docs: clean benchmark section + proper model-index (mean metrics, no per-dataset wall)

README.md

@@ -8,114 +8,87 @@ tags:
 - pytorch
 pipeline_tag: image-to-image
 model-index:
-- name: sen2sr (main — SEN2SRLite RGBN 4x)
+- name: sen2sr main (CNN RGBN 4x)
   results:
   - task:
       type: image-to-image
       name: Super Resolution 4x
     dataset:
-      type: opensr-test-naip
-      name: opensr-test NAIP
+      name: opensr-test
+      type: opensr-test
     metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.001016}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 0.471423}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.072529}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.056076}
-    - {type: omission,    name: "Omission ↓",           value: 0.069583}
-    - {type: improvement, name: "Improvement ↑",        value: 0.874341}
-  - task:
-      type: image-to-image
-      name: Super Resolution 4x
-    dataset:
-      type: opensr-test-spot
-      name: opensr-test SPOT
-    metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.001976}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 0.628618}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.076452}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.073418}
-    - {type: omission,    name: "Omission ↓",           value: 0.127390}
-    - {type: improvement, name: "Improvement ↑",        value: 0.799192}
-  - task:
-      type: image-to-image
-      name: Super Resolution 4x
-    dataset:
-      type: opensr-test-spain-crops
-      name: opensr-test Spain Crops
-    metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.001188}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 0.546995}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.077599}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.073548}
-    - {type: omission,    name: "Omission ↓",           value: 0.085900}
-    - {type: improvement, name: "Improvement ↑",        value: 0.840552}
-  - task:
-      type: image-to-image
-      name: Super Resolution 4x
-    dataset:
-      type: opensr-test-spain-urban
-      name: opensr-test Spain Urban
-    metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.002167}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 0.899134}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.065831}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.115622}
-    - {type: omission,    name: "Omission ↓",           value: 0.189015}
-    - {type: improvement, name: "Improvement ↑",        value: 0.695363}
-- name: sen2sr (mamba-rgbn-x4 — SEN2SR Mamba RGBN 4x)
+    - type: reflectance
+      name: Reflectance L1 (↓)
+      value: 0.001587
+    - type: spectral
+      name: Spectral Angle (↓)
+      value: 0.636542
+    - type: synthesis
+      name: Synthesis (↑)
+      value: 0.073103
+    - type: hallucination
+      name: Hallucination (↓)
+      value: 0.079666
+    - type: omission
+      name: Omission (↓)
+      value: 0.117972
+    - type: improvement
+      name: Improvement (↑)
+      value: 0.802362
+- name: sen2sr mamba-rgbn-x4 (Mamba RGBN 4x)
   results:
   - task:
       type: image-to-image
       name: Super Resolution 4x
     dataset:
-      type: opensr-test-naip
-      name: opensr-test NAIP
+      name: opensr-test
+      type: opensr-test
     metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.001255}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 0.674253}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.072526}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.056509}
-    - {type: omission,    name: "Omission ↓",           value: 0.070804}
-    - {type: improvement, name: "Improvement ↑",        value: 0.872687}
-  - task:
-      type: image-to-image
-      name: Super Resolution 4x
-    dataset:
-      type: opensr-test-spot
-      name: opensr-test SPOT
-    metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.002166}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 0.728039}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.076409}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.074770}
-    - {type: omission,    name: "Omission ↓",           value: 0.131327}
-    - {type: improvement, name: "Improvement ↑",        value: 0.793903}
-  - task:
-      type: image-to-image
-      name: Super Resolution 4x
-    dataset:
-      type: opensr-test-spain-crops
-      name: opensr-test Spain Crops
-    metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.001476}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 0.803401}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.077594}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.074450}
-    - {type: omission,    name: "Omission ↓",           value: 0.088468}
-    - {type: improvement, name: "Improvement ↑",        value: 0.837082}
+    - type: reflectance
+      name: Reflectance L1 (↓)
+      value: 0.001853
+    - type: spectral
+      name: Spectral Angle (↓)
+      value: 0.847490
+    - type: synthesis
+      name: Synthesis (↑)
+      value: 0.073092
+    - type: hallucination
+      name: Hallucination (↓)
+      value: 0.080977
+    - type: omission
+      name: Omission (↓)
+      value: 0.121563
+    - type: improvement
+      name: Improvement (↑)
+      value: 0.797460
+- name: sen2sr lite-main (CNN 10-band 4x)
+  results:
   - task:
       type: image-to-image
       name: Super Resolution 4x
     dataset:
-      type: opensr-test-spain-urban
-      name: opensr-test Spain Urban
+      name: opensr-test
+      type: opensr-test
     metrics:
-    - {type: reflectance, name: "Reflectance L1 ↓",    value: 0.002514}
-    - {type: spectral,    name: "Spectral Angle ↓",     value: 1.184267}
-    - {type: synthesis,   name: "Synthesis ↑",          value: 0.065839}
-    - {type: hallucination, name: "Hallucination ↓",    value: 0.118179}
-    - {type: omission,    name: "Omission ↓",           value: 0.195655}
-    - {type: improvement, name: "Improvement ↑",        value: 0.686166}
+    - type: reflectance
+      name: Reflectance L1 (↓)
+      value: 0.003309
+    - type: spectral
+      name: Spectral Angle (↓)
+      value: 2.006296
+    - type: synthesis
+      name: Synthesis (↑)
+      value: 0.073082
+    - type: hallucination
+      name: Hallucination (↓)
+      value: 0.079739
+    - type: omission
+      name: Omission (↓)
+      value: 0.119464
+    - type: improvement
+      name: Improvement (↑)
+      value: 0.800797
 ---
 
 
@@ -288,35 +261,34 @@ f2_model.safetensor / f2_hard_constraint.safetensor   (RSWIR 2x stage)
 
 ---
 
+
 ## Benchmark
 
-Evaluated on [opensr-test](https://github.com/ESAOpenSR/opensr-test) across 4 datasets (NAIP, SPOT, Spain Crops, Spain Urban).
-Venus is a 2x dataset and is excluded from 4x model evaluation.
-Full results: [eval/sen2sr_eval_results.csv](eval/sen2sr_eval_results.csv)
+Evaluated on [opensr-test](https://github.com/ESAOpenSR/opensr-test) — NAIP, SPOT, Spain Crops, Spain Urban (4× datasets). Means across all 4 datasets. Venus excluded (2× dataset, incompatible with 4× models).
+
+| Variant | Architecture | Reflectance↓ | Spectral↓ | Hallucination↓ | Omission↓ | Improvement↑ |
+|---|---|---|---|---|---|---|
+| `main` | CNN | 0.0016 | 0.64 | 0.080 | 0.118 | **0.802** |
+| `mamba-rgbn-x4` | Mamba | 0.0019 | 0.85 | 0.081 | 0.122 | 0.797 |
+| `lite-main` | CNN (10-band) | 0.0033 | 2.01 | 0.080 | 0.119 | 0.801 |
+
+> `mamba-main` results pending. Full per-dataset breakdown: [`eval/sen2sr_eval_results.csv`](eval/sen2sr_eval_results.csv)
 
 <p align="center">
-  <img src="eval/sen2sr_benchmark.png" width="100%">
+  <img src="eval/sen2sr_benchmark.png" width="95%">
 </p>
 
-| Variant | Reflectance↓ | Spectral↓ | Synthesis↑ | Hallucination↓ | Omission↓ | Improvement↑ |
-|---|---|---|---|---|---|---|
-| `main` (CNN RGBN 4x) | 0.0016 | 0.637 | 0.073 | 0.080 | 0.118 | 0.802 |
-| `mamba-rgbn-x4` (Mamba RGBN 4x) | 0.0019 | 0.848 | 0.073 | 0.081 | 0.122 | 0.797 |
-| `lite-main` (CNN 10-band 4x) | 0.0033 | 2.006 | 0.073 | 0.080 | 0.120 | 0.801 |
+### Visual comparison — Spain Urban (opensr-test)
 
-*Mean across NAIP, SPOT, Spain Crops, Spain Urban. `mamba-main` results pending (see open issues).*
+LR (Sentinel-2 bicubic) · SR (sen2sr `main`) · HR (SPOT reference)
 
----
+<p align="center">
+  <img src="eval_images/spain_urban.png" width="95%">
+</p>
 
-## Citation
+### Visual comparison — Spain Crops (opensr-test)
 
-If you use these models please cite the original work:
+<p align="center">
+  <img src="eval_images/spain_crops.png" width="95%">
+</p>
 
-```bibtex
-@software{sen2sr2024,
-  author  = {Aybar, Cesar and others},
-  title   = {SEN2SR: Sentinel-2 Super-Resolution},
-  url     = {https://github.com/ESAOpenSR/sen2sr},
-  year    = {2024}
-}
-```