Getting Started¶

Install workspace members from the repository root with uv. The root workspace contains shared libraries under lib/* and model packages under models/*; member-specific commands should select the package so extras and dependency source mappings resolve correctly.

Install¶

git clone https://github.com/creative-graphic-design/design-generators.git
cd design-generators
uv sync --all-packages

For a narrower environment, install only the package you want to run.

uv sync --package layout-dm
uv run --package layout-dm pytest models/layout-dm/tests -m "not vendor_parity and not integration"

Run A Converted Checkpoint¶

Most weight-backed packages use locally converted checkpoint directories until planned Hub repos are published. Each model package has a REPRODUCING.md file with the download, reference generation, agreement-check, conversion, and smoke-test commands that create the local path.

For LayoutDM, run the minimal download and conversion commands from the repository root. See the full LayoutDM reproducibility guide for original-implementation reference generation, agreement checks, and smoke tests.

Use from_pretrained to load the converted directory. Prompt-only packages use save_pretrained for reusable configuration and exemplars.

uv run --package layout-dm python models/layout-dm/scripts/download_original.py \
  --output-dir .cache/layout-dm/original

uv run --package layout-dm --extra convert python models/layout-dm/scripts/convert_original_checkpoint.py \
  --dataset rico25 \
  --starter-dir .cache/layout-dm/original/download \
  --output-dir .cache/layout-dm/converted/layoutdm-rico25

uv run --package layout-dm python

from layout_dm import LayoutDMPipeline

path = ".cache/layout-dm/converted/layoutdm-rico25"
# After Hub publication: from_pretrained("creative-graphic-design/layoutdm-rico25")
pipe = LayoutDMPipeline.from_pretrained(path)
out = pipe(batch_size=1, seed=0, sampling="deterministic")

print(out.bbox.shape)
print(out.labels.shape)
print(out.mask.shape)
print(out.id2label[0])

torch.Size([1, 25, 4])
torch.Size([1, 25])
torch.Size([1, 25])
Text

bbox uses normalized center xywh coordinates in [0, 1], labels are dataset-local integer ids, mask=True marks valid elements, and id2label decodes labels for the active dataset.

GPU Selection¶

Agreement checks against the original implementation (the repository's vendor_parity test suite) and heavyweight conversion commands that need CUDA use a placeholder GPU selector. Replace <gpu-index> with one visible CUDA device on your machine, such as 0 on a single-GPU host.

CUDA_VISIBLE_DEVICES=<gpu-index> uv run --package layout-dm pytest models/layout-dm/tests/vendor_parity

Prompt-Only Packages¶

LayoutGPT and LayoutPrompter do not convert learned checkpoints. Install their workspace member, configure the provider credentials required by the selected Pydantic AI model, and load prompt configuration or exemplars with save_pretrained and from_pretrained as documented on each model page.