CodeGoat24 commited on
Commit
5b5d84d
·
verified ·
1 Parent(s): b4540e4

Point to CodeGoat24/WorldReward-9B

Browse files
Files changed (1) hide show
  1. README.md +31 -5
README.md CHANGED
@@ -24,9 +24,14 @@ A human-annotated preference benchmark for **camera-conditioned world models**.
24
  source image and the *same* camera-action sequence, with human verdicts on three
25
  independent axes.
26
 
27
- Companion to [`CodeGoat24/WorldReward`](https://huggingface.co/CodeGoat24/WorldReward). Code: https://github.com/CodeGoat24/WorldReward
28
 
29
- ## Evaluation Dimensions
 
 
 
 
 
30
 
31
  | Axis | Question |
32
  |---|---|
@@ -47,8 +52,18 @@ videos/<pair_id>/left_overlay.mp4 # left.mp4 with the commanded action burned i
47
  videos/<pair_id>/right_overlay.mp4 # right.mp4 with the commanded action burned in
48
  ```
49
 
 
 
 
 
 
50
  The `*_overlay.mp4` files are a visualisation aid for inspecting trajectories by
51
- eye.
 
 
 
 
 
52
 
53
  ## Schema
54
 
@@ -66,6 +81,11 @@ eye.
66
  | `style` | string | `photo` / `game_anime` / `traditional` |
67
  | `label` | object | Human verdicts: `action`, `appearance`, `motion`, each `left`/`right`/`tie` |
68
 
 
 
 
 
 
69
  ## Composition
70
 
71
  Trajectory groups:
@@ -85,7 +105,8 @@ Visual styles:
85
  | `traditional` | 61 | 8.0% |
86
 
87
  Videos come from 9 generation systems compared pairwise. Ground-truth
88
- verdicts are close to balanced between the two positions.
 
89
 
90
  | Axis | left | right | tie |
91
  |---|---:|---:|---:|
@@ -119,6 +140,11 @@ axis a predictor does not model.
119
  | Qwen3.5-VL-27B | 63.68 | 44.34 | 62.76 | 65.07 | 37.33 | 65.75 | 65.05 | 51.61 | _63.44_ | 61.35 | 46.81 | 59.22 | 64.93 | 38.36 | 66.85 | 62.87 | 51.50 | 58.68 | 60.66 | 40.98 | 60.66 |
120
  | **WorldReward-9B** | **77.63** | **81.32** | **73.03** | **76.71** | 77.74 | **78.77** | **73.12** | **86.02** | **64.52** | **81.56** | **81.91** | **72.70** | **77.26** | 81.37 | **71.78** | **78.74** | **82.04** | **75.75** | _73.77_ | **77.05** | _65.57_ |
121
 
 
 
 
 
 
122
 
123
  ## Usage
124
 
@@ -132,7 +158,7 @@ python scripts/preprocess.py \
132
  python scripts/run_inference.py \
133
  --render-root outputs/rendered_chunks \
134
  --output outputs/chunk_predictions.json \
135
- --model CodeGoat24/WorldReward
136
  python worldreward-bench/score.py \
137
  --bench data/WorldReward-Bench/bench.jsonl \
138
  --predictions outputs/chunk_predictions.json
 
24
  source image and the *same* camera-action sequence, with human verdicts on three
25
  independent axes.
26
 
27
+ Companion to [`CodeGoat24/WorldReward-9B`](https://huggingface.co/CodeGoat24/WorldReward-9B). Code: https://github.com/CodeGoat24/WorldReward
28
 
29
+ ## What makes this benchmark different
30
+
31
+ Most video-generation benchmarks score a single video for overall quality. Here
32
+ each item is a **head-to-head comparison** under a *known camera trajectory*, and
33
+ the three axes are annotated separately, so a model that moves the camera
34
+ correctly but renders it badly is distinguishable from the reverse:
35
 
36
  | Axis | Question |
37
  |---|---|
 
52
  videos/<pair_id>/right_overlay.mp4 # right.mp4 with the commanded action burned in
53
  ```
54
 
55
+ `pair_id` runs from `wrb_0001` to `wrb_0760`. The ids are opaque and carry no
56
+ meaning: consecutive ids are unrelated, and the numbering encodes nothing about
57
+ trajectory, style, or which systems were compared. Use `trajectory_group`, `style`
58
+ and `left`/`right` `model` for those.
59
+
60
  The `*_overlay.mp4` files are a visualisation aid for inspecting trajectories by
61
+ eye. They are **not** model input; evaluation reads only `left.mp4` / `right.mp4`.
62
+
63
+ Source images keep their original format and pixels when they are at most
64
+ 2048px on the long side. Larger ones are downscaled to 2048px and saved
65
+ as JPEG, matching what the evaluation pipeline feeds the model anyway. Always
66
+ read the path from `bench.jsonl` rather than assuming an extension.
67
 
68
  ## Schema
69
 
 
81
  | `style` | string | `photo` / `game_anime` / `traditional` |
82
  | `label` | object | Human verdicts: `action`, `appearance`, `motion`, each `left`/`right`/`tie` |
83
 
84
+ Action tokens combine at most one translation (`forward`, `backward`, `left`,
85
+ `right`, `forward_left`, ...) with at most one rotation (`camera_up` = pitch up,
86
+ `camera_down` = pitch down, `camera_left` = yaw left, `camera_right` = yaw right,
87
+ plus `camera_ul`/`ur`/`dl`/`dr` diagonals), joined by `+`.
88
+
89
  ## Composition
90
 
91
  Trajectory groups:
 
105
  | `traditional` | 61 | 8.0% |
106
 
107
  Videos come from 9 generation systems compared pairwise. Ground-truth
108
+ verdicts are close to balanced between the two positions, so a predictor that
109
+ always answers `left` scores near chance:
110
 
111
  | Axis | left | right | tie |
112
  |---|---:|---:|---:|
 
140
  | Qwen3.5-VL-27B | 63.68 | 44.34 | 62.76 | 65.07 | 37.33 | 65.75 | 65.05 | 51.61 | _63.44_ | 61.35 | 46.81 | 59.22 | 64.93 | 38.36 | 66.85 | 62.87 | 51.50 | 58.68 | 60.66 | 40.98 | 60.66 |
141
  | **WorldReward-9B** | **77.63** | **81.32** | **73.03** | **76.71** | 77.74 | **78.77** | **73.12** | **86.02** | **64.52** | **81.56** | **81.91** | **72.70** | **77.26** | 81.37 | **71.78** | **78.74** | **82.04** | **75.75** | _73.77_ | **77.05** | _65.57_ |
142
 
143
+ `UnifiedReward-Think` / `-Flex` returned nothing for 2 and 28 pairs, excluded from
144
+ their denominators (758 / 732). Every other predictor covers all 760.
145
+
146
+ The aggregation rule behind these numbers, and how to reproduce them, are in
147
+ https://github.com/CodeGoat24/WorldReward.
148
 
149
  ## Usage
150
 
 
158
  python scripts/run_inference.py \
159
  --render-root outputs/rendered_chunks \
160
  --output outputs/chunk_predictions.json \
161
+ --model CodeGoat24/WorldReward-9B
162
  python worldreward-bench/score.py \
163
  --bench data/WorldReward-Bench/bench.jsonl \
164
  --predictions outputs/chunk_predictions.json