DreamCraft3D

Use this repo skill when a user is working with DreamCraft3D, a CUDA-first research checkout for hierarchical 3D generation from a reference image using staged geometry sculpting and bootstrapped texture refinement.

This skill is self-contained operating guidance: it distills commands, configs, data layouts, model artifacts, and troubleshooting from the repository. It does not include model weights or run training by itself.

Start here

1. Check whether the user's checkout matches this skill baseline in references/repo-provenance.md.
2. Read references/config-reference.md for the main CLI, registry names, canonical configs, and shared artifact conventions.
3. Use references/troubleshooting.md for cross-cutting CUDA, dependency, model-artifact, image/config, and output failures.
4. Run scripts/dreamcraft3d_static_check.py for a safe checkout/skill/config preflight; it does not import heavy ML packages or launch training.

Sub-skill routes

• Use sub-skills/image-preparation/SKILL.md for reference-image preprocessing, _rgba/_depth/_normal sidecars, recentering, caption sidecars, and image-load assertions.
• Use sub-skills/generation-pipeline/SKILL.md for the four canonical stages, launch.py commands, OmegaConf overrides, checkpoint chaining, and stage-specific troubleshooting.
• Use sub-skills/bootstrapped-texture/SKILL.md for optional Zero123++ multiview generation, DreamBooth/LoRA texture boosting, Stable Zero123/Stable Diffusion/DeepFloyd/Omnidata artifacts, and Janus-problem mitigation.
• Use sub-skills/export-and-evaluation/SKILL.md for mesh export, OBJ/MTL/textures, output directory summaries, validation/test assets, metrics, and progress-video utilities.
• Use sub-skills/interfaces-and-monitoring/SKILL.md for installation/backend triage, Docker/NVIDIA container guidance, Gradio launch/watch behavior, GPU memory checks, and safe environment diagnostics.

Installation and runtime route

DreamCraft3D is operated from a source checkout rather than an installable package distribution. For setup questions, route to interfaces-and-monitoring: it covers the documented Python, CUDA, PyTorch, Docker, NVIDIA Container Toolkit, compiled extension, and model-artifact requirements. A minimal public readiness check is to confirm the checkout has launch.py, requirements.txt, the four configs/dreamcraft3d-*.yaml files, and visible CUDA hardware before attempting full training.

Minimal operating checks

From a DreamCraft3D checkout, use the bundled static checker when you need a cheap preflight:

python <skill-dir>/scripts/dreamcraft3d_static_check.py --repo-root . --json

For a full operator preflight with GPU/model-path warnings, route to interfaces-and-monitoring and use its environment checker.

Canonical workflow at a glance

1. Prepare or validate a single image family: <stem>_rgba.png, <stem>_depth.png, and <stem>_normal.png.
2. Run coarse NeRF with configs/dreamcraft3d-coarse-nerf.yaml.
3. Run coarse NeuS with configs/dreamcraft3d-coarse-neus.yaml and system.weights=<coarse-nerf-last.ckpt>.
4. Run geometry refinement with configs/dreamcraft3d-geometry.yaml and system.geometry_convert_from=<coarse-neus-last.ckpt>.
5. Run texture refinement with configs/dreamcraft3d-texture.yaml and system.geometry_convert_from=<geometry-last.ckpt>.
6. Export a mesh from the final trial's configs/parsed.yaml and ckpts/last.ckpt using system.exporter_type=mesh-exporter.

Use the generation-pipeline command builder instead of composing long shell commands from memory.

Required runtime assumptions

DreamCraft3D's actual generation path requires a CUDA-capable NVIDIA GPU, compatible PyTorch CUDA wheels, GPU rendering/ML extensions, and large pretrained model artifacts. Static checks, command builders, and output summarizers can run without those dependencies, but they are not proof that full training or export succeeded.

Import and verification status

This skill was produced for a no-import run. Do not assume it has been installed into the managed live repo-skills library. The verification artifacts for this production run record that full CUDA native training/export was not executed.