Big Sleep

Big Sleep turns a text prompt into an image by optimizing a BigGAN latent with CLIP guidance. Use this skill when a user wants to run dream, compose positive or negative prompts, save progress frames, tune generation settings, or diagnose the CUDA-only runtime.

Start here

• Use scripts/check_runtime.py --check-cli to confirm the installed package, CUDA torch, and dream --help.
• Read references/workflows.md for end-to-end CLI and Python recipes.
• Read references/api-reference.md for verified signatures and method behavior.
• Read references/troubleshooting.md when import, CUDA, download, cache, or prompt-validation errors appear.
• Read references/repo-provenance.md to check whether this skill still matches the current repository snapshot.
• references/repo-routing-metadata.json feeds repo-skills-router; keep it aligned with this router.

Install

Big Sleep is effectively CUDA-only because big_sleep.big_sleep asserts that CUDA is available during import.

Verified on this host:

python -m pip install --index-url https://download.pytorch.org/whl/cu124 torch torchvision
python -m pip install big-sleep
python -m pip check
dream --help
python scripts/check_runtime.py --check-cli

If your driver needs a different CUDA wheel tag, install a matching CUDA-enabled torch/torchvision pair instead of the example cu124 wheel.

Main routes

• dream command-line runs → references/workflows.md
• Imagine(...) or BigSleep(...) from Python → references/api-reference.md
• Multi-prompt, negative prompt, save-best, save-progress, seeding, and file-naming questions → references/workflows.md
• Missing CUDA, missing libcudnn, stale torch wheels, or first-run downloads → references/troubleshooting.md
• Staleness / refresh decisions → references/repo-provenance.md

Supported workflows

• Text-to-image generation from the dream CLI.
• Python generation with Imagine.
• Multi-prompt generation with |.
• Negative-prompt suppression with text_min.
• Image-conditioned runs with img.
• Save-best, save-progress, seeding, and filename controls.
• Low-level BigSleep control for custom loops.

What this skill does not cover

• CPU-only execution.
• Diffusion, LoRA, or generic vision-model training stacks.
• Serving APIs or web apps.
• Repo-maintenance workflows.

Notes

• The Fire CLI spells flags with underscores in the verified help output, such as --save_progress, --save_best, --open_folder, --text_min, --max_classes, and --larger_model; use explicit boolean assignments such as --open_folder=False for headless runs, not space-separated --open_folder false.
• The default open_folder behavior opens the output directory after a run starts; disable it on headless or remote sessions.
• Generated images are written to the current working directory unless you change where the command is invoked.