pt25 update

pt25-cu124/attachments/!FIRST_RUN-compile-install.bat

@@ -1,16 +1,17 @@
 @REM Edit this to your GPU arch.
 @REM You don't need to add the "+PTX". Here it works as a fail-safe (providing forward compatibility).
+
 set TORCH_CUDA_ARCH_LIST=6.1+PTX
 
 @REM ===========================================================================
 
+set CMAKE_ARGS=-DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DWITH_NVCUVID=ON
+set PATH=%PATH%;%~dp0\python_embeded\Scripts
+
 @REM Compile-install PyTorch3D
 @REM PyTorch3D on Windows works best through compile-install. Binary-install will fail on some workflows.
 @REM e.g. "CUDA error: no kernel image is available for execution on the device"
 
-set CMAKE_ARGS=-DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DWITH_NVCUVID=ON
-set PATH=%PATH%;%~dp0\python_embeded\Scripts
-
 .\python_embeded\python.exe -s -m pip install --force-reinstall ^
  .\extras\pytorch3d
 
@@ -39,20 +40,9 @@ set PATH=%PATH%;%~dp0\python_embeded\Scripts
  .\extras\diff-gaussian-rasterization
 
 @REM Differential Octree Rasterization
-@REM Note that PIP will auto git clone submodules, no need to explicit clone it.
-.\python_embeded\python.exe -s -m pip install ^
- .\extras\diffoctreerast
-
-@REM (Optional) Flash Attention
-@REM flash-attn can ONLY be used on Ampere and later GPUs (RTX 30 series and beyond)
-@REM Safe to remove this if you are not using RTX 30/40 or A100+.
-
-@REM Limit Ninja jobs to avoid OOM
-
-set MAX_JOBS=4
 
 .\python_embeded\python.exe -s -m pip install ^
- flash-attn --no-build-isolation
+ .\extras\diffoctreerast
 
 @REM ===========================================================================
 

pt25-cu124/attachments/reinstall-deps-for-3d.bat

@@ -20,22 +20,33 @@ pause
 
 @echo on
 
-@REM In order to save your time on compiling, edit this line according to your GPU arch.
+@REM ===========================================================================
+
+@REM [IMPORTANT] To save your time on compiling, edit this line according to your GPU arch.
 @REM Ref: https://github.com/ashawkey/stable-dreamfusion/issues/360#issuecomment-2292510049
 @REM Ref: https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/
+
 set TORCH_CUDA_ARCH_LIST=5.2+PTX;6.0;6.1+PTX;7.5;8.0;8.6;8.9+PTX
 
-set CMAKE_ARGS=-DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DWITH_NVCUVID=ON
+@REM ===========================================================================
 
 set PATH=%PATH%;%~dp0\python_embeded\Scripts
 
+set CMAKE_ARGS=-DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DWITH_NVCUVID=ON
+
+.\python_embeded\python.exe -s -m pip install --force-reinstall ^
+ spconv-cu124
+
 if not exist ".\tmp_build" mkdir tmp_build
 
 .\python_embeded\python.exe -s -m pip install numpy==1.26.4
 
 git clone --depth=1 https://github.com/MrForExample/Comfy3D_Pre_Builds.git ^
  .\tmp_build\Comfy3D_Pre_Builds
 
+git clone --depth=1 https://github.com/autonomousvision/mip-splatting.git ^
+ .\tmp_build\mip-splatting
+
 .\python_embeded\python.exe -s -m pip wheel -w tmp_build ^
  .\tmp_build\Comfy3D_Pre_Builds\_Libs\pointnet2_ops
 
@@ -46,7 +57,14 @@ git clone --depth=1 https://github.com/MrForExample/Comfy3D_Pre_Builds.git ^
  .\tmp_build\Comfy3D_Pre_Builds\_Libs\vox2seq
 
 .\python_embeded\python.exe -s -m pip wheel -w tmp_build ^
- git+https://github.com/ashawkey/diff-gaussian-rasterization.git
+ .\tmp_build\mip-splatting\submodules\diff-gaussian-rasterization
+
+@REM Note that PIP will auto git clone submodules, no need to explicit clone it.
+.\python_embeded\python.exe -s -m pip wheel -w tmp_build ^
+ git+https://github.com/JeffreyXiang/diffoctreerast.git
+
+.\python_embeded\python.exe -s -m pip wheel -w tmp_build ^
+ git+https://github.com/EasternJournalist/utils3d.git
 
 .\python_embeded\python.exe -s -m pip wheel -w tmp_build ^
  git+https://github.com/ashawkey/kiuikit.git
@@ -63,4 +81,17 @@ del .\tmp_build\numpy-2*.whl
 
 for %%i in (.\tmp_build\*.whl) do .\python_embeded\python.exe -s -m pip install --force-reinstall "%%i"
 
+@REM ===========================================================================
+
+@REM (Optional) Flash Attention for TRELLIS demo
+@REM "flash-attn" can ONLY be used on Ampere and later GPUs (RTX 30 series / A100 and beyond).
+@REM WARNING: VERY long build time!
+
+@REM Limit Ninja jobs to avoid OOM. If have RAM larger than 96GB, just remove this line.
+rem set MAX_JOBS=4
+
+rem .\python_embeded\python.exe -s -m pip install flash-attn --no-build-isolation
+
+@REM ===========================================================================
+
 .\python_embeded\python.exe -s -m pip install numpy==1.26.4

pt25-cu124/attachments/run.bat

@@ -20,6 +20,9 @@ rem set HF_HUB_ENABLE_HF_TRANSFER=1
 @REM This command redirects HuggingFace-Hub to download model files in this folder.
 set HF_HUB_CACHE=%~dp0\HuggingFaceHub
 
+@REM This command redirects Pytorch Hub to download model files in this folder.
+set TORCH_HOME=%~dp0\TorchHome
+
 @REM This command will set PATH environment variable.
 set PATH=%PATH%;%~dp0\python_embeded\Scripts
 

pt25-cu124/attachments/run_trellis_demo.sh

@@ -1,26 +1,29 @@
 #!/bin/bash
 set -eu
 
-# The TRELLIS demo was written in Linux context in mind. So here we use bash to run it.
+# The TRELLIS demo was written with Linux context in mind. So here we use bash to run it.
 # You need to install <Git for Windows> with <Git Bash> (installed by default).
 # Download: https://git-scm.com/download/win
 
 ################################################################################
 # Edit this first! According to your GPU model.
 export TORCH_CUDA_ARCH_LIST="6.1+PTX"
 
-# "flash-attn" can ONLY be used on Ampere and later GPUs (RTX 30 series and beyond).
-# Change it to "xformers" if you can't start the program.
-export ATTN_BACKEND="flash-attn"
+################################################################################
+# Optional Optimizations
 
-# "auto" will be faster but will do benchmarking at the beginning.
 # If run only once, set to "native".
-export SPCONV_ALGO="auto"
+# "auto" will be faster but will do benchmarking at the beginning.
+export SPCONV_ALGO="native"
 
+# Default to "xformers" for compatibility
+# "flash-attn" for higher performance.
+# Flash Attention can ONLY be used on Ampere and later GPUs (RTX 30 series / A100 and beyond).
+export ATTN_BACKEND="xformers"
 
 ################################################################################
 
-# To set proxy, edit and uncomment the lines below
+# To set proxy, uncomment and edit the lines below
 # (remove '#' in the beginning of line).
 #export HTTP_PROXY=http://localhost:1081
 #export HTTPS_PROXY=$HTTP_PROXY
@@ -40,19 +43,16 @@ export SPCONV_ALGO="auto"
 #export PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
 #export HF_ENDPOINT="https://hf-mirror.com"
 
-# To set HuggingFace Access Token, uncomment and edit the line below.
-# https://huggingface.co/settings/tokens
-#export HF_TOKEN=
-
-# To enable HuggingFace Hub's experimental high-speed file transfer, uncomment the line below.
-# https://huggingface.co/docs/huggingface_hub/hf_transfer
-#export HF_HUB_ENABLE_HF_TRANSFER=1
+################################################################################
 
 workdir="$(pwd)"
 
 # This command redirects HuggingFace-Hub to download model files in this folder.
 export HF_HUB_CACHE="$workdir/HuggingFaceHub"
 
+# This command redirects Pytorch Hub to download model files in this folder.
+export TORCH_HOME="$workdir/TorchHome"
+
 # This command will set PATH environment variable.
 export PATH="${PATH}:$workdir/python_embeded/Scripts"
 
@@ -76,4 +76,10 @@ fi ;
 $workdir/python_embeded/Scripts/huggingface-cli.exe download JeffreyXiang/TRELLIS-image-large
 
 # Run the TRELLIS official Gradio demo
-./python_embeded/python.exe -s TRELLIS/app.py
+
+echo "########################################"
+echo "[INFO] Starting TRELLIS demo..."
+echo "########################################"
+
+cd TRELLIS
+../python_embeded/python.exe -s app.py

pt25-cu124/attachments/中文脚本/!首次运行-编译安装依赖项.bat

@@ -1,34 +1,52 @@
 @REM 编辑为你的 GPU 对应架构
 @REM 修改时无需保留 "+PTX" ，其用于向前兼容，防止用户忘了该步骤。
+
 set TORCH_CUDA_ARCH_LIST=6.1+PTX
 
-@REM 编译安装 PyTorch3D
-@REM PyTorch3D 对 Windows 不甚友好，所有二进制安装都可能在某个节点报错，极难排查，故最好提前编译安装。
+@REM 使用清华 PyPI 源
+set PIP_INDEX_URL=https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+
+@REM ===========================================================================
 
-set CMAKE_ARGS=-DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DWITH_NVCUVID=ON
 set PATH=%PATH%;%~dp0\python_embeded\Scripts
+set CMAKE_ARGS=-DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -DWITH_CUBLAS=ON -DWITH_NVCUVID=ON
 
-set PIP_INDEX_URL=https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
+@REM 编译安装 PyTorch3D
+@REM PyTorch3D 对 Windows 不甚友好，所有二进制安装都可能在某个节点报错，极难排查，故最好提前编译安装。
 
 .\python_embeded\python.exe -s -m pip install --force-reinstall ^
- "git+https://ghp.ci/https://github.com/facebookresearch/pytorch3d.git"
+ .\extras\pytorch3d
 
 @REM 编译安装 pointnet2_ops ，该组件用于 Triplane Gaussian
 
 .\python_embeded\python.exe -s -m pip install --force-reinstall ^
  .\extras\pointnet2_ops
 
-@REM 编译安装 diff-gaussian-rasterization ，该组件用于 Triplane Gaussian
 
-.\python_embeded\python.exe -s -m pip install --force-reinstall ^
- "git+https://ghp.ci/https://github.com/ashawkey/diff-gaussian-rasterization.git"
-
-@REM 编译安装 simple-knn
+@REM 编译安装 simple-knn ，该组件用于 Gaussian Splatting
 
 .\python_embeded\python.exe -s -m pip install --force-reinstall ^
  .\extras\simple-knn
 
-@REM 编译安装 vox2seq ，该组件用于 TRELLIS
+@REM ===========================================================================
+@REM 用于 TRELLIS 的组件
+@REM 注意我们跳过了 'utils3d' ，该组件不需要本地重新编译
+
+@REM vox2seq
 
 .\python_embeded\python.exe -s -m pip install --force-reinstall ^
  .\extras\vox2seq
+
+@REM diff-gaussian-rasterization
+
+.\python_embeded\python.exe -s -m pip install --force-reinstall ^
+ .\extras\diff-gaussian-rasterization
+
+@REM Differential Octree Rasterization
+
+.\python_embeded\python.exe -s -m pip install ^
+ .\extras\diffoctreerast
+
+@REM ===========================================================================
+
+.\python_embeded\python.exe -s -m pip install numpy==1.26.4

pt25-cu124/attachments/中文脚本/run_cn.bat

@@ -22,6 +22,9 @@ set HF_ENDPOINT=https://hf-mirror.com
 @REM 该环境变量指示 HuggingFace Hub 下载模型到"本目录\HuggingFaceHub"，而不是"用户\.cache"目录。
 set HF_HUB_CACHE=%~dp0\HuggingFaceHub
 
+@REM 该环境变量指示 Pytorch Hub 下载模型到"本目录\TorchHome"，而不是"用户\.cache"目录。
+set TORCH_HOME=%~dp0\TorchHome
+
 @REM 该命令配置 PATH 环境变量。
 set PATH=%PATH%;%~dp0\python_embeded\Scripts
 

pt25-cu124/attachments/中文脚本/run_trellis_demo_cn.sh

@@ -0,0 +1,85 @@
+#!/bin/bash
+set -eu
+
+# 因为 TRELLIS 官方 demo 是以 Linux 语境编写的，未考虑跨平台，这里需用 Bash 运行
+# 下载 Git for Windows: https://git-scm.com/download/win
+# 并在安装时选择 Git Bash（默认）
+
+################################################################################
+# 务必根据你的 GPU 型号配置！
+export TORCH_CUDA_ARCH_LIST="6.1+PTX"
+
+################################################################################
+# 性能优化（可选）
+
+# 如果仅单次运行，使用 "native" 即可
+# 如果长期运行，使用 "auto" 会有更好性能，但一开始会花时间进行性能测试。
+export SPCONV_ALGO="native"
+
+# 默认使用 "xformers" 以保证兼容性
+# 如果需要高性能，尝试改为 "flash-attn"
+# Flash Attention 只能用于 Ampere (RTX 30 系 / A100) 及更新的 GPU
+export ATTN_BACKEND="xformers"
+
+################################################################################
+
+# 如需配置代理，取消注释并编辑以下部分
+# (删除行首井号 # )
+#export HTTP_PROXY=http://localhost:1081
+#export HTTPS_PROXY=$HTTP_PROXY
+#export http_proxy=$HTTP_PROXY
+#export https_proxy=$HTTP_PROXY
+#export NO_PROXY="localhost,*.local,*.internal,[::1],fd00::/7,
+#10.0.0.0/8,127.0.0.0/8,169.254.0.0/16,172.16.0.0/12,192.168.0.0/16,
+#10.*,127.*,169.254.*,172.16.*,172.17.*,172.18.*,172.19.*,172.20.*,
+#172.21.*,172.22.*,172.23.*,172.24.*,172.25.*,172.26.*,172.27.*,
+#172.28.*,172.29.*,172.30.*,172.31.*,172.32.*,192.168.*,
+#*.cn,ghproxy.com,*.ghproxy.com,ghproxy.org,*.ghproxy.org,
+#gh-proxy.com,*.gh-proxy.com,ghproxy.net,*.ghproxy.net"
+#export no_proxy=$NO_PROXY
+#echo "[INFO] Proxy set to $HTTP_PROXY"
+
+# 配置使用国内镜像站点
+export PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
+export HF_ENDPOINT="https://hf-mirror.com"
+
+################################################################################
+
+workdir="$(pwd)"
+
+# 该环境变量指示 HuggingFace Hub 下载模型到"本目录\HuggingFaceHub"，而不是"用户\.cache"目录。
+export HF_HUB_CACHE="$workdir/HuggingFaceHub"
+
+# 该环境变量指示 Pytorch Hub 下载模型到"本目录\TorchHome"，而不是"用户\.cache"目录。
+export TORCH_HOME="$workdir/TorchHome"
+
+# 该命令配置 PATH 环境变量。
+export PATH="${PATH}:$workdir/python_embeded/Scripts"
+
+# 该环境变量使 .pyc 缓存文件集中保存在一个文件夹下，而不是随 .py 文件分布。
+export PYTHONPYCACHEPREFIX="$workdir/pycache"
+
+# 该命令会复制 u2net.onnx 到用户主目录下，以免启动时还需下载。
+if [ ! -f "${HOME}/.u2net/u2net.onnx" ]; then
+  if [ -f "./extras/u2net.onnx" ]; then
+    mkdir -p "${HOME}/.u2net"
+    cp "./extras/u2net.onnx" "${HOME}/.u2net/u2net.onnx"
+  fi
+fi
+
+# 下载 TRELLIS 模型（不会重复下载）
+if [ ! -f "$workdir/python_embeded/Scripts/.hf-hub-reinstalled" ] ; then
+    $workdir/python_embeded/python.exe -s -m pip install --force-reinstall huggingface-hub
+    touch "$workdir/python_embeded/Scripts/.hf-hub-reinstalled"
+fi ;
+
+$workdir/python_embeded/Scripts/huggingface-cli.exe download JeffreyXiang/TRELLIS-image-large
+
+# 运行 TRELLIS 官方 Gradio demo
+
+echo "########################################"
+echo "[INFO] Starting TRELLIS demo..."
+echo "########################################"
+
+cd TRELLIS
+../python_embeded/python.exe -s app.py