---

name: generation description: "Guides Zero123Plus single-image to multi-view generation, ControlNet variants, model choices, camera outputs, and postprocessing workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Zero123Plus Generation

Use this sub-skill when the task is about producing, adapting, or debugging Zero123Plus multiview outputs from a single image: base v1.1/v1.2 runs, depth ControlNet, normal ControlNet, six-view grid interpretation, matting postprocess, or reference-only upstream image synthesis for inputs.

Do not use this sub-skill for serving or deployment surfaces such as Streamlit, Gradio, Docker/Gitpod, or Cog. Route those questions to ../deployment/SKILL.md instead.

Read first

• references/workflows.md: copyable recipes for the base, depth ControlNet, normal ControlNet, matting, and the optional text-to-image input helper.
• references/api-reference.md: verified pipeline and helper signatures, parameter notes, and call behavior.
• references/model-and-camera.md: model ids, six fixed views, camera poses, field of view, and license notes.
• references/troubleshooting.md: cache, CUDA, VRAM, image-shape, diffusers version, and matting dependency failures.

Bundled scripts

• scripts/run_img_to_mv.py: v1.1 base single-image-to-six-view runner.
• scripts/run_depth_controlnet.py: v1.1 depth ControlNet runner.
• scripts/run_normal_gen.py: v1.2 normal generator runner with bundled matting postprocess.
• scripts/matting_postprocess.py: CPU-safe alpha/normal postprocess helper and CLI.

Routing procedure

1. Identify the requested flow: base multiview generation, depth ControlNet, normal ControlNet, postprocess-only work, or optional image synthesis before Zero123Plus.
2. Check GPU and cache readiness before a real run. The bundled scripts default to local-cache-only loading; pass --allow-download only when network access and model fetches are approved.
3. For a standard six-view grid, use the bundled script that matches the requested model family. The scripts already encode the source defaults for scheduler choice, width, height, and output layout.
4. For camera layout, view ordering, and license questions, use the model-and-camera reference. For exact callable signatures and preprocessing details, use the API reference.
5. If the user asks about Streamlit, Gradio, Docker, or Cog rather than the core generation path, switch to the deployment sub-skill instead of duplicating those details here.

Stop conditions

• No CUDA GPU is available for a real generation run.
• The requested model or custom pipeline is not already cached and downloads are not approved.
• The user is asking about training or finetuning, which this repository does not expose.
• The issue is really about serving or deployment rather than generation.