init

.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,138 @@
+---
+license: mit
+pretty_name: TrajectoryBench
+size_categories:
+- 1K<n<10K
+task_categories:
+- image-to-video
+tags:
+- world-models
+- video-generation
+- benchmark
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: metadata.jsonl
+dataset_info:
+  features:
+  - name: id
+    dtype: string
+  - name: category
+    dtype: string
+  - name: difficulty
+    dtype: string
+  - name: stylized
+    dtype: bool
+  - name: art_style
+    dtype: string
+  - name: num_frames
+    dtype: int64
+  - name: caption
+    dtype: string
+  - name: camera_prompt
+    dtype: string
+---
+
+# TrajectoryBench
+
+TrajectoryBench is a benchmark for long-horizon, camera-controlled video generation with complex trajectories. Each of its 2000 samples pairs a first frame with a camera trajectory and a text prompt. The model generates a video that starts from the frame and follows the trajectory, and is scored on 3D consistency, camera following, and video quality.
+
+Paper: [LoGo: Local-Global Rewards for Consistent Long-Horizon Video Generation](https://arxiv.org/abs/2610.03636)
+
+## Contents
+
+```
+metadata.jsonl        one row per sample
+data/
+  0001/
+    first_frame.png   832x480 conditioning frame
+    trajectory.npz    camera path to follow
+    prompt.json       caption and camera prompt
+  0002/
+  ...
+  2000/
+```
+
+## Metadata
+
+`metadata.jsonl` has one JSON object per line, one per sample, with these fields:
+
+- `id`: sample folder name under `data/`
+- `category`: `indoor`, `outdoor`, or `transition` (passing through a door into a new space)
+- `difficulty`: `easy`, `medium`, or `hard`
+- `stylized`: `true` for non-photorealistic first frames
+- `art_style`: the style of a stylized sample, one of `minecraft_blocks`, `japanese_anime`, `oil_painting`, `watercolor`, `low_poly_3d`, `studio_3d_cartoon`, `neon_cyberpunk`, `victorian_steampunk`, or `null`
+- `num_frames`: number of camera poses, and so the number of frames to generate
+- `caption`: a video-style description of the scene, about 90 words
+- `camera_prompt`: description of the camera motion
+
+`prompt.json` in each sample folder holds the same `caption` and `camera_prompt`.
+
+The benchmark covers:
+
+- indoor: 125 easy, 125 medium, 500 hard
+- outdoor: 125 easy, 125 medium, 500 hard
+- transition: 250 medium, 250 hard
+
+Each group is 80% photoreal and 20% stylized. For each hard indoor or outdoor scene, there are two trajectories: one focusing on expansion (generating content far from the initial view), and one focusing on revisit (revisiting the same region from different angles multiple times).
+
+## Difficulty
+
+- **easy**: one camera primitive (push in, pull back, move, orbit, or pan) over 81 frames, subset from WorldScore.
+- **medium**: medium includes transition scenes which go through a door into a new space. For indoor and outdoor, the medium split evaluates three chained primitives over 241 frames. This corresponds to WorldScore's small split of "composite camera motion", which is three separate 80-frame generations in a loop and thus not truly long-horizon. The indoor and outdoor initial images are sourced from this split of WorldScore for comparability.
+- **hard**: 241-frame generation with 7-8 camera motions. Indoor and outdoor initial images are sourced from [Pexels](https://www.pexels.com), with trajectories either focusing on expansion far from the initial view or looking away and revisiting the same region from multiple viewpoints. Transition initial images are generated with [Nano Banana](https://deepmind.google/models/gemini-image/) to specify the location and size of the door, so that the trajectories can reasonably pass through it and generate a new space.
+
+## Trajectory format
+
+`trajectory.npz` contains:
+
+- `w2c`: `(N, 4, 4)` float32 world-to-camera matrices, one per frame. Frame 0 is the identity, so poses are relative to the first frame.
+- `intrinsics`: `(N, 3, 3)` float32 pinhole intrinsics in pixels.
+- `image_wh`: `(2,)` int32 image width and height, `[832, 480]`.
+
+Translations are given at the scale the samples are evaluated at, so a model follows `w2c` as
+published without rescaling it.
+
+Camera axes follow the OpenCV convention: +x right, +y down, +z forward.
+
+## Usage
+
+```python
+import json
+import numpy as np
+import pandas as pd
+from huggingface_hub import snapshot_download
+from PIL import Image
+
+root = snapshot_download("ziqima/TrajectoryBench", repo_type="dataset")
+meta = pd.read_json(f"{root}/metadata.jsonl", lines=True, dtype={"id": str})
+
+sample = meta.iloc[0]
+d = f"{root}/data/{sample.id}"
+frame = Image.open(f"{d}/first_frame.png")
+traj = np.load(f"{d}/trajectory.npz")
+prompt = json.load(open(f"{d}/prompt.json"))
+
+w2c, K = traj["w2c"], traj["intrinsics"]  # (N, 4, 4), (N, 3, 3)
+```
+
+Generate one video per sample with `num_frames` frames at 832x480, conditioned on `first_frame.png` and following `w2c` and `intrinsics`.
+
+## Citation
+
+```bibtex
+@misc{ma2026logo,
+  title={LoGo: Local-Global Rewards for Consistent Long-Horizon Video Generation},
+  author={Ziqi Ma and Shreya Sharma and Mohamed El Banani and Katja Schwarz and Chongjie Ye and Chao-Yuan Wu and Li Fei-Fei and Ben Mildenhall and Georgia Gkioxari and Justin Johnson and Gowthami Somepalli},
+  year={2026},
+  eprint={2610.03636},
+  archivePrefix={arXiv},
+  primaryClass={cs.CV},
+  url={https://arxiv.org/abs/2610.03636},
+}
+```
+
+## License
+
+TrajectoryBench is released under the MIT License.

data/0001/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a forward movement through open white double doors with brass handles, entering a passport office. The dim hallway features creamy walls and a wooden bench. Inside, the bright space reveals dark wood desks, black office chairs, and computer monitors on the counters. A light blue back wall displays a royal crest, \"PASSPORT OFFICE - LONDON,\" and a \"WAIT HERE\" sign, while hanging signs indicate \"WINDOW 4\" and \"WINDOW 5\". Bright lighting illuminates this completely empty service area as the view progresses smoothly inward over the carpeted floor.",
+ "camera_prompt": "Walk forward through the open doorway into the space beyond."
+}

data/0001/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7188799b5edd9995eaaf4662424c5734e0ffd345f894df1b2781e3ffebc45102
+size 8868

data/0002/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a narrow European courtyard completely empty of people, defined by warm terracotta-orange facades, a light beige building, and light-yellow herringbone paving tiles. Notable objects include turquoise-green shutters, hanging clothing, air conditioning units, terracotta potted plants, and tarp-covered bicycles clustered against a low right wall. Vibrant, warm sunlight bathes the entire scene beneath a pale blue sky, emphasizing the architectural textures. As the view slowly moves forward through the intimate space, it reveals a tall back wall topped with a globe, communication antennae, and a dark satellite dish.",
+ "camera_prompt": "Back up and turn left to look toward one area, then turn back and move to revisit the starting spot."
+}

data/0002/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:907887592de6cd1e9da459fb384a4c5a9a9ca2240b739d6e8ac043271405d4bd
+size 24868

data/0003/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a dramatic coastal view from inside a rugged, dark sea cave, looking out toward a turbulent ocean and towering rock formations. The damp, uneven stone floor and jagged tunnel walls feature deep charcoal and brown shadows. These contrast with the pale grey overcast sky and dark blue, white-capped waters. Natural daylight pours through the opening, illuminating the exit while keeping the interior shadowed. As the viewpoint moves steadily forward, it emerges from the corridor, expanding the breathtaking view of the sea in this completely unpopulated natural space.",
+ "camera_prompt": "Walk forward through the doorway into the space beyond, move right, turn to face a new direction, and continue exploring the room."
+}

data/0003/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af2da9a21c21453533c99049c9d34dbcb37d9e6852b3d04d8a31277c63a11c33
+size 24868

data/0004/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a modern, minimalist living room entirely empty of people. As the view slowly moves through and reveals more of the elegant space, it highlights a neutral color palette of white and gray with warm wood tones and navy blue accents. Materials like a floor-to-ceiling marble-effect wall and dark wood-look floors anchor notable objects, including a light beige sofa, a round table with three brown slatted chairs, and a media console holding a television. Diffused natural light from sheer-curtained windows blends seamlessly with subtle recessed lighting.",
+ "camera_prompt": "Back up and turn left to look toward one area, then turn back and move to revisit the starting spot."
+}

data/0004/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01cecd478d92c10bbc619b4d5c1b1614eb61760742726110b0e58d6c57e1e132
+size 24868

data/0005/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases an empty pedestrian promenade in Chartres, France, rendered in a muted, desaturated color palette. Gliding forward over a pale beige gravel path, the view slowly advances to reveal a low concrete wall enclosing vibrant yellow wildflowers and silvery grasses. Soft, diffuse light from a pale orange and yellow sky casts long shadows across the elegant space. The gently moving perspective passes perfectly spherical trees in concrete planters, grey multi-story buildings—one bearing a \"La Lingerie\" sign—a small red kiosk, and an open-air building with a glass roof and metal supports.",
+ "camera_prompt": "Back out of the space, then turn right and strafe sideways in steps, sweeping across to reveal more of the scene."
+}

data/0005/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75fe376f8bf70891e90adca0bc730c28ee9cceb8bd3ca66485b7b43f322f0d97
+size 24868

data/0006/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a serene marshland at dawn or dusk, remaining entirely empty of people. Vibrant purple, pink, and bright orange hues paint the sky, reflecting softly in a shallow foreground puddle. Earthy browns and greens tint the damp soil and tall, dry grasses. Slender bare trees and a budding bush stand silhouetted against the glowing horizon, with a gentle mist hovering over distant fields. Under soft, dramatic natural lighting, the view slowly glides forward, pushing through the bristly brush to reveal deeper layers of this tranquil, foggy wetland environment.",
+ "camera_prompt": "Pull back away from the scene, pan the camera left, move left."
+}

data/0006/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f7f8979c6fa60eff693472ed8a72669b3e053a0c7d5598a1c2eb03cebf09368
+size 24868

data/0007/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a tranquil study space rendered in thick strokes of paint, completely empty of people. Dominated by rich, textured blue hues, the scene features wooden shelves heavily laden with books sporting warm orange, green, and dark blue spines. Below the shelves sits a desk holding a stacked trio of light blue books, a cup of pencils, and an open laptop. No text is visible. A glowing, bulb-like fixture casts a luminous yellow halo. Slowly, the view glides inward, intimately exploring the quiet dimensions and revealing more of the painted room.",
+ "camera_prompt": "Push in toward the scene."
+}

data/0007/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bc0c32c3e74f8a3f7f5b55d39eb732250b4dac1ccd95834db4ba957bc1fc80f
+size 8868

data/0008/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a pristine alpine landscape completely empty of people. Vibrant green grasses and dark evergreen trees contrast sharply with the light aquamarine water of a fast-flowing river. Scattered grey and brown boulders line the rugged riverbed. In the distance, a massive grey mountain peak, lightly dusted with snow, rises under a pale blue sky. Brilliant natural daylight illuminates the organic materials of stone, rushing water, and lush vegetation. Devoid of any signage or text, the view slowly moves forward, exploring the wild terrain and revealing more of the winding waterway and vast forest.",
+ "camera_prompt": "Orbit right around the scene."
+}

data/0008/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ba63a4bdfc6a0d6522944044d2ab39093652234fe047414d82b6a787d940953
+size 8868

data/0009/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a serene alpine valley, entirely empty of people. Vibrant blue skies contrast with crisp white snow dusting rugged rocky mountain peaks. Below, a clear river flows between banks lined with leafless, orange-brown bushes and towering dark green evergreen trees. To the left, quaint wooden houses nestle against the forested hillside. Without any visible signage, the scene is bathed in bright, natural daylight highlighting textures of wood, stone, and water. The view slowly glides forward along the winding river, gently revealing deeper, majestic layers of the snow-capped mountain range ahead.",
+ "camera_prompt": "Pan the camera left, move left, pull back away from the scene."
+}

data/0009/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a615b22138fa9f17ba27bf9731a5aac127b782d31ef42ba4585835bca7eba748
+size 24868

data/0010/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a serene flight over a vast, undulating landscape filled with distinctive conical rock formations. Warm hues of golden yellow, deep orange, and earthy brown color the rugged stone, contrasted perfectly by patches of green vegetation. Several hot air balloons, crafted from red and white patterned fabric without legible text, float gracefully above the valleys. Bathed in soft, warm golden light under a pale peach sky, the slow, forward-gliding view gradually reveals deeper, shadowed canyons across a majestic terrain that appears completely empty of visible people.",
+ "camera_prompt": "Orbit right around the scene."
+}

data/0010/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ba63a4bdfc6a0d6522944044d2ab39093652234fe047414d82b6a787d940953
+size 8868

data/0011/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases the serene, breathtaking Lake Louise in Banff National Park, completely empty of people. Gliding forward over the deep turquoise blue water, the view slowly reveals towering mountains flanking the lake. Dense green evergreen forests cover gray and beige rock formations, while white snow patches and a distant hanging glacier emerge on the rugged peaks. To the left, a small red-orange wooden dock holds bright red canoes, vividly contrasting the natural materials. Lacking any text or signage, the space is bathed in moody, diffused light from an overcast sky.",
+ "camera_prompt": "Back out of the space, then turn left and strafe sideways in steps, sweeping across to reveal more of the scene."
+}

data/0011/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98499feb6c79ccfb3a7f954bb63cba35c4a14b50bf5ccae2a2314cf113d916d1
+size 24868

data/0012/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a luxurious, people-free bedroom and a visible modern bathroom. Cool grey bed linens, flowing curtains, and an off-white paneled wall contrast with a blush-pink tufted headboard and light wood floors. Notable objects include a vase of pink and white flowers, a freestanding bathtub against pale-beige tiles, and standard wall switches devoid of text. Rose-gold pendant lights and a recessed LED ceiling strip provide soft illumination. As the view slowly moves through, it reveals more of the peaceful space, uncovering bathroom niches holding a potted plant and candle.",
+ "camera_prompt": "Back up and turn right to look toward one area, then turn back and move to revisit the starting spot."
+}

data/0012/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30465ab230ad2170ea02dacd4ff8c5a8bca5298341769f5e928bc65e1cd3d529
+size 24868

data/0013/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a modern, minimalist kitchen and dining area empty of people. The space features matte black, dark gray, and neutral off-white tones against light wood flooring. Materials include sleek wood cabinetry and stainless steel appliances. Notable objects are a light wood dining table, X-backed black chairs, a ribbed island, linear pendant lights, and potted plants, with no visible signage. Bathed in bright natural light, the view slowly moves through the space, revealing a glimpse of a living room with a beige sofa and green plant through a large window.",
+ "camera_prompt": "Back out of the space, then turn right and strafe sideways in steps, sweeping across to reveal more of the scene."
+}

data/0013/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19579eb390b6375a6d030b5e60b63f47bd9005ca8ba1fa09d410b4d079907cb5
+size 24868

data/0014/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a bright, modern kitchen empty of people. Crisp white shiplap walls and upper cabinets contrast elegantly with gray lower units and warm wood butcher block countertops. A stainless steel Frigidaire electric stove sits below a matching hood featuring a warm under-light. Nearby, a dark gooseneck faucet arcs over a metallic sink, complemented by matte black cabinet hardware throughout. Abundant natural daylight streams through a window and a glass door. The view slowly glides forward, exploring the compact layout and revealing more of the airy, adjacent living space.",
+ "camera_prompt": "Pull back away from the scene, move right, pan the camera right."
+}

data/0014/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:559f9b1e0f4add9983dab8c552611b3cf24445b00e4e920737b4f3a89886f389
+size 24868

data/0015/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases a steep San Francisco street completely empty of people, as the view slowly moves through and reveals more of the space. To the left, a textured brick building features a window poster reading \"BOLD,\" standing near a white glass-and-steel structure and a gray stone building. A whimsical yellow, red, and green tiered pagoda structure adds vibrant color. Along the paved road, parked cars and a moving white van flank dark red cable car tracks. Lush trees and a bright blue sky bask in a warm late-afternoon glow.",
+ "camera_prompt": "Back out of the space, then turn left and strafe sideways in steps, sweeping across to reveal more of the scene."
+}

data/0015/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98499feb6c79ccfb3a7f954bb63cba35c4a14b50bf5ccae2a2314cf113d916d1
+size 24868

data/0016/prompt.json

@@ -0,0 +1,4 @@
+{
+ "caption": "The video showcases the ornate facade of the Shaykh Lotfollah Mosque in Isfahan, Iran, completely empty of people. Turquoise, navy blue, and white mosaic tiles form complex geometric and floral patterns across the massive multi-lobed arched doorway, which holds a terracotta red latticed wooden screen. A pronounced band of Arabic calligraphy crowns the facade. Diffused lighting from an overcast sky casts soft shadows across the gray stone courtyard and a sunken circular depression. As the video continues, the view slowly moves through the paved courtyard, revealing more of this tranquil, atmospheric space.",
+ "camera_prompt": "Back up and turn right to look toward one area, then turn back and move to revisit the starting spot."
+}

data/0016/trajectory.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db9c8f07f035b280317524130b3af0167ab4127ab3e630ecacee777e915986ac
+size 24868