Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
122
# Core runtime
torch>=2.0.0
torchvision
pydantic
# VLM / planner / critic
transformers
accelerate
sentencepiece
protobuf
safetensors
einops
qwen-vl-utils
# Video / image IO
imageio
imageio[ffmpeg]
pillow
# Ditto / DiffSynth runtime
modelscope
ftfy
pynvml
peft
# Data utilities
pandas
datasets
accelerate==1.10.1
aiohappyeyeballs==2.6.1
aiohttp==3.13.4
aiosignal==1.4.0
anaconda-anon-usage @ file:///croot/anaconda-anon-usage_1710965072196/work
annotated-doc==0.0.4
annotated-types==0.7.0
anyio==4.13.0
archspec @ file:///croot/archspec_1709217642129/work
asttokens @ file:///opt/conda/conda-bld/asttokens_1646925590279/work
astunparse==1.6.3
attrs @ file:///croot/attrs_1695717823297/work
av==17.0.0
beautifulsoup4 @ file:///croot/beautifulsoup4-split_1718029820055/work
blessed==1.20.0
boltons @ file:///work/ci_py311/boltons_1677685195580/work
Brotli @ file:///croot/brotli-split_1714483155106/work
certifi @ file:///home/task_176765898832781/croot/certifi_1767659141017/work/certifi
cffi @ file:///croot/cffi_1714483155441/work
chardet @ file:///work/ci_py311/chardet_1676830276092/work
charset-normalizer @ file:///tmp/build/80754af9/charset-normalizer_1630003229654/work
click==8.3.2
comm==0.2.2
conda @ file:///croot/conda_1736277811838/work
conda-build @ file:///croot/conda-build_1716991285767/work
conda-content-trust @ file:///croot/conda-content-trust_1714483159009/work
conda-libmamba-solver @ file:///croot/conda-libmamba-solver_1706733287605/work/src
conda-package-handling @ file:///croot/conda-package-handling_1718138267740/work
conda_index @ file:///croot/conda-index_1719338209492/work
conda_package_streaming @ file:///croot/conda-package-streaming_1718136078615/work
controlnet_aux==0.0.10
cryptography @ file:///croot/cryptography_1714660666131/work
cuda-bindings==13.2.0
cuda-pathfinder==1.5.0
cuda-toolkit==13.0.2
datasets==4.8.4
debugpy==1.8.5
decorator @ file:///opt/conda/conda-bld/decorator_1643638310831/work
DeepCache==0.1.1
diffusers==0.35.1
dill==0.4.1
distro @ file:///croot/distro_1714488253808/work
dnspython==2.6.1
einops==0.8.2
executing @ file:///opt/conda/conda-bld/executing_1646925071911/work
expecttest==0.2.1
filelock @ file:///croot/filelock_1700591183607/work
frozendict @ file:///croot/frozendict_1713194832637/work
frozenlist==1.8.0
fsspec==2024.6.1
ftfy==6.3.1
git-filter-repo==2.47.0
gmpy2 @ file:///work/ci_py311/gmpy2_1676839849213/work
gpustat==1.1.1
h11==0.16.0
hf-xet==1.4.3
httpcore==1.0.9
httpx==0.28.1
huggingface_hub==0.36.2
hypothesis==6.108.4
idna @ file:///croot/idna_1714398848350/work
ImageIO==2.37.3
imageio-ffmpeg==0.6.0
importlib_metadata==8.7.0
ipykernel==6.29.5
ipython @ file:///croot/ipython_1718287989724/work
jedi @ file:///croot/jedi_1721058342488/work
Jinja2 @ file:///croot/jinja2_1716993405101/work
jsonpatch @ file:///croot/jsonpatch_1714483231291/work
jsonpointer==2.1
jsonschema @ file:///croot/jsonschema_1699041609003/work
jsonschema-specifications @ file:///croot/jsonschema-specifications_1699032386549/work
End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Low-High New

1. 项目概述

本项目旨在探索一个面向 low-cost -> high-cost 视频生成/编辑的 agent 化框架。

核心思想不是把视频编辑视为一次性的 prompt-to-video 过程,而是将其改造成一个自适应闭环:

  1. 理解低成本输入视频与用户 prompt。
  2. 规划哪些内容需要增强、保留或重写。
  3. 调用视频编辑模型执行生成。
  4. 使用 critic 对结果进行评估。
  5. 根据评估结果进行局部修正或全局重规划。
  6. 将成功和失败经验写入 memory,供后续样本复用。

从长期目标来看,这个 story 与 SPIRAL 一类的闭环生成框架一致,但本项目面向的不是长时序动作世界模型,而是 low-cost / high-cost 视频增强与编辑。

2. 问题定义

传统视频编辑流程通常依赖:

  • 一段用户 prompt
  • 一个 low-cost 输入视频
  • 一次单轮编辑生成

但对于 low-cost -> high-cost 任务,仅靠 prompt 往往不够。系统还需要自动判断:

  • 视频里哪些区域是真正的 low-cost,应该重点增强
  • 哪些主体、结构和布局必须保持稳定
  • 不同区域应该增强到什么强度
  • 当前结果是否满足身份一致性、时序一致性、审美质量和物理真实感

因此,这个任务天然适合引入 agent、critic 和 memory,形成可迭代优化的闭环系统。

3. 目标 Story

本项目的目标系统应具备以下能力:

  • 能按编辑维度自适应规划,例如主体、脸部、服装、道具、背景、光照、氛围等
  • 能结合 Wan / Ditto 等视频编辑底座执行编辑
  • 能通过 critic 对 prompt 对齐、结构保持、身份一致、时序连贯、审美质量、artifact 等维度进行评估
  • 当计划基本正确但实现较弱时,进行局部 prompt refinement
  • 当当前计划本身不合理时,触发全局 replan
  • 能跨样本积累 scene archetype、失败模式、prompt recipe、风格偏好等记忆
  • 最终支持 RL / GRPO 类训练,将 planner/critic 的反馈逐步内化进生成模型本身

4. 当前仓库状态

当前仓库并不是一个空框架,而是已经具备一个可运行的闭环原型。

目前已经实现的模块包括:

  • Planner

    • 输入 prompt 和 low-cost 视频
    • 分析视频内容、选择风格族、生成候选策略并汇总成结构化编辑计划
    • 关键文件:
      • src/planner/core/video_planner.py
      • src/planner/core/pipeline.py
  • Edit Model Wrapper

    • 负责调用 DiffSynth/Wan 或 Ditto LoRA 变体执行视频编辑
    • 关键文件:
      • src/edit/video_editor.py
  • Critic

    • 对编辑结果进行评估
    • 输出结构化评分、失败标签和修正建议
    • 关键文件:
      • src/critic/evaluator.py
      • src/critic/reflector.py
  • Memory

    • 存储 run 记录、evaluation 结果、failure pattern 和 prompt recipe
    • 为后续规划与修正提供历史经验检索
    • 关键文件:
      • src/memory/store.py
      • src/memory/retriever.py
      • src/memory/evolver.py
  • Closed-Loop Runner

    • 串联执行 planner -> edit -> critic -> reflection -> memory
    • 当前支持三类结果分支:
      • accept
      • local_refine_prompt
      • global_replan
    • 关键文件:
      • src/loop/run_loop.py
      • scripts/run_planner_and_edit.py

5. 当前实现与 Agent Story 的对应关系

当前实现已经能映射到一个基础版的 agent 闭环:

  • PlanAgent

    • 由 planner 模块承担
    • 负责决定 low-cost 视频应该如何向 high-cost 转化
  • World Model / Edit Model

    • 当前由现有视频编辑底座实现
    • 接收编辑计划或 prompt,生成编辑后视频
  • CriticAgent

    • 负责对结果打分,并输出失败标签和文本反馈
    • 决定当前结果应该被接受、局部修正还是全局重规划
  • Memory

    • 负责存储显式经验
    • 当前更接近“可检索经验库”,还不是参数化或隐式记忆

6. 当前已经具备的能力

目前仓库已经支持:

  • 单样本级别的视频规划
  • 多种编辑 backend 变体调用
  • critic 结构化评分
  • 基于 reflection 的 prompt patch
  • 从局部修正退化到全局重规划
  • memory 随运行过程持续增长
  • planner / critic / memory / edit 的 smoke test 脚本

因此,当前阶段更准确的定位是:

Phase 0:agent 化 low/high 视频编辑原型

7. 与最终研究目标相比的主要缺口

尽管闭环已经搭起来了,但距离最终研究目标还有明显差距。

当前主要缺口包括:

  • planner 仍然是原型,还不是专门为 low/high edit 训练出来的规划器
  • 规划维度还不够明确,尤其缺少更系统的并行 edit dimension 设计
  • critic 还没有针对 low/high 偏好和视频编辑目标进行专门训练
  • memory 目前是显式、符号化的,还没有偏好条件化或参数化能力
  • 还没有接入真正的 RL / GRPO 训练流程
  • 数据构建、批量化处理、标注与训练管线还没有完整并入当前仓库
  • 缺少标准化、研究级的 benchmark 和评估协议

8. 推荐技术路线

8.1 Planner

planner 后续应该从“结构化 prompt 编译器”演化成“专门面向 low/high 任务的编辑规划器”。

推荐的规划维度包括:

  • 主体身份
  • 脸部与发型
  • 服装与材质丰富度
  • 前景道具
  • 背景与 world-building
  • 光照与氛围特效
  • 与运动相关的敏感区域

planner 需要回答的问题包括:

  • 哪些内容必须 preserve
  • 哪些内容适合轻量增强
  • 哪些内容应该强重写
  • transformation 应该挂载在哪些 visible carrier 上
  • 每个区域的增强强度应该是多少

关键设计判断:

  • 更推荐走 并行编辑维度规划
  • 不应过度依赖动作序列式串行规划

本项目的核心不是动作分解,而是面向区域和语义维度的自适应增强。

8.2 Critic

critic 不应只是一个通用打分器,而应该紧贴 low/high 编辑目标。

推荐重点评估维度:

  • prompt alignment
  • structure preservation
  • subject ID consistency
  • pose / geometry consistency
  • temporal coherence
  • world/background realization
  • transformation strength
  • aesthetics
  • artifact level / realism

critic 的输出应包括:

  • 总分
  • 分维度评分
  • failure tags
  • 文本化修正建议
  • local refineglobal replan 的决策

8.3 Memory

memory 建议分两步发展。

第一阶段:显式记忆

  • scene archetypes
  • 常见 failure patterns
  • prompt compilation recipes
  • style conversion priors
  • model behavior priors

第二阶段:偏好记忆

  • 不同用户/数据风格下的 high-cost 偏好
  • 对“什么算 high-cost”的可复用先验
  • 如果实验结果足够支持,再考虑做 latent / parameterized memory

短期建议是先把显式 memory 做扎实,不要一开始就直接上隐式记忆。

8.4 RL / GRPO

RL 不应该等所有模块都完善后才考虑。

更合理的路径是尽早验证以下问题:

  • critic 分数是否真的和更好的视觉结果相关
  • reflection 相比 one-shot baseline 是否确实有收益
  • grouped candidates 是否可以稳定排序
  • 相对偏好优化是否能带来可测提升

推荐最短路径:

  1. 先得到一个可靠的 critic
  2. 在小规模数据上验证 reward 是否有效
  3. 引入 grouped candidate comparison

9. 外部依赖与非 Git 目录

当前仓库有几类目录虽然在本地存在,但默认不会被 Git 跟踪
这是有意为之,不是目录缺失。

对应规则在:

当前被整体忽略的目录包括:

  • DiffSynth-Studio/
  • Ditto/
  • models/
  • datas/
  • out/
  • reference/

这意味着:

  • 你在本地能看到这些目录
  • git status / GitHub / 代码审查页面里通常看不到它们的内容
  • 如果没有额外说明,别人很容易误以为项目缺文件

9.1 DiffSynth-Studio

  • 路径:
  • 上游仓库:
    • https://github.com/modelscope/DiffSynth-Studio
  • 作用:
    • 本地 Wan / VACE 推理与训练依赖
    • 当前 edit-model 的真实训练入口优先走这里
  • 当前用途:
    • wan14b
    • wan14b+4steplora
    • base_sft
    • planner_conditioned

它属于外部代码依赖,不是本仓库原生源码,因此默认不纳入 Git。

9.2 Ditto

  • 路径:
  • 上游仓库:
    • https://github.com/ezioby/Ditto
  • 作用:
    • 参考 Ditto 原始训练/推理实现
    • 某些编辑 variant 和训练兼容逻辑仍会复用这里的代码
  • 当前用途:
    • wan14b+dittolora
    • wan14b+dittolora+4steplora
    • Ditto 论文式 low/high 编辑参考实现

它也是外部代码依赖,因此默认不纳入 Git。

9.3 models

  • 路径:
  • 作用:
    • 存放本地模型权重、LoRA、tokenizer 和训练产物挂载

典型内容包括:

  • Wan2.1-VACE-14B
  • Wan2.1-Distill-Loras
  • Ditto_models
  • edit-model/base_sft
  • edit-model/planner_conditioned
  • critic/evaluator
  • critic/reflector
  • critic/pairwise

这个目录通常非常大,而且常包含:

  • 基座模型
  • LoRA checkpoint
  • 本地训练输出软链接

因此默认被 .gitignore 排除。

9.4 reference

  • 路径:
  • 作用:
    • 存放论文、笔记、对话记录、参考实现、临时研究材料

例如:

  • SPIRAL 论文
  • 项目讨论笔记
  • 风格/规划/critic 设计参考材料

这些内容通常不是项目必须的可执行源码,而且可能持续变动、体积较大或带版权限制,所以默认不纳入 Git。

9.5 datas

  • 路径:
  • 作用:
    • 存放 Ditto-1M 原始数据、pair 数据、训练集导出结果

例如:

  • datas/Ditto-1M
  • datas/ditto
  • datas/ditto_pairs.jsonl

这是数据目录,不应作为代码仓库内容提交。

9.6 out

  • 路径:
  • 作用:
    • 存放所有实验输出

包括:

  • planner 单测输出
  • edit-model 出片结果
  • critic 测试结果
  • 训练日志
  • checkpoint
  • 对比实验视频

这是典型实验工作目录,也不应进入 Git。

9.7 如果别人只拿到 Git 仓库,会缺什么

如果别人只 git clone 本仓库,而没有你本地这些忽略目录,那么默认会缺:

  • DiffSynth-Studio/
  • Ditto/
  • models/
  • datas/
  • reference/

也就是说,代码框架本身能看到,但:

  • 不能直接本地训练
  • 不能直接本地推理
  • 不能直接复现实验结果

必须额外补:

  • 外部代码依赖
  • 本地模型权重
  • Ditto 数据
  • 参考材料

9.8 推荐理解方式

可以把当前仓库分成两层:

  • Git 管理层

    • src/
    • scripts/
    • configs/
    • README.md
  • 本地工作区层

    • DiffSynth-Studio/
    • Ditto/
    • models/
    • datas/
    • out/
    • reference/

前者是代码与流程定义。
后者是实际运行、训练和实验所需的本地资源。 4. 在 reward 足够可信后,再引入 RL / GRPO

9. 分阶段项目计划

Phase 0:原型基线

目标:

  • 跑通 planner -> edit -> critic -> memory 闭环

当前状态:

  • 当前仓库已经基本完成

产出:

  • 单样本 smoke test
  • planner / critic / edit / memory 的中间产物

Phase 1:稳定闭环原型

目标:

  • 让当前原型稳定、可复现、便于后续扩展

主要任务:

  • 统一 planner / critic / memory 的输入输出 schema
  • 明确支持的 model variant 和运行方式
  • 在多个代表性样本上验证全链路可用性
  • 建立 baseline 对比:
    • 无 agent 的 direct edit
    • 只有 planner 的 edit
    • 完整 planner + critic + reflection 闭环

交付物:

  • 可稳定复现的内部基线系统

Phase 2:形式化 low/high 规划维度

目标:

  • 把 planner 从粗粒度 prompt 分解,升级为明确的 low/high 编辑维度规划器

主要任务:

  • 定义 region / dimension taxonomy
  • 支持主体 / 道具 / 背景 / 光照等并行规划
  • 明确 preserve / rewrite / light-touch target 的定义
  • 改进 planner prompt 和输出 schema

交付物:

  • 与 low/high 编辑目标直接对齐的 planner 输出

Phase 3:构建 low/high 对齐的 Critic

目标:

  • 让 critic 真正对 low/high 编辑质量负责

主要任务:

  • 定义 low/high 任务下的评分标准
  • 设计 scalar 或 pairwise 的评价目标
  • 加入 ID、一致性、时序和审美等维度
  • 校准 accept / refine / replan 阈值

交付物:

  • 更接近人类偏好的 critic 输出

Phase 4:数据与标注管线

目标:

  • 为训练和评估提供可扩展的数据基础

主要任务:

  • 做批量化数据处理与清洗
  • 整理 low/high 视频对或可编辑视频素材
  • 构建 prompt 标注和 edit-dimension 标注
  • 准备 planner 监督数据和 critic 偏好数据

交付物:

  • 可用于训练和评估的数据资产

Phase 5:训练 Planner 与 Critic

目标:

  • 用可训练模块替换掉当前主要依赖 prompt engineering 的部分

主要任务:

  • planner 的 SFT
  • critic 的 SFT
  • 在适合时加入 pairwise preference modeling
  • 验证不同场景、不同风格下的泛化能力

交付物:

  • 可训练、可迭代的 planner / critic 模块

Phase 6:升级 Memory

目标:

  • 让 memory 对后续样本的规划和修正产生实质性帮助

主要任务:

  • 提升 memory retrieval 质量
  • 将 memory 更深入地接入失败恢复流程
  • 引入 preference-aware memory
  • 测试 memory 是否能提升重复场景、重复风格下的表现

交付物:

  • 能带来可测收益的经验复用机制

Phase 7:验证 RL / GRPO

目标:

  • 验证 planner/critic 的反馈是否能真正内化进生成模型

主要任务:

  • 生成 grouped candidates
  • 计算组内相对 reward
  • 使用 GRPO 类目标优化 edit model 或 adapter
  • 与 one-shot baseline、reflection-only baseline 进行对比

交付物:

  • 对 RL 是否有效的实验性结论

10. 近期 To-Do

短期内建议优先推进以下事项:

  1. 在多个 smoke sample 上跑通现有全链路。
  2. 固定 planner / critic / memory 的 schema,避免接口频繁变化。
  3. 更明确地定义 low/high 的规划维度。
  4. 重构 critic 指标,补上 ID、姿态、时序和审美维度。
  5. 把当前原型接到已有数据与批量化处理流程上。
  6. 开始小规模标注,为 planner 监督和 critic 训练做准备。
  7. 在引入 RL 之前,先验证 reflection 是否显著优于 one-shot baseline。

11. 仓库入口

常用脚本:

  • scripts/run_planner_and_edit.py
  • scripts/test_video_planner.py
  • scripts/test_edit_model.py
  • scripts/test_critic.py
  • scripts/test_memory.py

核心源码目录:

  • src/planner/
  • src/edit/
  • src/critic/
  • src/memory/
  • src/loop/

配置文件:

  • configs/planner_qwen.json
  • configs/base_skills.json

12. 环境安装与测试命令

12.1 安装依赖

建议在项目根目录执行:

pip install -r requirements.txt

这份依赖中已经包含本地 planner 所需的 qwen-vl-utils,否则运行 local_qwen 路径时会报错。

如果后续需要单独安装本地 DiffSynth 包,也可以执行:

pip install -e DiffSynth-Studio

12.2 检查可用编辑变体

在测试 edit model 之前,建议先检查当前本地可运行的 variant:

python scripts/test_edit_model.py --list-variants

12.3 测试 Ditto + Wan 版本

当前仓库中,Ditto + Wan 对应的 variant 名称为:

  • wan14b+dittolora
  • wan14b+dittolora+4steplora

示例命令如下。

测试 Ditto + Wan

python scripts/test_edit_model.py \
  --variant wan14b+dittolora \
  --low-video smoke-data/1_low-cost.mp4 \
  --output-video out/test_ditto_wan/sample1_edited.mp4 \
  --prompt-file smoke-data/prompt1.txt \
  --device cuda

测试 Ditto + Wan + 4step LoRA

python scripts/test_edit_model.py \
  --variant wan14b+dittolora+4steplora \
  --low-video smoke-data/1_low-cost.mp4 \
  --output-video out/test_ditto_wan/sample1_edited_4step.mp4 \
  --prompt-file smoke-data/prompt1.txt \
  --device cuda

12.4 从 planner 输出继续测试 edit model

如果已经先跑过 planner,也可以直接使用 planner 输出的 normalized_prompt

python scripts/test_edit_model.py \
  --variant wan14b+dittolora \
  --low-video smoke-data/1_low-cost.mp4 \
  --output-video out/test_ditto_wan/from_planner.mp4 \
  --planner-response out/sample_1/planner.response.json \
  --device cuda

12.5 运行完整闭环

完整闭环入口如下:

python scripts/run_planner_and_edit.py \
  --sample-id 1 \
  --prompt-file smoke-data/prompt1.txt \
  --low-video smoke-data/1_low-cost.mp4 \
  --output-dir out/full_loop_sample1 \
  --variant wan14b+dittolora \
  --max-attempts 3

如果 --output-dir 已经存在,脚本会自动在该目录下创建一个新的时间戳子目录,例如:

out/full_loop_sample1/rerun_20260330_120501

这样重复运行不会覆盖旧的 attempt 结果。

scripts/run_planner_and_edit.py 现在默认会开启两项 GPU 常驻优化:

  • LOW_HIGH_AGENT_KEEP_VLM_ON_GPU=1
  • LOW_HIGH_AGENT_KEEP_EDIT_ON_GPU=1

这意味着在同一个 Python 进程里:

  • 本地 Qwen VLM 会优先常驻 GPU,减少重复加载时间
  • Ditto / DiffSynth 的 edit pipeline 会优先常驻 GPU,减少重复初始化和 LoRA 重挂载时间

如果想显式指定,也可以这样运行:

LOW_HIGH_AGENT_KEEP_VLM_ON_GPU=1 \
LOW_HIGH_AGENT_KEEP_EDIT_ON_GPU=1 \
python scripts/run_planner_and_edit.py

如果显存不够,想回到更保守的 offload 模式,可以关闭:

LOW_HIGH_AGENT_KEEP_VLM_ON_GPU=0 \
LOW_HIGH_AGENT_KEEP_EDIT_ON_GPU=0 \
python scripts/run_planner_and_edit.py

当前默认 memory 文件位置为:

out/memory/global_memory.json

如果不手动覆盖 --memory-path,闭环运行时会默认读写这个文件。

12.6 从 memory 导出可晋升 custom skills

如果希望从 memory 中筛选高质量 custom skills,并导出成中间 JSON 文件,可以执行:

python scripts/export_promotable_skills.py

默认行为:

  • out/memory/global_memory.json 读取 memory
  • memory.custom_skill_records 中筛选可晋升 skill
  • 导出到 out/promoted_skills/

默认筛选规则是保守版:

  • accepted == true
  • overall_score >= 0.78
  • count >= 1
  • skill_payload 必填字段完整

例如,使用更严格的筛选条件:

python scripts/export_promotable_skills.py --min-count 2 --min-score 0.82

12.7 将导出的 promoted skill JSON 转成正式 skill 文件

如果确认导出的 skill 值得沉淀到正式 skill 库中,可以执行:

python scripts/promote_skill_json_to_template.py

默认行为:

  • out/promoted_skills/ 读取导出的 JSON
  • 转换为 src/skills/templates/<style_family>/<skill_id>.py

只晋升某个 style family:

python scripts/promote_skill_json_to_template.py --style-family sci_fi

如果需要覆盖已有同名 skill 文件:

python scripts/promote_skill_json_to_template.py --overwrite

12.8 导出 skill execution summaries

如果希望离线查看哪些 skill / carrier 在某个 style_family + archetype 下更稳定有效,可以执行:

python scripts/export_skill_execution_summaries.py

默认行为:

  • out/memory/global_memory.json 读取 memory
  • memory.skill_execution_summaries 读取聚合结果
  • 导出到 out/skill_execution_summaries/

例如,只看 sci_fi 下的执行统计:

python scripts/export_skill_execution_summaries.py --style-family sci_fi

例如,只看某个 archetype:

python scripts/export_skill_execution_summaries.py --style-family sci_fi --archetype-id single_person_indoor_closeup

12.9 常用脚本汇总

规划器:

  • python scripts/test_video_planner.py

编辑模型:

  • python scripts/test_edit_model.py --list-variants
  • python scripts/test_edit_model.py --variant wan14b+dittolora --low-video smoke-data/1_low-cost.mp4 --output-video out/test_ditto_wan/sample1_edited.mp4 --prompt-file smoke-data/prompt1.txt --device cuda

Critic:

  • python scripts/test_critic.py

完整闭环:

  • python scripts/run_planner_and_edit.py

memory 分析与回放:

  • python scripts/test_memory.py

从 memory 导出可晋升 skill:

  • python scripts/export_promotable_skills.py

导出 skill execution summaries:

  • python scripts/export_skill_execution_summaries.py

将导出的 promoted skill 落入 skill 模板库:

  • python scripts/promote_skill_json_to_template.py

13. Reference 资料目录

reference/ 存放论文、外部项目快照和 benchmark 参考资料,不是运行时依赖。为了避免把大体积参考材料直接放进主仓库,它被拆到一个 private Hugging Face dataset repo,并在主仓库中记录为 submodule:

https://huggingface.co/datasets/Ouzhang/low-high-reference

首次拉取主仓库后,如果需要本地展开 reference/,先准备有访问权限的 Hugging Face token:

export HF_TOKEN='your_hf_token'

推荐方式是按 submodule 拉取:

cd /path/to/low-high-new
GIT_LFS_SKIP_SMUDGE=1 git -c http.extraHeader="Authorization: Bearer ${HF_TOKEN}" \
  submodule update --init --depth 1 reference
git -C reference -c http.extraHeader="Authorization: Bearer ${HF_TOKEN}" lfs pull

如果当前机器的 Git 访问 Hugging Face 不稳定,也可以用 HF CLI 直接下载内容到同名目录:

cd /path/to/low-high-new
rm -rf reference
HF_TOKEN="${HF_TOKEN}" hf download Ouzhang/low-high-reference \
  --repo-type dataset \
  --local-dir reference

注意:

  • reference/ 只用于研究参考和资料查阅,训练、推理和评测脚本不依赖它。
  • 这个 repo 是 private 的;没有 HF token 或权限时,主仓库仍然可以正常使用,只是 reference/ 不会展开。
  • 不要把 token 写进 .gitmodules、README 或任何提交文件里。

14. 当前阶段判断

当前仓库更适合被视为一个:

面向 low/high 视频编辑的 agent 原型系统

而不是一个已经完备的训练框架。

因此,现阶段最合理的推进顺序是:

  • 先稳定闭环
  • 再形式化 planner 和 critic 的目标
  • 再接数据与训练
  • 最后验证 RL / GRPO

这样可以避免在目标尚未定义清楚时,就提前投入大规模训练。

15. 总结

本项目的目标是构建一个自适应、agent 驱动的 low-cost 到 high-cost 视频生成/编辑系统。

它的核心贡献方向,是将一次性的视频编辑流程改造成一个包含以下模块的闭环:

  • planning
  • execution
  • critique
  • reflection
  • memory

当前仓库已经具备这一思路的有效原型。下一阶段的重点,是把原型逐步推进成一个可训练、可扩展、可验证的研究系统。

Downloads last month
6