---

name: configuration-and-cli description: "Author and validate imagen-pytorch configs and safely use the imagen config, train, and sample CLI commands." disable-model-invocation: true metadata: disco-role: operating license: MIT

configuration-and-cli

Use this sub-skill when the task mentions imagen config JSON, default_config.json, UnetConfig, ImagenConfig, ElucidatedImagenConfig, ImagenTrainerConfig, CLI train/sample commands, config validation, command generation, or safe preflight checks for the imagen console command.

Fast route

1. Read configuration to choose the JSON schema (type: "original" or "elucidated"), cascade length, image sizes, channel count, dataset/checkpoint keys, and sample/save/validation interval keys.
2. Use default-config-template.json only as a LAION-scale template. It is not a safe local training default.
3. From this sub-skill directory, validate before running anything expensive:

   python scripts/validate_imagen_config.py path/to/imagen_config.json --mode train --unet 1
   

4. From this sub-skill directory, check the installed CLI surface without training or sampling:

   python scripts/imagen_cli_quickcheck.py
   

5. Generate or run commands from CLI reference. Use the exact --epoches spelling for training.
6. If validation or CLI use fails, consult troubleshooting.

Command workflow

• Create a starter config: imagen config --path ./imagen_config.json.
• Edit the config, especially dataset_name, dataset.batch_size, labels, checkpoint_path, imagen.image_sizes, imagen.unets, interval keys, and practical model sizes.
• Preflight with the bundled validator; do not let imagen train be the first validator because it loads datasets and can enter expensive training.
• Train with imagen train --config ./imagen_config.json --unet 1 --epoches 1 only after dataset and hardware are intentionally chosen.
• Sample with imagen sample --model ./imagen.pt --cond_scale 5 --load_ema true "prompt" only from a commandable trainer checkpoint and on a CUDA-capable runtime.

Boundaries and handoffs

• Training loop internals, checkpoint save/load mechanics, EMA behavior, and distributed training: route to training-and-checkpointing.
• Image sampling tensor semantics, prompt conditioning, inpainting image tensors, and quality/performance tuning: route to image-generation.
• Video configs and 3D/video tensor details beyond the video: true config switch: route to video-and-inpainting.
• Dataset construction, collators, URL/image/text labels, and T5 embedding/data-loader details: route to data-and-text-conditioning.

Keep this sub-skill focused on static configuration, CLI flags, command preflights, and safe command construction. Do not use it as proof of generation quality; CUDA import smoke was verified for the package, but realistic Imagen training and sampling remain practical CUDA-scale work.