#!/usr/bin/env python3 """Run InDepth from the command line. Usage: python run_indepth.py "cat wearing green crown" --steps 4 --cfg 1.5 --seed 13372000 python run_indepth.py --prompts a.txt --out sprites/ # one prompt per line python run_indepth.py --local /path/to/indepth-10M ... # local dir instead of the hub Requires: torch, safetensors, huggingface_hub (for hub download), pillow. """ import argparse import json from pathlib import Path import numpy as np from PIL import Image def main(): parser = argparse.ArgumentParser(description="Generate 16x16 RGBA sprites with InDepth") parser.add_argument("prompt", nargs="?", help="text prompt (lowercase, short works best)") parser.add_argument("--prompts", help="text file with one prompt per line") parser.add_argument("--out", default="out/", help="output directory") parser.add_argument("--local", help="local model directory (skips hub download)") parser.add_argument("--steps", type=int, default=4, help="sampling steps (1-4)") parser.add_argument("--cfg", type=float, default=1.5, help="classifier-free guidance (1 = off)") parser.add_argument("--seed", type=int, default=13372000) parser.add_argument("--device", default="cpu", help="'cpu' or 'cuda'") parser.add_argument("--scale", type=int, default=8, help="PNG upscale factor (nearest)") args = parser.parse_args() if not args.prompt and not args.prompts: parser.error("give a prompt or --prompts file") from modeling_indepth import decode_to_rgba, generate, load_indepth from tokenizer import IndepthTokenizer model_dir = args.local or "DexagonAI/indepth-10M" model, config, resolved_dir = load_indepth(model_dir) tokenizer = IndepthTokenizer(resolved_dir) print(f"loaded {sum(p.numel() for p in model.parameters()):,} params, " f"vocab {len(tokenizer.stoi)} words, {config.subword_buckets} subword buckets") prompts = [args.prompt] if args.prompt else [l.strip() for l in Path(args.prompts).read_text().splitlines() if l.strip()] device = args.device model = model.to(device) token_rows, sub_rows = tokenizer.batch(prompts) import torch tokens = torch.tensor(token_rows, dtype=torch.long) subwords = torch.tensor(sub_rows, dtype=torch.long) image = generate(model, config, tokens, subwords, seed=args.seed, steps=args.steps, guidance=args.cfg, device=device) rgba = decode_to_rgba(image) out_dir = Path(args.out) out_dir.mkdir(parents=True, exist_ok=True) for i, prompt in enumerate(prompts): sprite = Image.fromarray(rgba[i], "RGBA").resize( (16 * args.scale, 16 * args.scale), Image.Resampling.NEAREST) slug = "-".join(w for w in __import__("re").split(r"[^a-z0-9]+", prompt.lower()) if w)[:40] or "sprite" path = out_dir / f"{slug}_{args.seed}_{i}.png" sprite.save(path) print("wrote", path) if __name__ == "__main__": main()