WikiArt-Shards

Small conditional image generation model for abstract art-style 256×256 images.

This repository is inference-only. It contains the runtime code, model weights, label mappings, scheduler configuration, and examples needed to generate images.

Status: This model is still in training. The current weights are a very early checkpoint and are shared mainly for testing, experimentation, and reproducible inference. Output quality, conditioning accuracy, and generation stability are expected to improve in future checkpoints.

Conditions

The model supports three optional condition fields:

• artist
• genre
• style

Omit any field to use its ANY_* token.

Quick start

pip install -r requirements.txt
python generate.py --model_path . --artist "pablo-picasso" --genre "portrait" --style "Cubism" --scheduler ddim --num_inference_steps 50 --guidance_scale 2.0 --num_images 16 --grid --seed 42

List labels

python generate.py --model_path . --list_labels --label_limit 25

Supported generation options

--model_path             Path to the cloned model repository.
--artist                 Artist label. Omit for ANY_ARTIST.
--genre                  Genre label. Omit for ANY_GENRE.
--style                  Style label. Omit for ANY_STYLE.
--num_images             Number of images to generate.
--batch_size             Images per generation batch. 0 means all images at once.
--scheduler              ddpm, ddim, or dpm.
--num_inference_steps    Number of denoising steps.
--guidance_scale         Classifier-free guidance scale. 1.0 disables guidance.
--output_dir             Directory for generated images.
--grid                   Save a grid image.
--grid_rows              Grid rows. 0 means automatic layout.
--grid_cols              Grid columns. 0 means automatic layout.
--seed                   Random seed. If omitted, a random seed is generated.
--device                 auto, cuda, or cpu.
--allow_tf32             Enable or disable TF32 on compatible NVIDIA GPUs.
--save_config            Save generation_config.json.
--no-save_config         Do not save generation_config.json.
--list_labels            Print available label info and exit.
--label_limit            Number of labels to preview with --list_labels.

Notes

This is a compact experimental image generation model. It is intended for demos, exploration, and reproducible local inference. It is not comparable to large text-to-image foundation models.