Revise card: intro (2 characteristics), difficulty L1/L2 + L3 note, reasoning types, license wording

README.md

@@ -17,10 +17,13 @@ size_categories:
 
 # ReasonText
 
-**ReasonText** is a benchmark for **reasoning-required text-aware image super-resolution (TAISR)**, from the paper
-*"Reading the Unreadable: Text-Aware Image Super-Resolution Needs Reasoning"* (NeurIPS 2026). Some degraded text
-cannot be read from local pixels alone; humans read it by reasoning over context, prior knowledge, or logic.
-ReasonText makes this measurable with human-annotated word-level text and a per-instance **difficulty level**.
+**ReasonText** (from the paper *"Reading the Unreadable: Text-Aware Image Super-Resolution Needs Reasoning"*,
+NeurIPS 2026) is the first benchmark for **text-aware image super-resolution (TAISR)** with the following two
+characteristics:
+
+1. **High quality** — all annotations are produced by humans, with no automatic pipeline.
+2. **Per-instance difficulty** — each text instance is labeled by difficulty: **Level 1** (perception) and
+   **Level 2** (reasoning-required).
 
 - 📄 Paper: `<paper link — fill in>`
 - 💻 Code (metric **GTTCA** + method **RTC**, full usage): `<GitHub link — fill in>`
@@ -47,7 +50,7 @@ ReasonText-meta.json    # word-level annotations, keyed by image stem
       "text_rotation":   0,                // 0 / 90 / 180 / 270 (counter-clockwise)
       "text_flipped":    false,
       "text_difficulty": "Level 1",        // Level 1 / Level 2 / Level 3
-      "reason_types":    ["Context","Prior"] // Level 2 only: subset of Prior / Context / Logic
+      "reason_types":    ["Context","Prior"] // Level 2 only; see "Reasoning types" below
     }
   ]
 }
@@ -55,23 +58,33 @@ ReasonText-meta.json    # word-level annotations, keyed by image stem
 
 ## Difficulty levels
 
-Assigned by 10 annotators in two phases (cropped text region, then full LR image):
+Each text instance is labeled by how it can be read:
+
+| Level | Type | Readable from | Instances |
+|---|---|---|---|
+| **Level 1** | perception | the cropped LR text region | 1,754 |
+| **Level 2** | reasoning-required | the full LR image (via image context, prior knowledge, or logic) | 709 |
+
+Text that human annotators could not identify even from the full LR image is labeled **Level 3** (1,522
+instances); it is not the focus of this benchmark. In total, ReasonText has **513 images / 3,985 annotations**
+(the paper's evaluation set; LR–HR pairs whose text is *only* Level 3 are excluded).
+
+### Reasoning types (Level 2)
+
+For a Level 2 instance, `reason_types` records which cue makes the text recoverable:
 
-| Level | Meaning | Instances |
-|---|---|---|
-| **Level 1** | locally recoverable (readable from the text region alone) | 1,754 |
-| **Level 2** | **reasoning-required** (not L1, but recoverable using context / prior knowledge / logic) | 709 |
-| **Level 3** | unrecoverable even from the full LR image | 1,522 |
+- **Context** — image context
+- **Prior** — prior knowledge
+- **Logic** — logical pattern (logical inference)
 
-**513 images / 3,985 annotations** in total. This is the paper's evaluation set (LR–HR pairs containing *only*
-Level 3 instances are excluded). Level 2 is the focus of the paper; its `reason_types` mark the decisive cue
-(**Prior** knowledge, image **Context**, or **Logic**) and should be treated as a set.
+A single Level 2 instance can be recoverable through more than one of these cues, so `reason_types` is a **list**
+(a subset of the three).
 
 ## Provenance & license
 
-ReasonText images are drawn from **RealSR (v3)** [Cai et al., ICCV 2019] and **DRealSR** [Wei et al., ECCV 2020];
-the difficulty/text annotations are our contribution. Released under **CC BY-NC 4.0 (non-commercial research use
-only)**. Please cite our paper together with RealSR-v3 and DRealSR, and refer to those datasets for their terms.
+We release ReasonText under **CC BY-NC 4.0** (non-commercial research use only). Because ReasonText images are
+drawn from **RealSR (v3)** [Cai et al., ICCV 2019] and **DRealSR** [Wei et al., ECCV 2020], copyright of the
+original images belongs to the RealSR-v3 and DRealSR datasets.
 
 ## Citation
 
@@ -79,4 +92,4 @@ only)**. Please cite our paper together with RealSR-v3 and DRealSR, and refer to
 <bibtex — fill in>
 ```
 
-RealSR-v3 [Cai et al., ICCV 2019] and DRealSR [Wei et al., ECCV 2020] — please cite both.
+Please also cite **RealSR-v3** [Cai et al., ICCV 2019] and **DRealSR** [Wei et al., ECCV 2020].