---

name: environment-and-inputs description: "Prepare and diagnose Make-It-3D CUDA dependencies, model assets, Hugging Face access, DPT weights, and alpha-mask reference images before training." metadata: disco-role: operating disable-model-invocation: true license: NO_LICENSE

Environment and Inputs

Use this sub-skill before any Make-It-3D training/refinement run, or when a user reports setup, dependency, model asset, credential, or image-format failures. This route owns the prerequisites that determine whether the later training commands are safe to run.

What This Sub-Skill Covers

• CUDA/PyTorch/tiny-cuda-nn/raymarching/PyTorch3D/DPT/Stable Diffusion/CLIP/BLIP2 dependency readiness.
• DPT depth model weights and Hugging Face or local cache/token requirements.
• Reference image alpha-channel validation and practical foreground-mask checks.
• Choosing when to pass --text to avoid BLIP2 captioning.
• Setup failures that would make main.py fail before training.

Route coarse-stage command tuning to coarse-training and refine/test/mesh export commands to refinement-and-export.

Required Reads and Scripts

• Read references/dependency-map.md when selecting package versions or explaining old README pins on a modern host.
• Read references/input-formats.md when validating or preparing the reference image and prompt.
• Read references/troubleshooting.md for missing module, DPT weight, Hugging Face, CUDA extension, and alpha-mask failures.
• Run scripts/validate_alpha_input.py on a candidate --ref_path image.
• For broad environment diagnostics, use the root ../../scripts/make_it_3d_env_check.py.

Setup Decision Workflow

1. Ask for or infer the user's target machine constraints. GPU model, driver/toolkit, Python version, torch version, network permissions, and whether model weights are already cached all matter.
2. Check CUDA first. A successful CPU import is not enough for Make-It-3D's main path. The source sets opt.cuda_ray = True, so raymarching CUDA support must be treated as a real requirement.
3. Check import-time blockers before running main.py. In this source, main.py can fail at import time if PyTorch3D or other refine dependencies are missing, even if the user only wants --help.
4. Check model assets. DPT hybrid weights are expected at dpt_weights/dpt_hybrid-midas-501f0c75.pt by the main pipeline. Stable Diffusion and BLIP2 need cache/network/token unless already available.
5. Validate the alpha input. If the image lacks alpha, stop and ask the user to create an alpha PNG or accept that the repo's input assumptions need a deliberate source patch.
6. Pass a prompt when possible. Recommend --text "..." for reproducibility and to avoid BLIP2 download/memory cost.

Quick Checks

python /path/to/skill/scripts/make_it_3d_env_check.py --repo-root /path/to/Make-It-3D --dpt-weights /path/to/Make-It-3D/dpt_weights/dpt_hybrid-midas-501f0c75.pt
python /path/to/skill/sub-skills/environment-and-inputs/scripts/validate_alpha_input.py --image /path/to/ref_alpha.png

The second command should report a present alpha channel and nonzero foreground coverage. If it fails, fix the image before constructing coarse training commands.