James040's picture
Update app.py
6ad12f0 verified
import os, cv2, numpy as np, onnxruntime as ort, gradio as gr
from huggingface_hub import hf_hub_download
# --- MODEL SETUP (APISR RRDB GAN - High Fidelity) ---
def load_model():
# Using APISR RRDB GAN for significantly better edge reconstruction than base Real-ESRGAN
path = hf_hub_download(repo_id="Xenova/2x_APISR_RRDB_GAN_generator-onnx", filename="onnx/model.onnx")
opts = ort.SessionOptions()
opts.intra_op_num_threads = 2
return ort.InferenceSession(path, opts, providers=['CPUExecutionProvider'])
session = load_model()
def upscale_image_tiled(frame, tile_size=128, overlap=16):
h, w, c = frame.shape
output_h, output_w = h * 2, w * 2
upscaled_img = np.zeros((output_h, output_w, c), dtype=np.uint8)
stride = tile_size - (overlap * 2)
for y in range(0, h, stride):
for x in range(0, w, stride):
y1, y2 = max(0, y - overlap), min(h, y + stride + overlap)
x1, x2 = max(0, x - overlap), min(w, x + stride + overlap)
tile = frame[y1:y2, x1:x2]
# --- MANDATORY APISR FIX: Pad to Multiple of 8 ---
th, tw = tile.shape[:2]
pad_h = (8 - (th % 8)) % 8
pad_w = (8 - (tw % 8)) % 8
if pad_h > 0 or pad_w > 0:
tile = cv2.copyMakeBorder(tile, 0, pad_h, 0, pad_w, cv2.BORDER_REFLECT)
# ------------------------------------------------
# AI Inference
img_input = cv2.cvtColor(tile, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
img_input = np.transpose(img_input, (2, 0, 1))[np.newaxis, :]
output = session.run(None, {session.get_inputs()[0].name: img_input})[0]
# Post-process tile
tile_out = np.clip(np.squeeze(output), 0, 1).transpose(1, 2, 0)
tile_out = (tile_out * 255.0).astype(np.uint8)
tile_out = cv2.cvtColor(tile_out, cv2.COLOR_RGB2BGR)
# Remove the AI padding and the overlap padding, then paste
# We only want the part corresponding to the original unpadded tile
tile_out = tile_out[:(th*2), :(tw*2)]
oy1, oy2 = (y - y1) * 2, (y2 - y) * 2
ox1, ox2 = (x - x1) * 2, (x2 - x) * 2
py1, py2 = y * 2, min(output_h, (y + stride) * 2)
px1, px2 = x * 2, min(output_w, (x + stride) * 2)
upscaled_img[py1:py2, px1:px2] = tile_out[oy1 : oy1 + (py2-py1), ox1 : ox1 + (px2-px1)]
return upscaled_img
def run_upscale(img_data, sharpen_amount):
if img_data is None: return None
img = img_data["composite"]
if img.shape[2] == 4: img = cv2.cvtColor(img, cv2.COLOR_RGBA2RGB)
# AI Upscale (2x APISR)
res = upscale_image_tiled(img)
# Sharpening for that "Pro" bite
if sharpen_amount > 0:
blurred = cv2.GaussianBlur(res, (0, 0), 3)
res = cv2.addWeighted(res, 1 + sharpen_amount, blurred, -sharpen_amount, 0)
return res
# --- UI ---
with gr.Blocks(theme=gr.themes.Default(primary_hue="blue")) as demo:
gr.Markdown("## ๐Ÿ’Ž Pro APISR-RRDB Upscaler (CPU Optimized)")
gr.Markdown("Uses the advanced APISR engine for cleaner edges and better texture restoration.")
with gr.Row():
with gr.Column():
img_in = gr.ImageEditor(label="Input (Crop allowed)", type="numpy")
sharp_slider = gr.Slider(0, 1, value=0.15, label="Sharpness Strength")
btn = gr.Button("UPSCALE 2X", variant="primary")
with gr.Column():
img_out = gr.Image(label="High Fidelity Result")
btn.click(run_upscale, [img_in, sharp_slider], img_out)
if __name__ == "__main__":
demo.queue().launch()