Upload folder using huggingface_hub

README.md

@@ -19,14 +19,24 @@ tags:
 
 Benvenuto su **Universal Upscale**, l'applicazione per l'upscaling e super-resolution di immagini a qualsiasi risoluzione o fattore di scala.
 
+Questo Space esegue il checkpoint **LIIF v4 GAN** (4.60 M parametri): un unico modello a scala
+arbitraria, esportato come `model.safetensors` + `model_scripted.pt` nel repo
+[tailot/universal-upscale](https://huggingface.co/tailot/universal-upscale).
+
 ## Come usare lo Space / How to use this Space
 
 1. **Carica un'immagine / Upload an image**: trascina o carica un'immagine dal tuo dispositivo.
 2. **Imposta la scala / Set the scale**:
-   - **Scale Factor**: scegli un fattore di scala da **1x fino a 10x** (es. 2x, 4x, 10x).
+   - **Scale Factor**: scegli un fattore di scala da **1x fino a 9.2x** (passo 0.2x, es. 2x, 3.6x, 4x, 9.2x).
    - **Target Long Side (px)**: in alternativa, specifica la dimensione in pixel del lato più lungo desiderato (es. 2048, 3000, 4096).
 3. **Genera / Upscale**: clicca su **Upscale Image** per ottenere l'immagine ad alta risoluzione.
 
+> **Limite di scala testato / Tested scale limit**: 9.2x è il valore massimo del cursore
+> ed il fattore di scala più alto raccomandato per questi pesi (la tabella completa del benchmark,
+> fino a 10x, si trova nel model card del modello). Richieste oltre questo valore vengono riportate
+> a 9.2x dal cursore; usando "Target Long Side" puoi comunque arrivare a 4096 px sul lato
+> lungo.
+
 ## Caratteristiche principali / Features
 - **Scale Arbitrarie (LIIF)**: Ingrandimento continuo a qualsiasi fattore o dimensione target senza vincoli rigidi (es. solo 2x/4x).
 - **Tiled Inference**: Elaborazione a tasselli sovrapposti per gestire immagini ad alta risoluzione senza esaurire la memoria (OOM).

app.py

@@ -37,6 +37,20 @@ def load_model(device: str):
 
 
 MAX_SIDE = 4096
+# Largest scale factor verified end-to-end for the published checkpoint. The slider is capped
+# here; "Target Long Side" still accepts any size up to MAX_SIDE.
+MAX_SCALE = 9.2
+
+
+def resolve_factor(input_size, scale, target_long_side):
+    """Scale factor to apply: an explicit target size wins, scale requests stay under MAX_SCALE."""
+    target = float(target_long_side) if target_long_side is not None else 0.0
+    if target > 0:
+        return target / float(max(input_size)), ""
+    factor = float(scale if scale is not None else 4.0)
+    if factor > MAX_SCALE:
+        return MAX_SCALE, f" (clamped to the tested maximum {MAX_SCALE:g}x)"
+    return factor, ""
 
 
 @gpu_decorator
@@ -46,10 +60,7 @@ def upscale(image, scale: float = 4.0, target_long_side: float = 0.0):
     img = image if isinstance(image, Image.Image) else Image.open(image)
     img = img.convert("RGB")
 
-    if target_long_side is not None and float(target_long_side) > 0:
-        factor = float(target_long_side) / float(max(img.size))
-    else:
-        factor = float(scale if scale is not None else 4.0)
+    factor, clamped = resolve_factor(img.size, scale, target_long_side)
 
     out_w, out_h = max(1, int(round(img.width * factor))), max(1, int(round(img.height * factor)))
     if max(out_w, out_h) > MAX_SIDE:
@@ -64,7 +75,7 @@ def upscale(image, scale: float = 4.0, target_long_side: float = 0.0):
         out = model(lr, out_h, out_w)
     arr = (out.squeeze(0).permute(1, 2, 0).cpu().numpy() * 255.0).round().astype(np.uint8)
     res_img = Image.fromarray(arr)
-    status = f"Input: {img.width}x{img.height} | Output: {res_img.width}x{res_img.height} (Scale: {factor:.2f}x) | Device: {device}"
+    status = f"Input: {img.width}x{img.height} | Output: {res_img.width}x{res_img.height} (Scale: {factor:.2f}x) | Device: {device}{clamped}"
     return res_img, status
 
 
@@ -78,7 +89,7 @@ demo = gr.Interface(
     fn=upscale,
     inputs=[
         gr.Image(type="pil", label="Input Image"),
-        gr.Slider(minimum=1.0, maximum=10.0, value=4.0, step=0.5, label="Scale Factor"),
+        gr.Slider(minimum=1.0, maximum=9.2, value=4.0, step=0.2, label="Scale Factor"),
         gr.Number(value=0, label="Or Target Long Side (px) [0 = use scale factor]"),
     ],
     outputs=[
@@ -87,8 +98,9 @@ demo = gr.Interface(
     ],
     title="Universal Upscale",
     description=(
-        "Arbitrary-scale super-resolution: upscale images to any scale factor (1x-10x) or exact target resolution. "
-        "Built-in tiled inference enables high-resolution upscaling without memory issues."
+        "Arbitrary-scale super-resolution: upscale images to any scale factor from 1x up to the tested maximum "
+        "9.2x, or to an exact target resolution. Built-in tiled inference enables high-resolution "
+        "upscaling without memory issues (outputs are capped at 4096 px on the long side)."
     ),
     examples=example_files if example_files else None,
 )

config.json

@@ -11,10 +11,11 @@
     "arbitrary": true,
     "trained_range": [
       1.5,
-      10.0
-    ]
+      6.0
+    ],
+    "tested_max": 9.2
   },
-  "parameters": 1261955,
+  "parameters": 4603897,
   "input": {
     "format": "RGB float32 tensor, shape (B, 3, H, W), values in [0, 1]"
   },
@@ -25,22 +26,37 @@
     "signature": "model(lr, target_h, target_w)",
     "tile_size": 256,
     "overlap": 16,
-    "tiling": "overlapping tiles + feather blending, compiled inside model_scripted.pt"
+    "tiling": "overlapping tiles + feather blending, compiled inside model_scripted.pt",
+    "max_scale": 9.2,
+    "max_scale_note": "largest scale factor verified end-to-end for these weights; beyond it the model still runs but is extrapolating outside its training range"
   },
   "training": {
-    "epochs": 100,
-    "batch_size": 8,
-    "patch_size": 128,
+    "stage": "2",
+    "iterations": 4000,
+    "batch_size": 2,
+    "patch_size": 192,
     "seed": 42,
-    "amp": true,
+    "optimizer": "adam",
+    "lr": 4e-05,
+    "min_lr": 2e-05,
+    "scheduler": "cosine",
+    "grad_clip": 1.0,
+    "amp": false,
     "losses": [
       "L1",
-      "VGG perceptual",
-      "PatchGAN"
+      "edge (Sobel gradient)",
+      "PatchGAN adversarial"
     ],
+    "loss_weights": {
+      "l1": 1.0,
+      "edge": 0.1,
+      "gan": 0.05
+    },
+    "num_buckets": 16,
+    "init_from": "checkpoints/liif_v4/best_liif.pth",
     "dataset": "DIV2K train + Flickr2K (3450 HR images)",
     "degradation": "Real-ESRGAN-style (blur, noise, downscale, JPEG)",
-    "checkpoint_epoch": 100
+    "checkpoint_step": 3500
   },
   "artifacts": {
     "torchscript": "model_scripted.pt",
@@ -49,65 +65,82 @@
   },
   "export": {
     "torch_version": "2.14.1+cu130",
-    "exported_at": "2026-10-23T14:46:57+00:00",
+    "exported_at": "2026-10-23T20:40:12+00:00",
     "method": "torch.jit.script + torch.jit.freeze",
     "verification": {
       "checks": [
         {
           "input": [
-            128,
-            160
+            160,
+            106
           ],
           "target": [
-            256,
-            320
+            320,
+            212
           ],
           "output_shape": [
             1,
             3,
-            320,
-            256
+            212,
+            320
           ],
-          "max_abs_diff": 1.0132789611816406e-06
+          "max_abs_diff": 1.1920928955078125e-07
         },
         {
           "input": [
-            128,
-            160
+            160,
+            106
           ],
           "target": [
-            413,
-            497
+            509,
+            335
           ],
           "output_shape": [
             1,
             3,
-            497,
-            413
+            335,
+            509
           ],
-          "max_abs_diff": 1.735985279083252e-06
+          "max_abs_diff": 1.1920928955078125e-07
         }
       ],
-      "max_abs_diff": 1.735985279083252e-06,
+      "max_abs_diff": 1.1920928955078125e-07,
       "forward_single_shape": [
         1,
         3,
-        320,
-        256
+        212,
+        320
       ],
       "forward_single_ok": true,
       "passed": true,
       "cuda_shape_check": [
         1,
         3,
-        320,
-        256
+        212,
+        320
       ],
       "cuda_finite": true
     },
     "examples": {
-      "source_image": null,
-      "files": {}
+      "source_image": "0801.png",
+      "files": {
+        "input.png": [
+          320,
+          240
+        ],
+        "output_4x.png": [
+          1280,
+          960
+        ],
+        "output_9.2x.jpg": [
+          1472,
+          1104
+        ],
+        "compare.jpg": [
+          496,
+          142
+        ]
+      }
     },
     "demo": {
       "source_image": "a.png",

examples/metrics.json

@@ -0,0 +1,106 @@
+{
+  "scale": 4,
+  "crop_size": 512,
+  "num_images": 5,
+  "per_image": [
+    {
+      "image": "0801.png",
+      "lr": [
+        128,
+        128
+      ],
+      "hr": [
+        512,
+        512
+      ],
+      "psnr_bicubic": 32.351,
+      "psnr_model": 32.632,
+      "ssim_bicubic": 0.8845,
+      "ssim_model": 0.8809,
+      "detail_hr": 2.1535,
+      "detail_bicubic": 0.1025,
+      "detail_model": 0.2354
+    },
+    {
+      "image": "0802.png",
+      "lr": [
+        128,
+        128
+      ],
+      "hr": [
+        512,
+        512
+      ],
+      "psnr_bicubic": 31.206,
+      "psnr_model": 31.042,
+      "ssim_bicubic": 0.8267,
+      "ssim_model": 0.8128,
+      "detail_hr": 3.0415,
+      "detail_bicubic": 0.0675,
+      "detail_model": 0.0998
+    },
+    {
+      "image": "0803.png",
+      "lr": [
+        128,
+        128
+      ],
+      "hr": [
+        512,
+        512
+      ],
+      "psnr_bicubic": 30.805,
+      "psnr_model": 31.34,
+      "ssim_bicubic": 0.8733,
+      "ssim_model": 0.8817,
+      "detail_hr": 2.3215,
+      "detail_bicubic": 0.1306,
+      "detail_model": 0.2517
+    },
+    {
+      "image": "0804.png",
+      "lr": [
+        128,
+        128
+      ],
+      "hr": [
+        512,
+        512
+      ],
+      "psnr_bicubic": 25.839,
+      "psnr_model": 26.072,
+      "ssim_bicubic": 0.7229,
+      "ssim_model": 0.7284,
+      "detail_hr": 13.3373,
+      "detail_bicubic": 0.199,
+      "detail_model": 0.4006
+    },
+    {
+      "image": "0805.png",
+      "lr": [
+        128,
+        128
+      ],
+      "hr": [
+        512,
+        512
+      ],
+      "psnr_bicubic": 27.77,
+      "psnr_model": 27.59,
+      "ssim_bicubic": 0.703,
+      "ssim_model": 0.6846,
+      "detail_hr": 8.9048,
+      "detail_bicubic": 0.1344,
+      "detail_model": 0.1757
+    }
+  ],
+  "average": {
+    "psnr_bicubic": 29.5942,
+    "psnr_model": 29.7352,
+    "ssim_bicubic": 0.8021,
+    "ssim_model": 0.7977,
+    "detail_hr": 5.9517,
+    "detail_bicubic": 0.1268,
+    "detail_model": 0.2326
+  }
+}

examples/verification.json

@@ -2,53 +2,53 @@
   "checks": [
     {
       "input": [
-        128,
-        160
+        160,
+        106
       ],
       "target": [
-        256,
-        320
+        320,
+        212
       ],
       "output_shape": [
         1,
         3,
-        320,
-        256
+        212,
+        320
       ],
-      "max_abs_diff": 1.0132789611816406e-06
+      "max_abs_diff": 1.1920928955078125e-07
     },
     {
       "input": [
-        128,
-        160
+        160,
+        106
       ],
       "target": [
-        413,
-        497
+        509,
+        335
       ],
       "output_shape": [
         1,
         3,
-        497,
-        413
+        335,
+        509
       ],
-      "max_abs_diff": 1.735985279083252e-06
+      "max_abs_diff": 1.1920928955078125e-07
     }
   ],
-  "max_abs_diff": 1.735985279083252e-06,
+  "max_abs_diff": 1.1920928955078125e-07,
   "forward_single_shape": [
     1,
     3,
-    320,
-    256
+    212,
+    320
   ],
   "forward_single_ok": true,
   "passed": true,
   "cuda_shape_check": [
     1,
     3,
-    320,
-    256
+    212,
+    320
   ],
   "cuda_finite": true
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d5f0b33a4448f4c936d26f05df6f2421566343b4f1abb0701582d8a200b94d4b
-size 5054380
+oid sha256:8c4781fa3375251b61aef71a6cc6c53fa3bc769f65d11e5733d4bf0d0b648c3f
+size 18439020

model_scripted.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:48deb75baf89d8de169be979afcf8989bb4504967af025d2a72d3e7cb54f0393
-size 5086206
+oid sha256:f20527462f30c27304a600e73f2eaa3ae0b60015312b59abbe51049ddbfce621
+size 18530078