0 表示 padding
known_frame_mask: [B, T]  # I2V 时第一帧为 1
fps: [B]

默认:

channels: 3
frames: 16
height: 256
width: 256
patch_size: 16

H=W=256, P=16

L = (H/P) × (W/P) = 16 × 16 = 256 patches/frame
patch video tensor = [B, T, L, C_hidden]

禁止将全部 token 展开为 [B, T*L, C] 后做 full attention。


3. Patch-level 视频建模

3.1 空间 MM-DiT:逐帧复用原模型

先将视频展平到 frame batch:

x_bt = x.reshape(B * T, C, H, W)

逐帧 patchify:

patches = unfold(x_bt, kernel_size=P, stride=P)
patches: [B*T, L, C*P*P]
s = s_embedder(patches)
s: [B*T, L, hidden_size]

文本 embedding 先按 batch 计算一次:

y_shared: [B, L_txt, hidden_size]

对每个原有 MMDiTBlockT2I

s_bt = s_video.reshape(B*T, L, C_hidden)
y_bt = repeat(y_shared, T)  # [B*T, L_txt, C_hidden]

s_bt, y_bt = spatial_mmdit_block(
    s_bt,
    y_bt,
    condition_bt,
    spatial_rope,
    text_rope,
    attention_mask,
)

s_video = s_bt.reshape(B, T, L, C_hidden)
y_shared = y_bt.reshape(B, T, L_txt, C_hidden).mean(dim=1)

这样:

  • T=1 时与原始模型完全一致;
  • T>1 时每帧做相同的空间语义建模;
  • 文本 token 在每层后跨帧平均,避免每一帧形成不同的 text stream;
  • 不需要对 T×L 做 joint attention。

必须写一个 ablation 配置:

text_temporal_reduce: mean   # mean / first / frozen

默认 mean


3.2 TemporalDiTBlock

每隔若干个 spatial block 插入一个 patch temporal block。

默认:

temporal_block_interval: 2
patch_depth: 14
temporal_depth: 7

输入:

s_video: [B, T, L, C]

reshape:

z = s_video.permute(0, 2, 1, 3).reshape(B * L, T, C)

对每一个空间 patch 位置沿时间做 attention:

batch = B × L
sequence = T
channel = C

block 结构:

z = z + gate_attn * TemporalAttention(
    AdaLN(RMSNorm(z), temporal_condition),
    temporal_rope,
    frame_mask,
)

z = z + gate_mlp * MLP(
    AdaLN(RMSNorm(z), temporal_condition)
)

要求:

  • 使用 1D temporal RoPE;
  • 支持 variable length 和 frame_mask
  • T=1 时直接返回输入;
  • temporal AdaLN / gate 全部 zero-init;
  • 可以复用原 RotaryAttention,但需要实现 precompute_freqs_cis_1d
  • fps embedding 与 diffusion timestep embedding 相加后作为 temporal condition;
  • 不要给每一个 frame 独立采样 diffusion timestep,同一个 clip 必须共享同一个 timestep。

新增模块建议:

pixdit_core/video_modules.py
    precompute_freqs_cis_1d
    FPSConditioner
    FrameTypeEmbedder
    TemporalDiTBlock
    TemporalAttention

3.3 Temporal block 初始化

目标是加载 image checkpoint 后模型初始行为近似逐帧 PixelDiT。

必须:

nn.init.zeros_(temporal_block.adaLN_modulation[-1].weight)
nn.init.zeros_(temporal_block.adaLN_modulation[-1].bias)

或者单独:

self.temporal_gate = nn.Parameter(torch.zeros(...))

优先使用与原 DiT block 相同风格的 AdaLN-Zero。

写测试确认:

temporal module 输出 residual 初始为 0

4. Pixel-level 视频建模

4.1 保留原 PiT 的空间细化

原 PiT 的每帧逻辑保持不变:

RGB pixel
→ per-pixel linear embedding
→ 按 P×P 分组
→ [B*T*L, P², C_pixel]
→ 每个 patch 压缩
→ 跨空间 L 个 patch attention
→ 展开回 P² 个 pixel token

调用原 PiTBlock 时,将视频帧放到 batch:

x_pixels: [B*T*L, P2, C_pixel]
s_cond:   [B*T*L, C_hidden]

不要将所有视频 pixel token 做 full temporal attention。


4.2 新增 CompressedTemporalPiTBlock

在每个或每两个 spatial PiT block 后插入一个 temporal PiT block。

输入:

x_pixels: [B*T*L, P², C_pixel]
s_video:  [B, T, L, C_hidden]

第一步,将每个 patch 内全部 pixel token 压缩:

pixel_flat = x_pixels.view(B*T*L, P2*C_pixel)
pixel_comp = compress_to_temporal(pixel_flat)
pixel_comp: [B*T*L, C_temporal_pixel]

第二步,沿时间重排:

pixel_comp = pixel_comp.view(B, T, L, Ctp)
pixel_comp = pixel_comp.permute(0, 2, 1, 3)
pixel_comp = pixel_comp.reshape(B*L, T, Ctp)

第三步,做 local/global temporal attention:

temporal_out = temporal_attention(
    pixel_comp,
    temporal_rope,
    frame_mask,
)

第四步,展开回 patch 内全部 pixel token:

temporal_out = temporal_out.view(B, L, T, Ctp)
temporal_out = temporal_out.permute(0, 2, 1, 3)
temporal_out = temporal_out.reshape(B*T*L, Ctp)

pixel_residual = expand_from_temporal(temporal_out)
pixel_residual = pixel_residual.view(B*T*L, P2, C_pixel)

x_pixels = x_pixels + zero_initialized_gate * pixel_residual

可再接一个 per-pixel MLP,但第一版只需要一个 temporal attention residual。

建议默认:

pixel_temporal_depth: 2
pixel_temporal_hidden_size: 512
pixel_temporal_num_heads: 8
pixel_temporal_window: 5

对于 T≤16,patch-level temporal attention 可以全局;pixel-level 默认只看 5 帧局部窗口。


4.3 为什么 pixel temporal block 要先压缩

原始高分辨率 pixel token 的时间 attention 成本过高:

[B, T, H, W, C_pixel]

而压缩后:

[B, L, T, Ctp]

只需对每个空间 patch 的 T 个压缩 token 做 attention。

该模块负责:

  • 局部纹理不随帧随机漂移;
  • 人脸、衣服纹理、文字边缘、小物体保持;
  • 减少 temporal flicker;
  • 不承担大范围运动对应,大范围运动主要由 patch temporal block 建模。

5. 统一 T2V 与 I2V

5.1 T2V 模式

正常 flow matching:

known_frame_mask = zeros([B, T])
loss_mask = ones([B, T])

5.2 I2V 模式:masked video diffusion

不要第一版新增独立图像 encoder。直接把第一帧作为已知 clean frame。

训练时:

known_frame_mask[:, 0] = 1
x_t[:, 0] = x_0[:, 0]
loss_mask[:, 0] = 0

同时加入 frame type embedding:

unknown/generated frame embedding
known/condition frame embedding
padding frame embedding

将 frame type embedding 加到 patch token:

s_video = s_video + frame_type_embedding[:, :, None, :]

采样时每一步都 clamp:

x_current[:, 0] = condition_frame

必须确认:

  • 第一帧最终逐像素不变;
  • loss 不回传第一帧 flow prediction;
  • T2V 时 known-frame module 不改变原行为;
  • 支持将来扩展 keyframe mask,不要把逻辑写死为只能第一帧。

6. 模型文件组织

建议新增:

pixdit_core/
├── modules.py
├── pixeldit_t2i.py
├── video_modules.py
└── video_pixeldit_t2i.py

video/
├── README.md
├── train.py
├── inference.py
├── sample.py
├── configs/
│   ├── debug_8f_128.yaml
│   ├── temporal_only_8f_256.yaml
│   ├── joint_16f_256.yaml
│   ├── joint_16f_512.yaml
│   └── i2v_16f_256.yaml
├── datasets/
│   ├── video_dataset.py
│   ├── image_video_mixed_dataset.py
│   ├── transforms.py
│   ├── manifest.py
│   └── synthetic_moving_shapes.py
├── diffusion/
│   ├── video_flow.py
│   └── video_sampler.py
├── losses/
│   ├── flow_loss.py
│   ├── temporal_loss.py
│   └── repa_video.py
├── eval/
│   ├── generate_prompts.py
│   ├── run_vbench.py
│   ├── run_t2v_compbench.py
│   ├── run_fvd.py
│   ├── run_i2v_detail_benchmark.py
│   └── aggregate_results.py
├── tools/
│   ├── build_manifest.py
│   ├── inspect_dataset.py
│   ├── detect_shot_cuts.py
│   ├── compute_motion_score.py
│   └── convert_image_checkpoint.py
└── tests/
    ├── test_t1_equivalence.py
    ├── test_video_shapes.py
    ├── test_temporal_zero_init.py
    ├── test_known_frame_clamp.py
    ├── test_video_dataset.py
    └── test_overfit_tiny.py

如果当前 repo 已有通用 trainer,不要复制 trainer;通过 subclass、wrapper 或新增 task mode 接入。


7. Tensor shape 规范

项目中统一使用:

视频输入/输出: [B, T, C, H, W]
图像网络内部:   [B*T, C, H, W]
patch 视频:     [B, T, L, C_hidden]
patch attention:[B*T, L, C_hidden]
temporal attn:  [B*L, T, C_hidden]
pixel group:    [B*T*L, P², C_pixel]

禁止在不同文件中混用 [B,C,T,H,W][B,T,C,H,W]

所有入口都加 shape assertion,并在 debug 模式打印一次 tensor shape。


8. Flow Matching

8.1 基础目标

保持与现有 PixelDiT scheduler 一致。

默认:

x_\tau = (1-\tau)x_0 + \tau\epsilon
v^* = \epsilon - x_0
L_flow = ||v_\theta(x_\tau,\tau,c)-v^*||²

其中:

  • tau 对整个 clip 共享;
  • epsilon 每个 pixel 独立采样;
  • 第一版不引入 temporal-correlated noise;
  • I2V known frame 不加噪并从 loss mask 中移除。

预测 clean video:

x0_pred = x_t - tau * v_pred

需要 clamp 或仅为 auxiliary loss 使用,不要在主 flow loss 中截断。


8.2 Loss mask

总 mask:

valid_mask = frame_mask * (1 - known_frame_mask)

广播到:

[B, T, 1, 1, 1]

所有 loss 都必须正确处理 padding frame 和 known frame。


9. 辅助损失

MVP 先保证 flow loss 能稳定下降。辅助 loss 全部通过 config 开关控制。

9.1 Multi-scale flow loss

对预测 flow 和目标 flow做空间 average pooling:

L_ms_2 = mse(avg_pool_2(v_pred), avg_pool_2(v_target))
L_ms_4 = mse(avg_pool_4(v_pred), avg_pool_4(v_target))

默认:

multiscale_flow_weight: 0.25
multiscale_scales: [2, 4]

目的是让模型更早学习低频结构和运动,而不是只优化高频 pixel noise。


9.2 Temporal difference loss

只在中低噪声 timestep 使用,例如:

aux_mask_t = (tau < 0.7)

定义:

delta_pred = x0_pred[:, 1:] - x0_pred[:, :-1]
delta_gt   = x0[:, 1:] - x0[:, :-1]

L_delta = charbonnier(delta_pred - delta_gt)

默认:

temporal_delta_weight: 0.05
temporal_aux_max_t: 0.7

9.3 High-frequency temporal loss

先提取 Laplacian / Sobel 高频:

hf_pred = laplacian(x0_pred)
hf_gt = laplacian(x0)

再匹配相邻帧变化:

L_hf = L1(
    hf_pred[:, 1:] - hf_pred[:, :-1],
    hf_gt[:, 1:] - hf_gt[:, :-1],
)

默认:

temporal_hf_weight: 0.02

权重必须很小,避免锐化伪影。


9.4 REPA

复用现有 image REPA。

节省计算的默认策略:

  • 每个视频随机抽 1–4 帧;
  • 对抽中的帧计算 DINO feature;
  • 对应 patch-level token 做 REPA;
  • 不对所有 16 帧在线跑 DINO;
  • Stage 1 可使用 repa_weight=0.5
  • 后续高分辨率 stage 可关闭。

未来可添加 V-JEPA temporal representation alignment,但不要作为 MVP 阻塞项。


9.5 默认总损失

L =
L_flow
+ 0.25 L_multiscale
+ 0.05 L_delta
+ 0.02 L_hf
+ 0.5 L_REPA

注意:

  • 以上只是初始值;
  • temporal-only warmup 阶段可只用 L_flow + L_multiscale
  • 每个 loss 单独记录到 W&B;
  • 每个 loss 检查 finite;
  • 发生 NaN 时保存当前 batch metadata 和 timestep。

10. 数据集方案

10.1 推荐默认路线

开发与单元测试:Synthetic Moving Shapes

自己程序生成:

  • 1–4 个彩色几何物体;
  • 平移、旋转、缩放;
  • 前后遮挡;
  • 固定或移动相机背景;
  • 8–16 帧;
  • 64/128/256 分辨率;
  • 自动生成 caption 和 instance mask。

用途:

  • 测试 temporal block 是否能学习运动;
  • 测试 I2V clamping;
  • 测试小数据 overfit;
  • 测试 temporal loss;
  • 无下载和版权问题。

必须实现 video/datasets/synthetic_moving_shapes.py


Debug benchmark:UCF101

用途:

  • 低成本验证真实视频训练;
  • class name 转成简单 caption;
  • action motion 较明显;
  • 可报告 action-conditioned FVD;
  • 不能作为高质量 T2V 主训练数据。

caption 模板示例:

A person is {action_name}.
A video of a person performing {action_name}.

只用于 debug / sanity,不用于最终高质量模型。


主训练数据:OpenVid-1M

默认主数据集。

推荐使用顺序:

OpenVid subset 50K
→ OpenVid 200K–500K
→ OpenVid-1M
→ OpenVidHD high-quality subset

优点:

  • 提供 expressive captions;
  • 适合 T2V;
  • 包含高分辨率视频;
  • OpenVidHD 可用于 512 分辨率 fine-tuning。

默认 MVP 不需要完整下载 4.5 TB 的 OpenVidHD。先依据官方 mapping/metadata 只下载所需 subset。

建议本项目生成固定 split:

train: 98%
validation: 1%
held-out test: 1%

按 video id hash 划分,禁止同一 source video 的相邻 clip 跨 train/test。


可选规模预训练:Panda-70M 2M subset

仅在需要扩大运动和场景覆盖时使用。

优先使用:

Panda-70M 2M split
desirability == desirable
single continuous shot
no screen recording
no screen-in-screen

Panda-70M 2M 元数据对应约 2.4M clips,官方估计原视频存储规模约 1.6 TB。视频来自公开来源,下载可用性可能变化,使用时必须遵循原始视频许可和机构政策。

推荐 full-scale 顺序:

Panda-2M at 256px pretraining
→ OpenVid-1M quality refinement
→ OpenVidHD at 512px fine-tuning

如果算力和存储有限,完全跳过 Panda,只用 OpenVid。


Image/Video mixed data

联合微调 spatial block 后,加入现有 PixelDiT image dataset:

video_probability: 0.75
image_probability: 0.25

image sample 直接作为:

T=1 video

不要复制成静止的 16 帧视频,否则会让模型偏向无运动。

作用:

  • 保持单帧清晰度;
  • 防止 text-image alignment 退化;
  • 防止 spatial backbone catastrophic forgetting。

10.2 数据清洗标准

为每个视频构建 JSONL/Parquet manifest:

{
  "video_path": ".../clip.mp4",
  "video_id": "...",
  "source_id": "...",
  "caption": "...",
  "fps": 30.0,
  "duration": 5.2,
  "frame_count": 156,
  "width": 1920,
  "height": 1080,
  "shot_count": 1,
  "motion_score": 0.34,
  "aesthetic_score": 5.8,
  "has_audio": true,
  "split": "train"
}

过滤规则建议:

min_duration_sec: 2.0
max_duration_sec: 12.0
min_short_side_for_256: 360
min_short_side_for_512: 576
require_single_shot: true
min_motion_score: 0.02
max_motion_score: 0.95
min_caption_words: 3
max_caption_words: 100
reject_corrupt_decode: true
reject_extreme_aspect_ratio: true
max_aspect_ratio: 2.0

shot cut:

  • Panda 优先使用官方 shot boundary annotation;
  • OpenVid 可用 TransNetV2 离线检测;
  • MVP 也可以先用 frame histogram difference 作为轻量检查;
  • 不允许 clip 中间存在硬切镜头。

10.3 采帧策略

默认:

num_frames: 16
target_fps: 8
temporal_stride_choices: [1, 2, 3, 4]

流程:

  1. 根据视频原 fps 和目标时长选择起始位置;
  2. 随机选择 temporal stride;
  3. 连续采样 T 帧;
  4. 帧不足则优先重新选择 clip,不要循环播放;
  5. 必须 padding 时,复制最后一帧并将 frame_mask=0
  6. 所有空间增强对整个 clip 使用同一参数。

禁止:

  • 每帧独立 random crop;
  • 每帧独立 horizontal flip;
  • 每帧独立 color jitter;
  • caption 描述方向性动作时做 temporal reverse。

10.4 空间增强

对整个 clip 一致执行:

resize
random crop / center crop
horizontal flip
轻量 color jitter(可选)
normalize 到 PixelDiT 当前使用的范围

多 aspect ratio 训练在 256 阶段稳定后再打开。

第一版顺序:

fixed 256 square
→ multi-aspect 256
→ fixed/多比例 512

10.5 数据加载

优先:

  • decordPyAV
  • decode failure 可重试,但要有最大重试次数;
  • 每个 worker 设置独立 seed;
  • 支持本地 MP4 + manifest;
  • 后续支持 WebDataset/tar shard;
  • 不要在 dataloader 中在线跑 shot detection、RAFT、caption model。

需要 dataset inspection 脚本:

python video/tools/inspect_dataset.py \
  --manifest data/openvid_train.jsonl \
  --num_samples 100 \
  --output_dir outputs/dataset_preview

输出:

  • contact sheet;
  • caption;
  • fps/duration/resolution;
  • sampled frame indices;
  • motion score;
  • decode failure 统计。

11. 训练课程

Stage 0:Synthetic overfit

配置:

model_size: tiny
frames: 8
resolution: 128
dataset: synthetic_moving_shapes
samples: 128

目标:

  • 100–1000 steps 内明显过拟合;
  • 采样运动方向与 caption 一致;
  • 无 NaN;
  • I2V 第一帧精确保持;
  • patch temporal 和 pixel temporal 都有非零梯度。

此阶段不通过,不得进入真实视频训练。


Stage 1:Temporal-only warmup

加载现有 1.3B image PixelDiT checkpoint。

冻结:

text encoder
text projection
spatial MMDiT blocks
original spatial PiT blocks
RGB/pixel embedder
final output layer

训练:

patch temporal blocks
pixel temporal blocks
fps embedding
frame type embedding
known-frame embedding

建议配置:

frames: 8
resolution: 256
dataset: OpenVid 50K–200K
per_gpu_batch_size: 1
gradient_accumulation_steps: 8
lr: 1.0e-4
weight_decay: 0
bf16: true
gradient_checkpointing: true
repa_weight: 0
temporal_delta_weight: 0
temporal_hf_weight: 0

训练目标先只用:

flow loss
multi-scale flow loss

验收:

  • 单帧质量不明显低于原 PixelDiT;
  • 视频不再是完全独立帧;
  • temporal parameter gradient 正常;
  • frozen parameter gradient 为 None;
  • 8-frame sample 无严重闪烁。

Stage 2:Partial joint fine-tuning

解冻:

后 1/3 patch-level spatial blocks
全部 temporal blocks
全部 pixel blocks
final layer

image/video 混合训练。

建议:

frames: 16
resolution: 256
video_probability: 0.75
image_probability: 0.25
lr_new_temporal: 5.0e-5
lr_pretrained: 5.0e-6
repa_weight: 0.5
multiscale_flow_weight: 0.25
temporal_delta_weight: 0.05
temporal_hf_weight: 0.02
caption_dropout: 0.1

必须使用 parameter groups,为新模块和预训练模块设置不同 lr。


Stage 3:Full 256 training

数据:

OpenVid-1M
或 Panda-2M + OpenVid-1M

解冻全部 denoiser,但 text encoder 默认继续冻结。

建议:

frames: 16
resolution: 256
global_batch_size_target: 64
lr: 1.0e-5
warmup_steps: 2000
gradient_clip: 0.2
flow_shift: 3.0
weighting_scheme: logit_normal

保持与原 PixelDiT scheduler 尽量一致。


Stage 4:512 high-quality fine-tuning

数据:

OpenVidHD filtered subset

建议:

frames: 8 or 16
resolution: 512
per_gpu_batch_size: 1
gradient_accumulation_steps: 16
lr: 2.0e-6
repa_weight: 0

先从 8 帧 512 开始,再决定是否升到 16 帧。

优先训练:

temporal modules
pixel-level modules
最后若干 patch blocks

不建议第一轮就全参数 16×512。


Stage 5:I2V specialization

在 Stage 2 或 Stage 3 checkpoint 上:

task_mix:
  t2v: 0.5
  i2v: 0.5

I2V 数据直接使用 OpenVid:

first sampled frame = condition image
remaining frames = generation target
caption = provided caption

加入不同 motion strength 的 frame stride sampling。


12. 优化和显存

必须打开:

bf16
scaled_dot_product_attention / FlashAttention
gradient checkpointing
gradient accumulation
EMA(如果原项目已有)

优先 checkpoint:

  • patch temporal block;
  • spatial MM-DiT block;
  • pixel PiT block;
  • temporal PiT block。

避免保存大中间量:

  • REPA 只抽部分帧;
  • auxiliary x0 只在对应 loss 开启时构建;
  • validation sample 数量受 config 控制;
  • 不要每 step 保存 GIF/MP4。

可选后续优化:

FSDP
sequence parallel along time
context parallel
CPU offload
torch.compile

MVP 不要因 sequence parallel 阻塞。


13. 采样

13.1 Video flow sampler

复用当前 FlowDPMSolver 或 flow sampler,但 state shape 改为:

[B, T, C, H, W]

模型内部负责 flatten/reshape。

CFG:

  • unconditional text 对整个 clip 共用;
  • 不要每帧不同 CFG;
  • 支持 CFG interval;
  • 默认沿用 PixelDiT 的 cfg scale 和 flow shift,再单独调视频。

输出 MP4 时记录:

{
  "prompt": "...",
  "negative_prompt": "...",
  "seed": 123,
  "num_frames": 16,
  "fps": 8,
  "height": 256,
  "width": 256,
  "steps": 50,
  "cfg_scale": 3.0,
  "flow_shift": 3.0,
  "checkpoint": "...",
  "git_commit": "..."
}

13.2 I2V clamp

每个 solver update 后:

x = x * (1 - known_mask) + clean_condition * known_mask

最后输出第一帧必须与输入图逐像素一致,除非用户显式要求允许颜色处理。


14. 评测数据与协议

14.1 Validation set

从主训练数据中按 source video id 固定 hold out。

建议:

OpenVid validation: 2,000 clips
OpenVid held-out test: 2,000 clips

不能仅随机按 clip 路径划分,避免相同长视频切出的相邻片段泄漏。


14.2 VBench

T2V 主 benchmark。

运行标准 prompt suite,至少报告:

  • subject consistency;
  • background consistency;
  • temporal flickering;
  • motion smoothness;
  • dynamic degree;
  • aesthetic quality;
  • imaging quality;
  • object class;
  • multiple objects;
  • human action;
  • spatial relationship;
  • scene;
  • overall consistency。

保存:

generated videos
per-prompt results
per-dimension JSON
aggregate JSON
exact model/sampling config

不要只报告 total score;重点观察 pixel 模型相关的:

temporal flickering
subject consistency
imaging quality
motion smoothness

14.3 VBench-I2V / VBench++

I2V 模式完成后,使用 VBench-I2V。

重点:

  • condition consistency;
  • subject identity;
  • background consistency;
  • motion quality;
  • temporal flicker。

VBench-2.0 可作为后期补充,用于 human fidelity、physics、commonsense 等更高层评估,不应阻塞 MVP。


14.4 T2V-CompBench

使用官方 1400 prompts,报告:

  • consistent attribute binding;
  • dynamic attribute binding;
  • spatial relationships;
  • motion binding;
  • action binding;
  • object interactions;
  • generative numeracy。

这对验证手指、计数、小物体、交互关系很重要。


14.5 UCF101 FVD

使用 UCF101 test split 做传统分布评测。

要求:

  • 明确 FVD feature extractor;
  • 明确 generated sample 数;
  • 至少报告 2,048 samples;
  • 算力允许时额外报告 10,000 samples;
  • 所有模型使用相同分辨率、帧数、fps 和 preprocessing;
  • 不把不同代码库、不同 I3D 权重得到的 FVD 直接横向比较。

FVD 不能作为唯一指标。


14.6 Pixel-specific I2V Detail Benchmark

为了证明 pixel-space 相比 latent-space 的优势,构建一个可复现的小 benchmark。

数据来源:

DAVIS 2017 validation videos
+
OpenVid held-out high-resolution clips

筛选包含:

  • 人脸;
  • 手和手指;
  • 文字/标牌;
  • 细小物体;
  • 高频重复纹理;
  • 快速和缓慢运动;
  • 相机运动。

每个 clip:

first frame 作为 I2V condition
生成后续 15 帧

指标:

First-frame identity preservation

  • DINO feature similarity;
  • face identity similarity(仅检测到脸时);
  • foreground masked similarity(DAVIS mask)。

Temporal consistency

  • RAFT warp error;
  • tLPIPS;
  • DINO feature consistency;
  • foreground identity consistency。

High-frequency preservation

  • Laplacian energy consistency;
  • edge density consistency;
  • text region OCR consistency;
  • foreground texture feature consistency。

生成一个 HTML report:

condition frame
ground-truth clip
V-PixelDiT result
latent baseline result
metric table
failure notes

15. Baseline 与消融

15.1 必须实现的内部 baseline

B0:Image PixelDiT frame-independent

相同 prompt 和不同 noise 独立生成各帧,用于说明没有 temporal module 的问题。

B1:Spatial-only video wrapper

将视频帧作为 batch,但不使用 temporal block。

B2:Patch temporal only

只加入 patch-level temporal block。

B3:Patch + pixel temporal

完整模型。

B4:No auxiliary temporal loss

关闭 L_deltaL_hf

B5:No image-video mixed training

验证 image mixing 对单帧质量和 text alignment 的作用。


15.2 结构消融

temporal block interval: 1 / 2 / 4
patch temporal global vs local
pixel temporal window: 3 / 5 / global
pixel temporal hidden: 256 / 512 / 1152
text temporal reduce: mean / first / frozen
zero-init vs random-init
T=8 vs T=16

15.3 Pixel vs latent 对比

后期实现一个同数据、同采样 protocol 的 latent baseline。

优先保证公平:

  • 尽量相同 text encoder;
  • 相近参数量的 temporal transformer;
  • 相同训练 clips、steps、fps、resolution;
  • 记录视频 VAE reconstruction quality;
  • 区分“生成模型误差”和“VAE reconstruction ceiling”。

至少先测:

GT video
→ video VAE encode/decode
→ pixel/detail benchmark

这样可以直接量化 latent pipeline 在生成之前已经损失的文字、手指、边缘和小纹理。


16. Unit Tests

16.1 T=1 等价测试

加载同一 image checkpoint。

输入相同:

x
t
text embedding
mask

比较:

image_out = image_model(x, t, y)
video_out = video_model(x[:, None], t, y)[:, 0]

要求:

torch.testing.assert_close(
    image_out,
    video_out,
    atol=1e-5,
    rtol=1e-5,
)

如果 bf16 导致误差,测试使用 fp32。


16.2 Shape tests

覆盖:

B=1/2
T=1/8/16
H,W=128/256
square and non-square
padding frame mask
T2V and I2V

16.3 Zero-init test

初始化后:

temporal residual norm == 0 或非常接近 0

同时检查 temporal parameters 存在且 requires_grad=True


16.4 Checkpoint conversion test

从 image checkpoint 加载:

unexpected_keys == []
missing_keys 仅属于允许的新模块

写明确 allowlist,不要简单 strict=False 后忽略所有错误。


16.5 Known-frame test

I2V 采样的输出:

assert max_abs(output[:, 0] - condition_frame) < 1e-6

16.6 Tiny overfit

固定 8 个 synthetic clips,训练至:

flow loss 明显下降
sample 能复现颜色、物体和运动

将 before/after sample 保存进测试 artifact。


17. 配置示例

创建 video/configs/temporal_only_8f_256.yaml

model:
  type: VideoPixDiT_T2I
  image_checkpoint: /path/to/pixeldit_t2i.pth
  in_channels: 3

  patch_size: 16
  hidden_size: 1536
  num_groups: 24
  patch_depth: 14

  pixel_hidden_size: 16
  pixel_attn_hidden_size: 1152
  pixel_num_groups: 16
  pixel_depth: 2

  temporal:
    enabled: true
    block_interval: 2
    hidden_size: 1536
    num_heads: 24
    window_size: 0
    rope_theta: 10000.0
    zero_init: true

  pixel_temporal:
    enabled: true
    depth: 2
    hidden_size: 512
    num_heads: 8
    window_size: 5
    zero_init: true

  conditioning:
    text_encoder: gemma-2-2b-it
    text_encoder_frozen: true
    caption_dropout: 0.1
    use_fps_embedding: true
    use_frame_type_embedding: true
    max_frames: 32

data:
  type: VideoDataset
  manifests:
    - /path/to/openvid_train.jsonl
  num_frames: 8
  resolution: 256
  target_fps: 8
  temporal_stride_choices: [1, 2, 3, 4]
  random_crop: true
  horizontal_flip: true
  require_single_shot: true
  num_workers: 8

task:
  t2v_probability: 0.5
  i2v_probability: 0.5

scheduler:
  predict_flow_v: true
  noise_schedule: linear_flow
  flow_shift: 3.0
  weighting_scheme: logit_normal
  logit_mean: 0.0
  logit_std: 1.0

loss:
  flow_weight: 1.0
  multiscale_flow_weight: 0.25
  multiscale_scales: [2, 4]
  repa_weight: 0.0
  temporal_delta_weight: 0.0
  temporal_hf_weight: 0.0
  temporal_aux_max_t: 0.7

train:
  mode: temporal_only
  mixed_precision: bf16
  fp32_attention: false
  train_batch_size: 1
  gradient_accumulation_steps: 8
  gradient_checkpointing: true
  gradient_clip: 0.2
  lr_temporal: 1.0e-4
  lr_pretrained: 0.0
  weight_decay: 0.0
  warmup_steps: 1000
  max_steps: 50000
  save_steps: 2000
  validation_steps: 500
  seed: 1

validation:
  prompts_file: video/prompts/validation_prompts.txt
  num_frames: 8
  fps: 8
  sampling_steps: 50
  cfg_scale: 3.0
  fixed_seeds: [0, 1, 2, 3]

18. 验证 prompts

创建固定 prompt 文件,覆盖:

A woman waves her right hand while standing in a kitchen.
A close-up video of a person slowly opening and closing both hands, with five fingers visible on each hand.
A child picks up a small red toy from a wooden table.
A black dog runs from left to right across a grassy field.
A camera slowly pans around a parked blue car.
A glass falls from a table and shatters on the floor.
Two people pass a basketball to each other.
Three red apples roll across a white table.
A street sign displaying the words "PIXEL VIDEO" while the camera moves closer.
A close-up of a person's face turning from left to right.
A bird lands on a thin tree branch.
A striped shirt remains visually consistent while a person walks forward.
A small silver key rotates on a dark surface.
A city street at night with moving cars and stable neon signs.
A fixed camera records ocean waves moving toward the beach.

每次 validation 使用相同 prompt、seed、fps 和 sampling config。


19. 日志和可复现性

每次运行保存:

resolved config
git commit
git diff
package versions
GPU type/count
global batch size
data manifest hash
checkpoint source
random seed
sampling config

W&B 至少记录:

loss/flow
loss/multiscale
loss/repa
loss/temporal_delta
loss/temporal_hf
grad_norm/temporal
grad_norm/spatial
lr/temporal
lr/pretrained
data/decode_failure_rate
data/mean_motion_score
system/max_memory_allocated

validation 保存:

MP4
first/middle/last frame contact sheet
prompt
seed
checkpoint step

20. 常见失败与处理

问题 1:每帧清晰但闪烁

检查:

  • temporal gate 是否仍接近 0;
  • temporal parameter 是否有梯度;
  • frame order 是否正确;
  • crop augmentation 是否逐帧独立;
  • timestep 是否对每帧独立采样;
  • pixel temporal block 是否实际启用。

优先:

增加 temporal warmup steps
加入小权重 temporal delta loss
检查 pixel temporal reshape

问题 2:视频很一致但几乎不动

检查:

  • 数据中过多静态视频;
  • motion score filtering;
  • image/video mixing 比例过高;
  • temporal loss 权重过大;
  • I2V first frame 是否错误地复制到所有帧;
  • frame stride 是否太小。

问题 3:加载图像 checkpoint 后单帧质量下降

检查:

  • T=1 equivalence test;
  • text stream 跨帧 reduce;
  • temporal gate 是否真正 zero-init;
  • frame embedding 在 T2I/T=1 下是否非零;
  • 原 PixelDiT weight name 是否改变;
  • spatial block 是否被意外随机初始化。

问题 4:训练 loss spike / NaN

检查:

  • PiT 是否使用官方 post-modulation option;
  • attention 是否 fp32;
  • gradient clip;
  • auxiliary x0 loss 是否在高噪声 t 计算;
  • temporal AdaLN scale;
  • bf16 下 RMSNorm;
  • variable-length attention mask shape;
  • corrupt clip;
  • 空视频或全 padding batch。

发生 NaN 时自动保存:

batch manifest rows
timestep
input min/max/std
每个 loss
最近一次 finite checkpoint

问题 5:512 分辨率 OOM

按顺序处理:

  1. 16 帧降为 8 帧;
  2. activation checkpoint;
  3. REPA 关闭;
  4. pixel temporal hidden 减小;
  5. pixel temporal window 减小;
  6. 只训练 temporal + pixel blocks;
  7. FSDP;
  8. temporal sequence parallel。

不要首先删掉 pixel-level pathway,否则失去项目核心。


21. Milestones 与最终验收

M0:仓库审计

交付:

video/REPO_MAPPING.md
现有类与新类映射
checkpoint key 结构
tensor layout

M1:数据与 synthetic pipeline

交付:

SyntheticMovingShapesDataset
VideoDataset
manifest builder
dataset inspection report

M2:模型 forward

交付:

TemporalDiTBlock
CompressedTemporalPiTBlock
VideoPixDiT_T2I
T=1 equivalence test
shape tests

M3:训练与采样

交付:

video flow trainer
T2V sampler
I2V masked sampler
known-frame clamp
tiny overfit result

M4:真实视频 MVP

交付:

OpenVid subset manifest
8×256 temporal-only config
16×256 joint config
fixed validation samples

M5:评测

交付:

VBench wrapper
T2V-CompBench wrapper
FVD wrapper
metric aggregation
HTML comparison report

M6:高分辨率与论文消融

交付:

8/16×512 config
OpenVidHD subset pipeline
patch-only vs patch+pixel temporal
pixel vs latent reconstruction comparison

22. Claude 最终需要输出的内容

完成实现后,Claude 必须给出:

  1. 新增和修改的完整文件列表;
  2. 模型结构与每个关键 tensor shape;
  3. image checkpoint 的加载报告;
  4. 所有 unit test 输出;
  5. synthetic tiny overfit 的 loss 曲线和样例路径;
  6. 单卡 smoke training 命令;
  7. 8 卡正式训练命令;
  8. T2V 推理命令;
  9. I2V 推理命令;
  10. VBench / T2V-CompBench / FVD 评测命令;
  11. 当前未解决风险;
  12. 下一步最值得优先做的一个实验。

23. 建议命令格式

# Unit tests
pytest video/tests -q

# Synthetic smoke test
torchrun --nproc_per_node=1 video/train.py \
  --config video/configs/debug_8f_128.yaml

# Temporal-only real-video warmup
torchrun --nproc_per_node=8 video/train.py \
  --config video/configs/temporal_only_8f_256.yaml \
  --model.image_checkpoint=/path/to/pixeldit_t2i.pth \
  --data.manifests="[/path/to/openvid_train.jsonl]" \
  --train.work_dir=/path/to/output

# Joint 16-frame training
torchrun --nproc_per_node=8 video/train.py \
  --config video/configs/joint_16f_256.yaml \
  --model.load_from=/path/to/temporal_only_checkpoint.pth

# T2V sampling
python video/inference.py \
  --config video/configs/joint_16f_256.yaml \
  --checkpoint /path/to/checkpoint.pth \
  --prompt "A black dog runs from left to right across a grassy field." \
  --num_frames 16 \
  --fps 8 \
  --height 256 \
  --width 256 \
  --steps 50 \
  --cfg_scale 3.0 \
  --seed 0

# I2V sampling
python video/inference.py \
  --config video/configs/i2v_16f_256.yaml \
  --checkpoint /path/to/checkpoint.pth \
  --condition_image /path/to/first_frame.png \
  --prompt "The person slowly turns their head and smiles." \
  --num_frames 16 \
  --fps 8 \
  --seed 0

# VBench
python video/eval/run_vbench.py \
  --checkpoint /path/to/checkpoint.pth \
  --output_dir /path/to/vbench_results

# T2V-CompBench
python video/eval/run_t2v_compbench.py \
  --checkpoint /path/to/checkpoint.pth \
  --output_dir /path/to/t2v_compbench_results

# FVD
python video/eval/run_fvd.py \
  --generated_dir /path/to/generated_ucf101 \
  --real_dir /path/to/ucf101_test \
  --num_samples 2048

24. 官方参考资料


25. 最优先实现顺序

严格按以下顺序,不要一开始追求完整大模型:

T=1 数值等价
→ synthetic 8-frame overfit
→ patch temporal block
→ pixel temporal block
→ I2V known-frame clamp
→ OpenVid 8×256 temporal-only
→ 16×256 partial joint tuning
→ VBench 和 pixel-detail benchmark
→ 512 fine-tuning
→ latent baseline

最关键的技术验收不是“代码能 forward”,而是:

1. 继承图像 PixelDiT 后单帧能力不被破坏;
2. patch temporal block 学到主要运动与身份一致性;
3. compressed temporal PiT 在不显著增加计算的情况下改善细节闪烁;
4. pixel 模型在文字、手指、小物体、边缘和纹理一致性上超过公平的 latent baseline。