Smart upscale: auto-detect hardware and pick best Real-ESRGAN path
Detection priority (probed once, cached): 1. Real-ESRGAN PyTorch + CUDA GPU → fastest, best quality 2. Real-ESRGAN PyTorch + Apple MPS → fast on Apple Silicon 3. Real-ESRGAN NCNN Vulkan binary → fast on any GPU via Vulkan (no CUDA needed) 4. Real-ESRGAN PyTorch CPU → works, slow (warned in UI) 5. Lanczos → always available, instant fallback Backend: - services/upscale.py: full capability probe (probe_upscale_capabilities), implementations for PyTorch (CUDA/MPS/CPU auto-device) and NCNN binary, upscale_sync() resolves method with fallback chain, async upscale_image() runs in thread pool - print_tools.py: /api/print/upscale uses new service; method="auto" by default; GET /api/print/upscale/available returns full capability map with device info and recommended_label; POST /api/print/upscale/refresh-caps busts cache without restart (useful after installing NCNN binary into container) Frontend: - upscale.js: fetches capability map on first open; builds method selector showing only available options; labels recommended method with ★; shows device info (CUDA/MPS/CPU/NCNN) in dialog; maps display label back to method key on submit; shows actual method used in success toast and undo history entry Scripts: - scripts/download_realesrgan.py: downloads NCNN Vulkan binary for current platform (Linux/macOS/Windows) to /app/data/models/realesrgan/; makes executable; run inside container or locally https://claude.ai/code/session_01B58MaJCU1R6KwBDJCp8AfN
This commit is contained in:
Claude
@@ -61,7 +61,8 @@ class FrameFitRequest(BaseModel):
|
||||
class UpscaleRequest(BaseModel):
|
||||
image: str # base64
|
||||
scale: float = 2.0 # 1.5, 2, 3, 4
|
||||
method: Literal["lanczos", "ai"] = "lanczos"
|
||||
# auto = pick best available; lanczos = always works; realesrgan_pytorch / realesrgan_ncnn = explicit
|
||||
method: str = "auto"
|
||||
|
||||
|
||||
# ── Frame sizes endpoint ───────────────────────────────────────────────────
|
||||
@@ -293,83 +294,63 @@ def _mirror_fill(canvas, mask, scaled, gap_dir, gap_a, gap_b, target_w, target_h
|
||||
|
||||
# ── Upscale ────────────────────────────────────────────────────────────────
|
||||
|
||||
@router.post("/upscale/refresh-caps")
|
||||
def upscale_refresh_caps():
|
||||
"""Bust the capability cache (call after installing Real-ESRGAN without restarting)."""
|
||||
from app.services.upscale import invalidate_caps_cache, probe_upscale_capabilities
|
||||
invalidate_caps_cache()
|
||||
return probe_upscale_capabilities()
|
||||
|
||||
|
||||
@router.get("/upscale/available")
|
||||
def upscale_available():
|
||||
"""
|
||||
Return capability probe: which upscale methods are available,
|
||||
which device will be used, and which method is recommended.
|
||||
Frontend uses this to populate the method selector.
|
||||
"""
|
||||
from app.services.upscale import probe_upscale_capabilities
|
||||
caps = probe_upscale_capabilities()
|
||||
return caps
|
||||
|
||||
|
||||
@router.post("/upscale")
|
||||
async def upscale(req: UpscaleRequest):
|
||||
"""
|
||||
Upscale image.
|
||||
method=lanczos — always available, fast, good for clean images
|
||||
method=ai — Real-ESRGAN if installed, else falls back to lanczos
|
||||
Upscale image. method values:
|
||||
auto — pick best available (recommended)
|
||||
realesrgan_pytorch — Real-ESRGAN via PyTorch (CUDA/MPS/CPU)
|
||||
realesrgan_ncnn — Real-ESRGAN NCNN Vulkan binary
|
||||
lanczos — always available, instant
|
||||
Any AI method falls back to the next best if unavailable.
|
||||
"""
|
||||
if not (1.1 <= req.scale <= 8.0):
|
||||
raise HTTPException(status_code=400, detail="scale must be 1.1–8.0")
|
||||
|
||||
valid_methods = {"auto", "realesrgan_pytorch", "realesrgan_ncnn", "lanczos"}
|
||||
if req.method not in valid_methods:
|
||||
raise HTTPException(status_code=400,
|
||||
detail=f"method must be one of {sorted(valid_methods)}")
|
||||
|
||||
try:
|
||||
image = Image.open(BytesIO(_decode(req.image))).convert("RGB")
|
||||
except Exception as e:
|
||||
raise HTTPException(status_code=400, detail=f"Could not decode image: {e}")
|
||||
|
||||
orig_w, orig_h = image.size
|
||||
new_w = round(orig_w * req.scale)
|
||||
new_h = round(orig_h * req.scale)
|
||||
|
||||
method_used = req.method
|
||||
|
||||
if req.method == "ai":
|
||||
try:
|
||||
result_bytes = await asyncio.get_event_loop().run_in_executor(
|
||||
None, _realesrgan_upscale, image, req.scale
|
||||
)
|
||||
result = Image.open(BytesIO(result_bytes)).convert("RGB")
|
||||
method_used = "realesrgan"
|
||||
except Exception as e:
|
||||
print(f"Real-ESRGAN failed, using Lanczos: {e}")
|
||||
result = image.resize((new_w, new_h), Image.Resampling.LANCZOS)
|
||||
method_used = "lanczos_fallback"
|
||||
else:
|
||||
result = image.resize((new_w, new_h), Image.Resampling.LANCZOS)
|
||||
try:
|
||||
from app.services.upscale import upscale_image
|
||||
result_bytes, method_used = await upscale_image(image, req.scale, req.method)
|
||||
result = Image.open(BytesIO(result_bytes))
|
||||
except Exception as e:
|
||||
import traceback; traceback.print_exc()
|
||||
raise HTTPException(status_code=500, detail=str(e))
|
||||
|
||||
return {
|
||||
"result": _encode(_to_png(result)),
|
||||
"result": _encode(result_bytes),
|
||||
"method": method_used,
|
||||
"original": {"width": orig_w, "height": orig_h},
|
||||
"output": {"width": result.width, "height": result.height},
|
||||
"scale": req.scale,
|
||||
"output": {"width": result.width, "height": result.height},
|
||||
"scale": req.scale,
|
||||
}
|
||||
|
||||
|
||||
def _realesrgan_upscale(image: Image.Image, scale: float) -> bytes:
|
||||
"""Run Real-ESRGAN upscaling. Raises if not installed."""
|
||||
from basicsr.archs.rrdbnet_arch import RRDBNet
|
||||
from realesrgan import RealESRGANer
|
||||
import torch
|
||||
import numpy as np
|
||||
|
||||
model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64,
|
||||
num_block=23, num_grow_ch=32, scale=4)
|
||||
upsampler = RealESRGANer(
|
||||
scale=4,
|
||||
model_path=None, # auto-download
|
||||
model=model,
|
||||
tile=400,
|
||||
tile_pad=10,
|
||||
pre_pad=0,
|
||||
half=torch.cuda.is_available(),
|
||||
)
|
||||
img_np = np.array(image)[:, :, ::-1] # RGB→BGR for cv2
|
||||
output, _ = upsampler.enhance(img_np, outscale=scale)
|
||||
result = Image.fromarray(output[:, :, ::-1]) # BGR→RGB
|
||||
buf = BytesIO()
|
||||
result.save(buf, format="PNG")
|
||||
return buf.getvalue()
|
||||
|
||||
|
||||
@router.get("/upscale/available")
|
||||
def upscale_available():
|
||||
"""Check which upscale methods are available."""
|
||||
ai_available = False
|
||||
try:
|
||||
import realesrgan # noqa: F401
|
||||
ai_available = True
|
||||
except ImportError:
|
||||
pass
|
||||
return {"lanczos": True, "realesrgan": ai_available}
|
||||
|
||||
@@ -0,0 +1,300 @@
|
||||
"""
|
||||
Upscale service — auto-detects best available method and runs it.
|
||||
|
||||
Priority (auto mode):
|
||||
1. Real-ESRGAN PyTorch + CUDA GPU — fastest, best quality
|
||||
2. Real-ESRGAN PyTorch + Apple MPS — fast on Apple Silicon
|
||||
3. Real-ESRGAN NCNN Vulkan binary — fast on any GPU (Intel/AMD/integrated)
|
||||
4. Real-ESRGAN PyTorch CPU — works, slow (warn user)
|
||||
5. Lanczos — always available, instant
|
||||
|
||||
Capability probe is run once at first call and cached.
|
||||
"""
|
||||
|
||||
import asyncio
|
||||
import os
|
||||
import shutil
|
||||
import subprocess
|
||||
import sys
|
||||
import tempfile
|
||||
from io import BytesIO
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
from PIL import Image
|
||||
|
||||
# ── Capability detection ──────────────────────────────────────────────────────
|
||||
|
||||
_caps: Optional[dict] = None
|
||||
|
||||
|
||||
def probe_upscale_capabilities() -> dict:
|
||||
"""
|
||||
Detect what upscaling hardware and software is available.
|
||||
Result is cached after first call.
|
||||
"""
|
||||
global _caps
|
||||
if _caps is not None:
|
||||
return _caps
|
||||
|
||||
caps = {
|
||||
"lanczos": True,
|
||||
"realesrgan_pytorch": False,
|
||||
"realesrgan_pytorch_device": None, # "cuda" | "mps" | "cpu"
|
||||
"realesrgan_ncnn": False,
|
||||
"realesrgan_ncnn_path": None,
|
||||
"recommended": "lanczos",
|
||||
"recommended_label": "Lanczos (no AI upscaler found)",
|
||||
"methods": ["lanczos"],
|
||||
}
|
||||
|
||||
# ── PyTorch path ──────────────────────────────────────────────────────────
|
||||
pytorch_device = None
|
||||
try:
|
||||
import torch
|
||||
if torch.cuda.is_available():
|
||||
pytorch_device = "cuda"
|
||||
elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
|
||||
pytorch_device = "mps"
|
||||
else:
|
||||
pytorch_device = "cpu"
|
||||
except ImportError:
|
||||
pass
|
||||
|
||||
if pytorch_device:
|
||||
try:
|
||||
import realesrgan # noqa: F401
|
||||
from basicsr.archs.rrdbnet_arch import RRDBNet # noqa: F401
|
||||
caps["realesrgan_pytorch"] = True
|
||||
caps["realesrgan_pytorch_device"] = pytorch_device
|
||||
caps["methods"].append("realesrgan_pytorch")
|
||||
except ImportError:
|
||||
pass
|
||||
|
||||
# ── NCNN Vulkan binary ────────────────────────────────────────────────────
|
||||
ncnn_path = _find_ncnn_binary()
|
||||
if ncnn_path:
|
||||
caps["realesrgan_ncnn"] = True
|
||||
caps["realesrgan_ncnn_path"] = str(ncnn_path)
|
||||
caps["methods"].append("realesrgan_ncnn")
|
||||
|
||||
# ── Pick recommended ──────────────────────────────────────────────────────
|
||||
if caps["realesrgan_pytorch"] and pytorch_device in ("cuda", "mps"):
|
||||
device_label = "CUDA GPU" if pytorch_device == "cuda" else "Apple Silicon"
|
||||
caps["recommended"] = "realesrgan_pytorch"
|
||||
caps["recommended_label"] = f"Real-ESRGAN ({device_label})"
|
||||
elif caps["realesrgan_ncnn"]:
|
||||
caps["recommended"] = "realesrgan_ncnn"
|
||||
caps["recommended_label"] = "Real-ESRGAN NCNN (Vulkan)"
|
||||
elif caps["realesrgan_pytorch"] and pytorch_device == "cpu":
|
||||
caps["recommended"] = "realesrgan_pytorch"
|
||||
caps["recommended_label"] = "Real-ESRGAN (CPU — may be slow)"
|
||||
else:
|
||||
caps["recommended"] = "lanczos"
|
||||
caps["recommended_label"] = "Lanczos (install Real-ESRGAN for AI quality)"
|
||||
|
||||
_caps = caps
|
||||
return caps
|
||||
|
||||
|
||||
def _find_ncnn_binary() -> Optional[Path]:
|
||||
"""Find realesrgan-ncnn-vulkan binary on the system."""
|
||||
# Check PATH first
|
||||
found = shutil.which("realesrgan-ncnn-vulkan")
|
||||
if found:
|
||||
return Path(found)
|
||||
|
||||
# Check known install locations
|
||||
candidates = [
|
||||
Path("/app/data/models/realesrgan/realesrgan-ncnn-vulkan"),
|
||||
Path("/usr/local/bin/realesrgan-ncnn-vulkan"),
|
||||
Path.home() / ".local/bin/realesrgan-ncnn-vulkan",
|
||||
# Windows
|
||||
Path(r"C:/realesrgan-ncnn-vulkan/realesrgan-ncnn-vulkan.exe"),
|
||||
# macOS Homebrew
|
||||
Path("/opt/homebrew/bin/realesrgan-ncnn-vulkan"),
|
||||
Path("/usr/local/bin/realesrgan-ncnn-vulkan"),
|
||||
]
|
||||
for p in candidates:
|
||||
if p.exists() and os.access(p, os.X_OK):
|
||||
return p
|
||||
|
||||
return None
|
||||
|
||||
|
||||
def invalidate_caps_cache():
|
||||
"""Call after installing new software so next probe picks it up."""
|
||||
global _caps
|
||||
_caps = None
|
||||
|
||||
|
||||
# ── Upscale implementations ───────────────────────────────────────────────────
|
||||
|
||||
def _to_png_bytes(img: Image.Image) -> bytes:
|
||||
buf = BytesIO()
|
||||
img.save(buf, format="PNG")
|
||||
return buf.getvalue()
|
||||
|
||||
|
||||
def upscale_lanczos(image: Image.Image, scale: float) -> tuple[bytes, str]:
|
||||
"""Pure Pillow Lanczos — instant, always available."""
|
||||
new_w = round(image.width * scale)
|
||||
new_h = round(image.height * scale)
|
||||
result = image.resize((new_w, new_h), Image.Resampling.LANCZOS)
|
||||
return _to_png_bytes(result), "lanczos"
|
||||
|
||||
|
||||
def upscale_realesrgan_pytorch(image: Image.Image, scale: float) -> tuple[bytes, str]:
|
||||
"""
|
||||
Real-ESRGAN via PyTorch.
|
||||
Uses CUDA > MPS > CPU automatically based on what's available.
|
||||
Scale factors: any float — upscales to nearest 2x or 4x model, then resizes to exact target.
|
||||
"""
|
||||
import torch
|
||||
from basicsr.archs.rrdbnet_arch import RRDBNet
|
||||
from realesrgan import RealESRGANer
|
||||
|
||||
caps = probe_upscale_capabilities()
|
||||
device = caps.get("realesrgan_pytorch_device", "cpu")
|
||||
|
||||
# Choose model: x2 for scale <= 2.5, x4 otherwise
|
||||
model_scale = 2 if scale <= 2.5 else 4
|
||||
model = RRDBNet(
|
||||
num_in_ch=3, num_out_ch=3, num_feat=64,
|
||||
num_block=23, num_grow_ch=32, scale=model_scale
|
||||
)
|
||||
|
||||
# Model path: check local cache first, then let RealESRGANer auto-download
|
||||
model_dir = Path("/app/data/models/realesrgan")
|
||||
model_dir.mkdir(parents=True, exist_ok=True)
|
||||
model_name = f"RealESRGAN_x{model_scale}plus.pth"
|
||||
model_path = model_dir / model_name
|
||||
if not model_path.exists():
|
||||
model_path = None # RealESRGANer will download to its default cache
|
||||
|
||||
upsampler = RealESRGANer(
|
||||
scale=model_scale,
|
||||
model_path=str(model_path) if model_path else None,
|
||||
model=model,
|
||||
tile=512,
|
||||
tile_pad=10,
|
||||
pre_pad=0,
|
||||
half=(device == "cuda"), # fp16 only on CUDA
|
||||
device=torch.device(device),
|
||||
)
|
||||
|
||||
import numpy as np
|
||||
img_bgr = np.array(image)[:, :, ::-1].copy() # RGB→BGR
|
||||
enhanced, _ = upsampler.enhance(img_bgr, outscale=scale)
|
||||
result = Image.fromarray(enhanced[:, :, ::-1]) # BGR→RGB
|
||||
|
||||
label = f"realesrgan_pytorch_{device}"
|
||||
return _to_png_bytes(result), label
|
||||
|
||||
|
||||
def upscale_realesrgan_ncnn(image: Image.Image, scale: float) -> tuple[bytes, str]:
|
||||
"""
|
||||
Real-ESRGAN via NCNN Vulkan binary — works on any GPU (Intel/AMD/integrated/Apple).
|
||||
Runs as subprocess with temp file I/O.
|
||||
"""
|
||||
caps = probe_upscale_capabilities()
|
||||
binary = caps.get("realesrgan_ncnn_path")
|
||||
if not binary:
|
||||
raise RuntimeError("realesrgan-ncnn-vulkan binary not found")
|
||||
|
||||
# NCNN only supports integer scales (2, 3, 4) natively
|
||||
# For non-integer scales: upscale to nearest integer, then resize to exact target
|
||||
model_scale = 4 if scale > 2.5 else 2
|
||||
target_w = round(image.width * scale)
|
||||
target_h = round(image.height * scale)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdir:
|
||||
in_path = Path(tmpdir) / "input.png"
|
||||
out_path = Path(tmpdir) / "output.png"
|
||||
|
||||
image.save(in_path, format="PNG")
|
||||
|
||||
# Model name for NCNN (bundled with binary)
|
||||
model_name = f"realesrgan-x{model_scale}plus"
|
||||
|
||||
cmd = [
|
||||
binary,
|
||||
"-i", str(in_path),
|
||||
"-o", str(out_path),
|
||||
"-s", str(model_scale),
|
||||
"-n", model_name,
|
||||
"-f", "png",
|
||||
]
|
||||
|
||||
result_proc = subprocess.run(
|
||||
cmd, capture_output=True, timeout=300
|
||||
)
|
||||
if result_proc.returncode != 0:
|
||||
raise RuntimeError(
|
||||
f"realesrgan-ncnn-vulkan failed: {result_proc.stderr.decode()}"
|
||||
)
|
||||
|
||||
result = Image.open(out_path).convert("RGB")
|
||||
|
||||
# Resize to exact target if scale was non-integer
|
||||
if result.width != target_w or result.height != target_h:
|
||||
result = result.resize((target_w, target_h), Image.Resampling.LANCZOS)
|
||||
|
||||
return _to_png_bytes(result), "realesrgan_ncnn"
|
||||
|
||||
|
||||
# ── Public entry point ────────────────────────────────────────────────────────
|
||||
|
||||
def upscale_sync(image: Image.Image, scale: float, method: str = "auto") -> tuple[bytes, str]:
|
||||
"""
|
||||
Upscale image synchronously. Call via run_in_executor from async context.
|
||||
|
||||
method values:
|
||||
"auto" — pick best available automatically
|
||||
"realesrgan_pytorch" — force PyTorch path
|
||||
"realesrgan_ncnn" — force NCNN binary path
|
||||
"lanczos" — force Lanczos
|
||||
|
||||
Returns (png_bytes, method_used_label).
|
||||
"""
|
||||
caps = probe_upscale_capabilities()
|
||||
|
||||
if method == "auto":
|
||||
method = caps["recommended"]
|
||||
|
||||
if method == "realesrgan_pytorch":
|
||||
if caps["realesrgan_pytorch"]:
|
||||
try:
|
||||
return upscale_realesrgan_pytorch(image, scale)
|
||||
except Exception as e:
|
||||
print(f"Real-ESRGAN PyTorch failed, falling back: {e}")
|
||||
# Fall through to next best
|
||||
if caps["realesrgan_ncnn"]:
|
||||
try:
|
||||
return upscale_realesrgan_ncnn(image, scale)
|
||||
except Exception as e:
|
||||
print(f"Real-ESRGAN NCNN fallback failed: {e}")
|
||||
return upscale_lanczos(image, scale)
|
||||
|
||||
if method == "realesrgan_ncnn":
|
||||
if caps["realesrgan_ncnn"]:
|
||||
try:
|
||||
return upscale_realesrgan_ncnn(image, scale)
|
||||
except Exception as e:
|
||||
print(f"Real-ESRGAN NCNN failed, falling back: {e}")
|
||||
# Fall through
|
||||
if caps["realesrgan_pytorch"]:
|
||||
try:
|
||||
return upscale_realesrgan_pytorch(image, scale)
|
||||
except Exception as e:
|
||||
print(f"Real-ESRGAN PyTorch fallback failed: {e}")
|
||||
return upscale_lanczos(image, scale)
|
||||
|
||||
# Default / lanczos
|
||||
return upscale_lanczos(image, scale)
|
||||
|
||||
|
||||
async def upscale_image(image: Image.Image, scale: float, method: str = "auto") -> tuple[bytes, str]:
|
||||
"""Async wrapper — runs upscale in thread pool to avoid blocking the event loop."""
|
||||
loop = asyncio.get_event_loop()
|
||||
return await loop.run_in_executor(None, upscale_sync, image, scale, method)
|
||||
Reference in New Issue
Block a user