""" Generate an image from a prompt with WVY-Image-Composition-101. python3 generate.py "a cat on top of a dog on a green background" python3 generate.py "a fish next to a pig on a purple background" -o fish_pig.png Prompts must follow one of the two sentence shapes it was trained on: a on top of a on a background a next to a on a background animals: dog cat bird fish pig rabbit colors: white black red orange yellow green blue purple Needs only numpy + Pillow. """ import sys, json, argparse import numpy as np from PIL import Image import model as M ap = argparse.ArgumentParser() ap.add_argument("prompt") ap.add_argument("-o", "--out", default="out.png") ap.add_argument("--scale", type=int, default=8, help="pixel upscale for viewing (model draws 48x48)") a = ap.parse_args() vocab = json.load(open("train_log.json"))["vocab"] w2i = {w: i for i, w in enumerate(vocab)} words = a.prompt.lower().strip().split() unknown = [w for w in words if w not in w2i] if unknown or len(words) > M.L: sys.exit(f"Can't read that prompt. Unknown words: {unknown}. Use: a on top of / next to a on a background") T = np.zeros((1, M.L), dtype=np.int64); T[0, :len(words)] = [w2i[w] for w in words] p = dict(np.load("wvy_compose_weights.npz")) y = M.forward(p, T)[0].reshape(48, 48, 3) Image.fromarray((np.clip(y, 0, 1) * 255).astype("uint8")).resize((48 * a.scale, 48 * a.scale), Image.NEAREST).save(a.out) print("saved", a.out)