# 环境重建 **推荐直接用一键脚本**:`bash environment/create_envs.sh`(可 `[opg|sam3|motion]` 单建; 结束时打印该填进 `.env` 的解释器路径)。本文档是它背后的版本表与逐条手动命令—— 脚本哪一步失败,就照下面手动补装哪一步。 OVOW 用 3 个互不兼容的 conda 环境(torch/CUDA 强耦合,**不能纯 pip 复现**,需对齐 CUDA)。 | env | python | torch | 用途 | 依赖清单 | |---|---|---|---|---| | `ovow-opg` | 3.11 | 2.4.1+cu121 | 主重建 / scale / 位姿 | `requirements-opg.txt` | | `ovow-sam3` | 3.12 | 2.7.0+cu118 | SAM3 分割 | `requirements-sam3.txt` | | `ovow-motion` | 3.11 | 2.4.1+cu124 | Motion324 + Hunyuan deform | `requirements-motion.txt` | > **三个环境的 torch 各绑一个 CUDA 版本(cu121 / cu118 / cu124)**——这是 pip wheel 自带的 > CUDA runtime,互不冲突,也不要求系统装三套 CUDA Toolkit。系统侧只需: > ① NVIDIA 驱动 ≥550(能跑 cu124 wheel 即可全兼容);② 一套 CUDA Toolkit(建议 12.1+, > 含 `nvcc`)用于现场编译扩展——opg 的一批(见下表)+ motion 的 custom_rasterizer; > sam3 无需编译,纯 pip 即装。 > `requirements-sam3.txt` / `requirements-motion.txt` 已内置对应的 > `--extra-index-url`(PyTorch 官方 wheel 源),直接 `pip install -r` 即可。 重建(以 opg 为例,**顺序重要**): ```bash conda create -n ovow-opg python=3.11 -y && conda activate ovow-opg # 1) 先按你的 CUDA 装 torch(示例 cu121) pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121 # 2) PyPI + git 源依赖 pip install -r environment/requirements-opg.txt # 3) 需现场编译的 CUDA 扩展(按目标 CUDA 编译,务必对齐 torch/CUDA) bash environment/build_cuda_exts.sh # 见下,封装了下面所有 clone+build # 4) 引擎自带子包(editable,从仓库内 engine/ 安装) for p in Amodal3R fpose fpose/fpose/bundlesdf/mycuda SpaTrackerV2 trellis DA3; do pip install -e engine/oneposeviagen/$p done ``` ## CUDA 扩展来源(已补全,2026-06-15) | 扩展 | 用途 | 来源 / 安装 | |---|---|---| | pytorch3d | 3D 算子 | `pip install --no-build-isolation git+https://github.com/facebookresearch/pytorch3d.git` | | flash-attn | 注意力 | `pip install flash-attn --no-build-isolation`(或用官方预编译 whl 对齐 cu121/torch2.4)| | nvdiffrast | 可微渲染 | `pip install git+https://github.com/NVlabs/nvdiffrast.git` | | diffoctreerast | TRELLIS 渲染 | `pip install git+https://github.com/JeffreyXiang/diffoctreerast.git` | | diff-gaussian-rasterization | 3DGS | clone `autonomousvision/mip-splatting` 后装其 `submodules/diff-gaussian-rasterization/` | | simple-knn | 3DGS | 同上 mip-splatting 的 `submodules/simple-knn/` | | **CuMesh**(原缺源)| TRELLIS2 网格/UV | `pip install --no-build-isolation git+https://github.com/JeffreyXiang/CuMesh.git`(需 torch≥2.4 + CUDA≥12.4)| | **nvdiffrec-renderutils**(原缺源)| TRELLIS2 渲染(模块名 `nvdiffrec_render`)| `git clone -b renderutils https://github.com/JeffreyXiang/nvdiffrec.git /tmp/nvdiffrec && pip install /tmp/nvdiffrec --no-build-isolation` | | o-voxel / flexgemm | TRELLIS2 | o-voxel 源在仓库内 `engine/oneposeviagen/TRELLIS2/o-voxel`;flexgemm 见 `microsoft/TRELLIS.2` 的 `setup.sh --flexgemm` | | utils3d | 工具 | `pip install git+https://github.com/EasternJournalist/utils3d.git` | | custom_rasterizer | Hunyuan 贴图(motion 环境)| 源在仓库内 `third_party/Motion324/scripts/hy3dgen/texgen/custom_rasterizer`,`pip install -e` 之 | | FoundationPose 扩展 | 位姿 | `cd engine/oneposeviagen/fpose && bash build_all_conda.sh` | > **重要**:CuMesh / nvdiffrec / o-voxel / flexgemm 仅 **TRELLIS2 backbone** 用到。 > 而 OVOW 入口(`scripts/run_video.sh` / `run_fullpipe_batch.sh`)的默认 backbone 是 **`uplegacy`** > (= legacy/Hi3DGen 生成后端 + 升级版位姿链路:romav2 匹配 + 零迭代 scale + 全部模型进程内常驻; > 见 `configs/paths.sh` 的 `OPG_MATCHER=romav2 / OPG_SCALE_REFINE_ITERS=0 / OPG_ROMAV2_INPROC=1 / OPG_FP_RESIDENT=1`)。 > uplegacy 的生成后端会折叠成 legacy/Hi3DGen,**不走 TRELLIS2**——所以**默认复现不需要 CuMesh/nvdiffrec**, > 上游 OnePoseviaGen 的 `setup.sh` 也只装前面的 CORE。仅当显式 `--recon-backbone trellis2` 时才需要: > `bash environment/build_cuda_exts.sh --trellis2`。权威脚本见 > [microsoft/TRELLIS.2 setup.sh](https://github.com/microsoft/TRELLIS.2/blob/main/setup.sh)。 > > OVOW 引擎自带 `engine/setup.sh`(= 上游官方安装流程)是权威参照;`build_cuda_exts.sh` 即对齐它。 注意: - **环境与 CUDA 强耦合,不能纯 pip 一键复现**;step 3/4 需现场编译,务必对齐 CUDA/torch 版本。 - `requirements-*.txt` 已剔除 dev 机器本地路径,只留 PyPI + git 源;原始 `pip freeze` 见 `*.freeze.bak`(仅版本参考)。 - 引擎自带 `setup.sh` 也编码了上述步骤,可对照参考。 - `flux` / `hdr` 环境为可选步骤,未纳入核心复现。