Initial PixelBoost AI upscaler (Real-ESRGAN)

.gitignore

@@ -0,0 +1,5 @@
+weights/
+__pycache__/
+*.pyc
+.gradio/
+flagged/

README.md

@@ -1,13 +1,35 @@
 ---
-title: Pixelboost Upscaler
-emoji: 👀
-colorFrom: red
+title: PixelBoost AI Upscaler
+emoji: ✨
+colorFrom: purple
 colorTo: pink
 sdk: gradio
-sdk_version: 6.14.0
-python_version: '3.13'
+sdk_version: 4.44.1
 app_file: app.py
 pinned: false
+license: mit
+short_description: Real-ESRGAN AI image upscaler for PixelBoost
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# PixelBoost AI Upscaler
+
+Real-ESRGAN (`realesr-general-x4v3`) behind a Gradio interface. This Space
+powers the **AI Enhance** mode of [PixelBoost](https://pixelboost-upscaler.pages.dev/).
+The Space exposes an `/upscale` API endpoint that the PixelBoost FastAPI
+backend calls via `gradio_client`.
+
+## Local dev
+
+```bash
+pip install -r requirements.txt
+python app.py
+```
+
+## Notes
+
+- Free CPU tier: expect 20-90s per image. Tiling (256px) keeps memory low.
+- Model weights download from the official Real-ESRGAN GitHub release on
+  first boot and are cached in `./weights`.
+- Source-of-truth lives in
+  [pixelboost-upscaler/hf-space](https://github.com/minhajulofficial/pixelboost-upscaler/tree/main/hf-space);
+  push changes from there.

app.py

@@ -0,0 +1,133 @@
+"""PixelBoost AI upscaler — HuggingFace Space (Gradio).
+
+Real-ESRGAN inference behind a Gradio interface. The PixelBoost backend
+(FastAPI) calls this Space's ``/upscale`` API endpoint when the user picks
+"AI Enhance" mode. The Gradio UI is also usable directly for manual testing.
+
+Model: realesr-general-x4v3 (5MB, SRVGG-based) — chosen for fast CPU
+inference on HF Spaces' free tier. Quality is excellent on photos with the
+mild denoise strength applied here.
+
+Scales:
+- 4x: native single forward pass through the 4x model.
+- 2x: 4x forward pass + LANCZOS downscale to 2x (cheaper than a separate 2x
+  model and produces sharper results than a pure 2x model on free CPU).
+- 6x: 4x forward pass + LANCZOS upscale to 6x.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+import types
+
+import gradio as gr
+import numpy as np
+import torch
+from PIL import Image
+
+# --- torchvision compat shim ---------------------------------------------------
+# basicsr (transitively via realesrgan) imports
+# ``torchvision.transforms.functional_tensor`` which was removed in
+# torchvision >= 0.17. Recreate the module pointing at ``functional`` before
+# basicsr is imported.
+try:
+    import torchvision.transforms.functional_tensor  # type: ignore[import-not-found]  # noqa: F401
+except ImportError:
+    from torchvision.transforms import functional as _tv_functional
+
+    _shim = types.ModuleType("torchvision.transforms.functional_tensor")
+    _shim.rgb_to_grayscale = _tv_functional.rgb_to_grayscale  # type: ignore[attr-defined]
+    sys.modules["torchvision.transforms.functional_tensor"] = _shim
+
+# --- model setup --------------------------------------------------------------
+from basicsr.archs.srvgg_arch import SRVGGNetCompact  # noqa: E402
+from basicsr.utils.download_util import load_file_from_url  # noqa: E402
+from realesrgan import RealESRGANer  # noqa: E402
+
+MODEL_NAME = "realesr-general-x4v3"
+MODEL_URL = (
+    "https://github.com/xinntao/Real-ESRGAN/releases/download/"
+    "v0.2.5.0/realesr-general-x4v3.pth"
+)
+WEIGHTS_DIR = os.environ.get("PIXELBOOST_WEIGHTS_DIR", "weights")
+TILE_SIZE = int(os.environ.get("PIXELBOOST_TILE_SIZE", "256"))
+MAX_INPUT_PIXELS = int(os.environ.get("PIXELBOOST_MAX_INPUT_PIXELS", str(4_000_000)))
+
+
+def _build_upscaler() -> RealESRGANer:
+    os.makedirs(WEIGHTS_DIR, exist_ok=True)
+    model_path = load_file_from_url(
+        url=MODEL_URL, model_dir=WEIGHTS_DIR, file_name=f"{MODEL_NAME}.pth"
+    )
+    arch = SRVGGNetCompact(
+        num_in_ch=3,
+        num_out_ch=3,
+        num_feat=64,
+        num_conv=32,
+        upscale=4,
+        act_type="prelu",
+    )
+    return RealESRGANer(
+        scale=4,
+        model_path=model_path,
+        model=arch,
+        tile=TILE_SIZE,
+        tile_pad=10,
+        pre_pad=0,
+        half=False,  # CPU does not support fp16
+        device=torch.device("cpu"),
+    )
+
+
+print(f"[pixelboost] loading {MODEL_NAME}...", flush=True)
+UPSAMPLER = _build_upscaler()
+print(f"[pixelboost] model loaded; tile={TILE_SIZE}", flush=True)
+
+
+# --- inference ----------------------------------------------------------------
+def upscale(image: Image.Image | None, scale: int = 4) -> Image.Image:
+    """Run Real-ESRGAN inference and resize to the requested scale."""
+    if image is None:
+        raise gr.Error("No image provided.")
+    if int(scale) not in {2, 4, 6}:
+        raise gr.Error("Scale must be 2, 4, or 6.")
+
+    image = image.convert("RGB")
+    if image.width * image.height > MAX_INPUT_PIXELS:
+        raise gr.Error(
+            f"Input too large ({image.width}x{image.height}). "
+            f"Max ~{MAX_INPUT_PIXELS // 1_000_000} megapixels in AI mode."
+        )
+
+    arr = np.array(image)
+    output, _ = UPSAMPLER.enhance(arr, outscale=4)
+    result = Image.fromarray(output)
+
+    target = (image.width * int(scale), image.height * int(scale))
+    if result.size != target:
+        result = result.resize(target, Image.Resampling.LANCZOS)
+    return result
+
+
+demo = gr.Interface(
+    fn=upscale,
+    inputs=[
+        gr.Image(type="pil", label="Input image"),
+        gr.Radio([2, 4, 6], value=4, label="Scale", type="value"),
+    ],
+    outputs=gr.Image(type="pil", label="Upscaled", format="png"),
+    title="PixelBoost AI Upscaler",
+    description=(
+        "Real-ESRGAN (realesr-general-x4v3) running on HuggingFace free CPU. "
+        "Expect 20-90s per image. Called from the PixelBoost backend when "
+        "users pick AI Enhance mode."
+    ),
+    api_name="upscale",
+    allow_flagging="never",
+    concurrency_limit=1,
+)
+demo.queue(max_size=20)
+
+if __name__ == "__main__":
+    demo.launch(show_api=True)

requirements.txt

@@ -0,0 +1,8 @@
+torch==2.2.2
+torchvision==0.17.2
+numpy<2.0
+pillow>=10.0.0
+opencv-python-headless>=4.8.0
+basicsr==1.4.2
+realesrgan==0.3.0
+gradio>=4.44.0