Stb Invokeai Marketplace Studio

Portable InvokeAI 6.13.6 workflow for turning one reference image into editable marketplace cards.

Public repository: born2be4/stb-invokeai-marketplace-studio

The reference subject is extracted once and reused as source pixels. The image API generates backgrounds only. Each B-E variant is packaged as an official InvokeAI .invk Canvas project with five independent layers:

1. Background
2. Product
3. Headline
4. Supporting Text
5. Accent

The first job contains unchanged baseline A plus four strategies:

• B: lifestyle scene
• C: premium studio
• D: problem-solution
• E: seasonal promo

Wildberries Template

The default production canvas is 1200x1600 (3:4). Every rendered Product layer is measured before export and must occupy at least 50% of the canvas height while remaining fully visible. Each .invk keeps all five raster layers independently selectable.

Install In WSL

Requirements: Ubuntu WSL, Python 3.12, uv, and enough storage for InvokeAI and the u2net subject-extraction model.

git clone https://github.com/born2be4/stb-invokeai-marketplace-studio.git
cd stb-invokeai-marketplace-studio
bash scripts/install-invokeai.sh
bash scripts/start-invokeai.sh

In InvokeAI open Models → External Providers → OpenAI and configure:

• Base URL: https://api.provod.ai
• API key: store it only in the local InvokeAI secret configuration

Then register the four approved models:

PYTHONPATH=src .venv/bin/python \
  scripts/setup_invokeai.py sync-models

The tracked catalog is config/models.yaml. Do not add guessed or historical model IDs.

Build Layered Cards

PYTHONPATH=src .venv/bin/python -m stb_studio build \
  --reference /path/to/reference.png \
  --output outputs/my-job \
  --title "PRIMARY BENEFIT" \
  --subtitle "Supporting proof" \
  --badge "NEW" \
  --invokeai-root /home/codex/workspace/invokeai

Use --offline to test composition without paid API calls. Existing complete variants are reused unless --force is supplied.

Open In Canvas

The automated command backs up the current Canvas, uploads the five layers, and installs the selected project:

PYTHONPATH=src .venv/bin/python -m stb_studio install-canvas \
  outputs/my-job/B-lifestyle/B-lifestyle.invk \
  --backup outputs/my-job/canvas-backup.json

Reload InvokeAI after installation. The project menu can also load any generated .invk manually.

Output

my-job/
|-- A-baseline/
|-- B-lifestyle/
|   |-- B-lifestyle.invk
|   |-- preview.png
|   |-- prompt.txt
|   `-- layers/*.png
|-- C-premium-studio/
|-- D-problem-solution/
|-- E-seasonal-promo/
|-- contact-sheet.png
`-- job.json

Codex And MCP

• Agent skill: codex-skill/SKILL.md
• MCP server: mcp/stb_invokeai_mcp.py
• MCP configuration: mcp/mcp-config.toml.example

MCP tools build jobs, validate .invk, list outputs, and install a project into local Canvas. They never return API keys.

Transferable Skill

Download the public .skill archive from GitHub Releases. Extract its stb-invokeai-marketplace-studio directory into the recipient's Codex skills directory, then restart Codex.

The public archive contains agent rules, API endpoint and model settings, workflow instructions, and MCP setup. It contains no API key. A private archive may additionally contain private/api_keys.yaml; keep that file outside Git, transfer it only through a trusted channel, and never post it in an issue or chat.

Build the public archive locally:

python scripts/package_skill.py \
  --source codex-skill \
  --output dist/stb-invokeai-marketplace-studio.skill

The packager rejects public files containing secret-like values.

Validation

PYTHONPATH=src .venv/bin/python -m unittest discover -s tests -v
PYTHONPATH=src .venv/bin/python -m compileall -q src mcp scripts

Generated cards are drafts. A designer must verify SKU fidelity, claims, typography, safe areas, and the test hypothesis before marketplace publication.