feat: unified Docker image with GPU auto-detection (#37)

Merge CPU, CUDA, and lite Docker images into a single unified image.
One tag (latest) works on all platforms: amd64 (NVIDIA CUDA) and arm64 (CPU).
GPU auto-detected at runtime. All ML models and packages baked in.

Key changes:
- Platform-conditional Dockerfile (nvidia/cuda on amd64, node on arm64)
- tini as PID 1 for proper signal handling
- Fix FILES_STORAGE_PATH data loss bug
- Fix RealESRGAN upscaler (was broken, always fell back to Lanczos)
- Fix PaddleOCR language codes and stdout corruption
- Simplified CI/CD (single build, single tag)
- Expanded model pre-download with verification
- Shutdown timeout, improved health endpoint
- Remove unused lama-cleaner

packages/ai/python/ocr.py

@@ -33,23 +33,39 @@ def run_tesseract(input_path, language):
 def run_paddleocr(input_path, language):
     """Run PaddleOCR."""
     os.environ["PADDLE_PDX_DISABLE_MODEL_SOURCE_CHECK"] = "True"
-    from paddleocr import PaddleOCR
-    from gpu import gpu_available
 
-    emit_progress(20, "Loading")
-    ocr = PaddleOCR(lang=language, use_gpu=gpu_available())
-    emit_progress(30, "Scanning")
-    result = ocr.ocr(input_path)
-    emit_progress(70, "Extracting text")
-    text = "\n".join(
-        [
-            line[1][0]
-            for res in result
-            if res
-            for line in res
-            if line and line[1]
-        ]
-    )
+    # Redirect stdout to stderr so PaddleOCR download/init messages
+    # cannot contaminate our JSON result on stdout.
+    stdout_fd = os.dup(1)
+    os.dup2(2, 1)
+
+    try:
+        from paddleocr import PaddleOCR
+        from gpu import gpu_available
+
+        # Map API language codes to PaddleOCR codes
+        paddle_lang_map = {"en": "en", "de": "latin", "fr": "latin", "es": "latin", "zh": "ch", "ja": "japan", "ko": "korean"}
+        paddle_lang = paddle_lang_map.get(language, "en")
+
+        emit_progress(20, "Loading")
+        ocr = PaddleOCR(lang=paddle_lang, use_gpu=gpu_available(), show_log=False)
+        emit_progress(30, "Scanning")
+        result = ocr.ocr(input_path)
+        emit_progress(70, "Extracting text")
+        text = "\n".join(
+            [
+                line[1][0]
+                for res in result
+                if res
+                for line in res
+                if line and line[1]
+            ]
+        )
+    finally:
+        # Restore stdout
+        os.dup2(stdout_fd, 1)
+        os.close(stdout_fd)
+
     return text, "paddleocr"
 
 

packages/ai/python/requirements-gpu.txt

@@ -1,6 +1,5 @@
 rembg==2.0.62
 realesrgan==0.3.0
-lama-cleaner==1.2.5
 paddleocr==2.9.1
 paddlepaddle-gpu==3.0.0
 mediapipe==0.10.21
@@ -8,3 +7,4 @@ onnxruntime-gpu==1.20.1
 numpy==1.26.4
 Pillow==11.1.0
 opencv-python-headless==4.10.0.84
+seam-carving==1.1.0

packages/ai/python/requirements.txt

@@ -1,6 +1,5 @@
 rembg[cpu]==2.0.62
 realesrgan==0.3.0
-lama-cleaner==1.2.5
 paddleocr==2.9.1
 paddlepaddle==3.0.0
 mediapipe==0.10.21

packages/ai/python/upscale.py

@@ -1,6 +1,7 @@
 """Image upscaling with Real-ESRGAN fallback to Lanczos."""
 import sys
 import json
+import os
 
 
 def emit_progress(percent, stage):
@@ -8,6 +9,12 @@ def emit_progress(percent, stage):
     print(json.dumps({"progress": percent, "stage": stage}), file=sys.stderr, flush=True)
 
 
+REALESRGAN_MODEL_PATH = os.environ.get(
+    "REALESRGAN_MODEL_PATH",
+    "/opt/models/realesrgan/RealESRGAN_x4plus.pth",
+)
+
+
 def main():
     input_path = sys.argv[1]
     output_path = sys.argv[2]
@@ -24,26 +31,40 @@ def main():
 
         # Try Real-ESRGAN first
         try:
-            from basicsr.archs.rrdbnet_arch import RRDBNet
-            from realesrgan import RealESRGANer
-            from gpu import gpu_available
-            import numpy as np
-            import torch
+            # Redirect stdout to stderr so basicsr/realesrgan init messages
+            # cannot contaminate our JSON result on stdout.
+            stdout_fd = os.dup(1)
+            os.dup2(2, 1)
+
+            try:
+                from basicsr.archs.rrdbnet_arch import RRDBNet
+                from realesrgan import RealESRGANer
+                from gpu import gpu_available
+                import numpy as np
+                import torch
+            finally:
+                # Restore stdout after imports
+                os.dup2(stdout_fd, 1)
+                os.close(stdout_fd)
+
+            if not os.path.exists(REALESRGAN_MODEL_PATH):
+                raise FileNotFoundError(f"RealESRGAN model not found: {REALESRGAN_MODEL_PATH}")
 
             use_gpu = gpu_available()
             device = torch.device("cuda" if use_gpu else "cpu")
 
+            # RealESRGAN_x4plus is a 4x model internally
             model = RRDBNet(
                 num_in_ch=3,
                 num_out_ch=3,
                 num_feat=64,
                 num_block=23,
                 num_grow_ch=32,
-                scale=scale,
+                scale=4,
             )
             upsampler = RealESRGANer(
-                scale=scale,
-                model_path=None,
+                scale=4,
+                model_path=REALESRGAN_MODEL_PATH,
                 model=model,
                 half=use_gpu,
                 device=device,
@@ -57,8 +78,8 @@ def main():
             emit_progress(95, "Saving result")
             result.save(output_path)
             method = "realesrgan"
-        except (ImportError, Exception):
-            # Fallback to Lanczos upscaling
+        except (ImportError, FileNotFoundError, RuntimeError, OSError):
+            # RealESRGAN unavailable or failed - fall back to Lanczos
             emit_progress(50, "Upscaling with Lanczos")
             img_upscaled = img.resize(new_size, Image.LANCZOS)
             emit_progress(95, "Saving result")