Spaces:
Running
Running
Download app.py from FrameXlabs/Trout: direct link, hf CLI and curl.
- Browser
- Download file 5.53 kB
-
https://huggingface.co/spaces/FrameXlabs/Trout/resolve/main/app.py
- Command line
-
hf download hf://spaces/FrameXlabs/Trout/app.py
-
curl -L -o app.py https://huggingface.co/spaces/FrameXlabs/Trout/resolve/main/app.py
5.53 kB
| """Trout Pixel Studio -- interactive demo for Trout. | |
| Trout is a 47.6M-parameter absorbing-state masked-diffusion language model for | |
| 16x16 pixel art, built and trained from scratch (architecture after | |
| Tessera-small by TheJagStudio; trained on Angais' Pixel Art Atlas 1.4M). | |
| A sprite is 256 tokens over a learned 512-color palette; the model unmasks all | |
| pixels in parallel over ~16 steps, conditioned on a frozen CLIP text tower. | |
| Runs on free CPU hardware: expect ~10-30 s per generation depending on | |
| steps/samples. | |
| """ | |
| import os | |
| import sys | |
| import time | |
| import torch | |
| from huggingface_hub import hf_hub_download | |
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) | |
| from trout import (DEFAULT_TEXT_ENCODER, FrozenTextEncoder, | |
| load_model_from_checkpoint, make_grid, sample, | |
| tokens_to_images, upscale) | |
| import gradio as gr | |
| MODEL_REPO = os.environ.get("MODEL_REPO", "FrameXlabs/Trout") | |
| DEVICE = "cpu" | |
| print(f"loading Trout from {MODEL_REPO} ...", flush=True) | |
| ckpt_path = hf_hub_download(MODEL_REPO, "trout-fp16.pt") | |
| model, palette, ck = load_model_from_checkpoint(ckpt_path, device=DEVICE, use_ema=True) | |
| enc = FrozenTextEncoder(ck.get("text_encoder", DEFAULT_TEXT_ENCODER), | |
| model.cfg.text_len, DEVICE) | |
| print(f"Trout ready: {model.num_params():,} parameters, step {ck.get('step')}", flush=True) | |
| EXAMPLES = [ | |
| "red health potion bottle", "golden crown with blue gems", "green slime monster", | |
| "iron sword with a leather hilt", "wooden treasure chest", "pink cherry blossom tree", | |
| "blue magic spellbook", "orange fox head", "glowing purple crystal", "cheeseburger", | |
| "skull with a candle", "rocket ship", "brown leather boots", "yellow rubber duck", | |
| "ice cream cone with strawberry scoop", "black cat", "tiny green dragon", "gold coin", | |
| ] | |
| HEADER = """ | |
| <div style="text-align:center; margin-bottom: 8px;"> | |
| <h1 style="margin-bottom:4px;">🐟 Trout Pixel Studio</h1> | |
| <p style="color:#666; font-size:15px;"> | |
| A <b>diffusion language model</b> for 16×16 pixel art — each sprite is a | |
| 256-token “sentence” over a learned 512-color palette, unmasked in parallel | |
| over ~16 steps. 47.6M parameters, built & trained from scratch on | |
| <b>Pixel Art Atlas 1.4M</b> (Angais, CC0), conditioned on a frozen | |
| CLIP text tower. Runs on CPU — please allow ~10–30 s per batch. | |
| </p> | |
| </div> | |
| """ | |
| def generate(prompt: str, num: int, steps: int, cfg: float, temperature: float, | |
| seed: int, polish: int): | |
| t0 = time.time() | |
| seed = int(seed) if seed is not None else 0 | |
| num = int(num) | |
| g = torch.Generator().manual_seed(seed) | |
| with torch.no_grad(): | |
| states, pooled = enc.encode([prompt] * num) | |
| toks = sample(model, states, pooled, steps=int(steps), cfg_scale=float(cfg), | |
| temperature=float(temperature), polish_steps=int(polish), | |
| generator=g) | |
| imgs = tokens_to_images(toks, palette) | |
| big = [upscale(im, 16, checker=True) for im in imgs] | |
| grid = make_grid(imgs, ncols=max(2, num), scale=10).convert("RGB") | |
| dt = time.time() - t0 | |
| info = (f"{num} sprite{'s' if num > 1 else ''} · {int(steps)} steps · " | |
| f"cfg {float(cfg):.1f} · {dt:.1f}s on CPU") | |
| return big, grid, info | |
| def build_ui(): | |
| with gr.Blocks(title="Trout Pixel Studio", theme=gr.themes.Soft()) as demo: | |
| gr.HTML(HEADER) | |
| with gr.Row(): | |
| with gr.Column(scale=5): | |
| prompt = gr.Textbox(label="Prompt", value="red health potion bottle", | |
| placeholder="describe a 16x16 game sprite...", | |
| lines=1, max_lines=2) | |
| with gr.Row(): | |
| num = gr.Slider(1, 8, value=4, step=1, label="Samples") | |
| steps = gr.Slider(4, 32, value=16, step=1, label="Diffusion steps") | |
| with gr.Row(): | |
| cfg = gr.Slider(1.0, 8.0, value=3.0, step=0.1, label="CFG scale") | |
| temperature = gr.Slider(0.0, 1.5, value=1.0, step=0.05, | |
| label="Temperature") | |
| with gr.Row(): | |
| seed = gr.Number(value=1234, label="Seed", precision=0) | |
| polish = gr.Slider(0, 2, value=0, step=1, label="Polish steps") | |
| go = gr.Button("🎣 Generate pixel art", variant="primary") | |
| examples = gr.Dataset(components=[gr.Textbox(visible=False)], | |
| samples=[[e] for e in EXAMPLES], | |
| label="Example prompts — click one to use it") | |
| examples.click(lambda s: (s[0] if isinstance(s, (list, tuple)) else s), | |
| inputs=[examples], outputs=[prompt]) | |
| with gr.Column(scale=4): | |
| info = gr.Markdown("") | |
| grid = gr.Image(label="Sprites grid", type="pil", height=420) | |
| gallery = gr.Gallery(label="Individual sprites (upscaled ×16)", | |
| columns=4, height=420, object_fit="contain") | |
| go.click(generate, | |
| inputs=[prompt, num, steps, cfg, temperature, seed, polish], | |
| outputs=[gallery, grid, info]) | |
| prompt.submit(generate, | |
| inputs=[prompt, num, steps, cfg, temperature, seed, polish], | |
| outputs=[gallery, grid, info]) | |
| return demo | |
| demo = build_ui() | |
| if os.environ.get("SPACE_TEST"): | |
| print("SPACE_TEST: UI built OK") | |
| else: | |
| demo.launch() | |