Sync deps: Dockerfile system deps + ffmpeg, single pip flow

Dockerfile

@@ -17,11 +17,15 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     libxcb-xfixes0 \
     && rm -rf /var/lib/apt/lists/*
 
-COPY requirements.txt .
-RUN pip install --no-cache-dir -r requirements.txt
+COPY requirements.txt requirements-docker.txt ./
+# Install CPU-only PyTorch first so torchmcubes can build without CUDA (Docker build has no GPU/CUDA).
+# If we install default torch (CUDA wheel), CMake fails: "Your installed Caffe2 version uses CUDA but I cannot find CUDA".
+RUN pip install --no-cache-dir torch torchvision --index-url https://download.pytorch.org/whl/cpu
+# Install rest of deps without touching torch/torchvision (requirements-docker.txt omits them)
+RUN pip install --no-cache-dir -r requirements-docker.txt
 # torchmcubes: build deps for --no-build-isolation (scikit-build-core, pybind11, ninja)
 RUN pip install --no-cache-dir scikit-build-core pybind11 ninja
-# torchmcubes must be built after torch is installed (CMake needs TorchConfig.cmake)
+# torchmcubes must be built after torch is installed (CMake needs TorchConfig.cmake); CPU Torch = no CUDA required at build.
 RUN pip install --no-cache-dir --no-build-isolation "git+https://github.com/tatsy/torchmcubes.git"
 
 # TripoSR repo (model downloaded at runtime from Hugging Face)

README.md

@@ -46,7 +46,7 @@ After that, run Skybox or Mesh again; the model will download on first use.
 - **"Could not load Stable Diffusion"** → Add `HF_TOKEN` in Settings → Variables and secrets, then restart the Space.
 - **"TripoSR not found"** → Rebuild the Space (the Dockerfile clones TripoSR). If you forked the Space, ensure the Dockerfile and `scripts/` are present.
 - **Mesh generation times out** → Use **Mesh resolution 256** and disable **Bake texture atlas** for a faster run; GPU Spaces are much faster than CPU.
-- **Build fails on torchmcubes** → The Dockerfile installs torch first, then builds torchmcubes with `--no-build-isolation` so CMake finds Torch. Do not change the order of `pip install` steps.
+- **Build fails on torchmcubes** → The Dockerfile installs **CPU-only PyTorch** first (`--index-url https://download.pytorch.org/whl/cpu`) so torchmcubes can build without CUDA (Docker build has no GPU). Do not change the order of `pip install` steps; do not remove the CPU-index torch install or the build will fail with "Your installed Caffe2 version uses CUDA but I cannot find CUDA".
 
 ---
 

requirements-docker.txt

@@ -0,0 +1,26 @@
+# Same as requirements.txt but WITHOUT torch/torchvision (installed from CPU index in Dockerfile first)
+# Evoneural MVP – Space (Docker build)
+
+# Core (torch/torchvision installed separately from PyTorch CPU index)
+Pillow>=10.0.0
+numpy>=1.24.0
+
+# Streamlit
+streamlit>=1.28.0
+
+# Stable Diffusion (skybox + text-to-image for mesh)
+diffusers>=0.25.0
+transformers>=4.35.0
+accelerate>=0.25.0
+safetensors>=0.4.0
+
+# TripoSR / mesh pipeline
+huggingface-hub>=0.20.0
+rembg[cpu]>=2.0.50
+xatlas>=0.0.9
+trimesh>=4.0.0
+einops>=0.7.0
+omegaconf>=2.3.0
+moderngl>=5.10.0
+imageio[ffmpeg]>=2.33.0
+scipy>=1.10.0