---

name: artwork-generation description: "Generate Disco Diffusion artwork with DiscoArt Python APIs, runtime/model choices, DocArray outputs, recovery, and generation troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Artwork Generation

Use this sub-skill when the user wants to generate or plan Disco Diffusion artwork through the DiscoArt Python API. It covers create(**kwargs), go_big(), init images/documents, model/backend choices, output artifacts, DocArray recovery, and generation-specific failures.

Trigger on requests like

• "Generate an artwork with DiscoArt" or "write a create() call for this prompt".
• "Choose safe DiscoArt size/steps/CLIP/diffusion settings for my GPU".
• "Where did DiscoArt save my images/protobuf?" or "recover a lost DiscoArt session".
• "Debug CUDA, model cache/download, OOM, W&B, or no-output issues during create()".
• "Use an existing DiscoArt Document/DocumentArray as init_document" or "run go_big()".

Route elsewhere

• Prompt schema, prompt scheduling, schedule grammar, load_config/save_config/show_config/export_python, and YAML validation: read ../configuration-and-prompts/SKILL.md.
• python -m discoart CLI, Jina service endpoints, Docker, and Jupyter/runtime launch patterns: read ../cli-and-serving/SKILL.md.
• Cross-cutting install/import and package-level environment failures: read ../../references/troubleshooting.md when available.
• Maintainer scripts, release automation, CI, and repository development tasks are outside this generated operating skill.

Read order

1. references/artwork-api.md for create()/go_big() workflows, parameter grouping, output files, DocArray recipes, and safe examples.
2. references/model-and-runtime.md for diffusion/CLIP model choices, CUDA/cache/environment/W&B constraints, and memory trade-offs.
3. references/troubleshooting.md for CUDA, model download/cache, OOM, missing output/protobuf, DocArray recovery, and runtime symptoms.
4. scripts/plan_create_request.py when you need a safe local summary of a config without running generation or downloading models.

Quick workflow

1. Confirm the surface. Use this sub-skill only for Python API generation. If the user asks for CLI/service execution, route to cli-and-serving; if they ask for prompt-schema or schedule syntax, route to configuration-and-prompts.
2. Plan without side effects. Normalize proposed kwargs through discoart.config.load_config or the bundled scripts/plan_create_request.py; do not call create() until the user intends to run generation and the environment is ready.
3. Choose safe generation defaults. Prefer an explicit name_docarray, explicit DISCOART_OUTPUT_DIR and DISCOART_CACHE_DIR, n_batches=1, batch_size=1, a size that is a multiple of 64, bounded steps, and a small CLIP model set for first runs.
4. Check runtime risk. DiscoArt can run on CPU but is practically CUDA-oriented. First create() calls may download diffusion, secondary, and CLIP weights unless caches are already populated.
5. Run and preserve artifacts. create() returns a DocArray DocumentArray; final/intermediate images and da.protobuf.lz4 are saved under <DISCOART_OUTPUT_DIR or .>/<name_docarray>/.
6. Recover or debug by name. For interrupted/lost sessions, use the known name_docarray, local da.protobuf.lz4, or DocumentArray.pull(name_docarray) when cloud backup was enabled and reachable.

Guardrails

• Never treat config planning as proof that generation will succeed; model downloads, CUDA memory, and cache contents are runtime gates.
• Do not run full native generation tests from this sub-skill draft. Later verification may use the tiny-create-smoke and readme-api-workflows candidates only when CUDA/cache/time limits make them bounded.
• Keep examples self-contained: import from the installed discoart package and do not require opening the original repository checkout.

Version facts distilled for this skill

• Package version target: DiscoArt 0.12.2.
• create runtime signature: create(**kwargs) -> Optional[DocumentArray], backed by 49 default arguments.
• go_big signature: go_big(doc, window_size=256, upscale_factor=2, skip_rate=0.8, stride_size=None, **kwargs) -> Document.
• A CUDA smoke check passed during skill construction, but every user runtime must still check its own torch.cuda.is_available() and model cache state.