huggingworld commited on
Commit
d67c139
·
verified ·
1 Parent(s): 944e36f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +80 -3
README.md CHANGED
@@ -1,10 +1,87 @@
1
  ---
2
  title: Image Upscale Lab WebGPU
3
- emoji: 🏆
4
- colorFrom: yellow
5
  colorTo: green
6
  sdk: static
7
  pinned: false
 
8
  ---
9
 
10
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  title: Image Upscale Lab WebGPU
3
+ emoji: 👀
4
+ colorFrom: purple
5
  colorTo: green
6
  sdk: static
7
  pinned: false
8
+ short_description: WebGPU image upscale
9
  ---
10
 
11
+
12
+ ## **TIER 1: TOP PERFORMERS** ⭐⭐⭐
13
+
14
+ 1. **4x UltraSharp V2 FP32** (51.8MB) - The gold standard for general-purpose upscaling
15
+ - Superior detail preservation across all content types
16
+ - Excellent handling of both photos and artwork
17
+ - Minimal artifacts, natural texture enhancement
18
+
19
+ 2. **4x PurePhoto RealPLSKR** (29.9MB) - Best for photographic content
20
+ - Exceptional on portraits and realistic images
21
+ - Advanced noise handling while preserving skin textures
22
+ - Recommended by multiple 2026 benchmarks
23
+
24
+ 3. **4x ClearReality V1 FP32** (1.7MB) - Best lightweight option
25
+ - Surprisingly excellent quality for its tiny size
26
+ - Fast processing with minimal resource usage
27
+ - Great balance of quality vs. performance
28
+
29
+ ## **TIER 2: EXCELLENT SPECIALISTS** ⭐⭐
30
+
31
+ 4. **4x Nomos8k DAT** (89.9MB) - JPEG compression specialist
32
+ - Specifically trained for compressed image restoration
33
+ - Excellent on heavily compressed photos
34
+ - Recommended for social media/web images
35
+
36
+ 5. **4x PurePhoto SPAN** (1.7MB) - Another excellent lightweight
37
+ - Very similar quality to ClearReality
38
+ - Great for mobile/edge deployment
39
+ - Recommended for real-time applications
40
+
41
+ 6. **4x LSDIR** (67.2MB) - Large dataset trained
42
+ - Good general performance across content types
43
+ - Solid choice for batch processing
44
+
45
+ ## **TIER 3: SOLID PERFORMERS** ⭐
46
+
47
+ 7. **4x UltraSharp FP32** (67MB) - Original UltraSharp
48
+ - Still excellent but superseded by V2
49
+ - Good for legacy compatibility
50
+
51
+ 8. **RealESR-General WDN x4v3** (4.9MB) - Compact and efficient
52
+ - Recommended in multiple 2026 reviews
53
+ - Good balance of speed and quality
54
+
55
+ 9. **4x Nomos2 HQ DAT2 FP32** (52.9MB) - High-quality alternative
56
+ - Similar performance to top tier but larger file size
57
+
58
+ ## **TIER 4: SPECIALIZED/NICHE**
59
+
60
+ 10. **4x Upscayl UltraSharp** (67.2MB) - Upscayl version
61
+ - Good performance but redundant with standalone UltraSharp
62
+
63
+ 11. **4x Upscayl High Fidelity** (67.2MB) - Upscayl variant
64
+ - Quality varies by source image
65
+
66
+ 12. **2x Real-ESRGAN x2plus** (67.2MB) - Only 2x scaling
67
+ - Limited by scale factor but good quality
68
+
69
+ ## **TIER 5: AVOID/OUTDATED**
70
+
71
+ 13-42. Most remaining models fall into these categories:
72
+ - **Outdated architectures** (original ESRGAN variants)
73
+ - **Overly specialized** (JPG restoration models for specific compression levels)
74
+ - **Redundant** (multiple similar Upscayl variants)
75
+ - **Limited scale factors** (1x models for restoration only)
76
+
77
+ ## **KEY RECOMMENDATIONS:**
78
+
79
+ **For General Use:** Start with **4x UltraSharp V2 FP32** - it consistently ranks #1 in 2026 benchmarks
80
+
81
+ **For Mobile/Limited Resources:** Use **4x ClearReality V1 FP32** or **4x PurePhoto SPAN** - tiny files, excellent quality
82
+
83
+ **For Compressed Images:** **4x Nomos8k DAT** handles JPEG artifacts better than anything else
84
+
85
+ **For Photography:** **4x PurePhoto RealPLSKR** is specifically optimized for photos and portraits
86
+
87
+ The field has matured significantly, with newer architectures like DAT2 and RealPLKSR clearly outperforming older ESRGAN-based models. File size doesn't always correlate with quality - some of the best models are surprisingly compact.