← 返回目录

huayi-three-task-samples

Huayi · paired video-to-video review

Three tasks.
Six paired samples.

Task 118 去雨、Task 121 室内去雾、Task 130 运动去模糊的统一 sample 视图。 每个 task 固定两组,均包含 input / source target 视频、准确 prompt 和 sample-level metadata

3tasks
6paired samples
12reference videos
0model outputs
证据边界:页面里的 target 均为发布方提供的成对 ground truth, 不是模型生成结果。视频为浏览器展示副本;Task 118 的 10 fps、Task 121 的 8 fps 与 Task 130 的 15 fps 是各自现有 preview 约定,不应跨 task 比较速度。
Task 118 · adverse weather restoration

Rain removal · RainSyn25

选择 Light 与 Complex 各一条 31 帧序列。输入是合成雨层覆盖的视频, target 是同一 scene、同一时刻的 clean source ground truth。

Gate A draft
01

RainSynLight25 · sequence 17

Reference rainsyn-light25-test-17

LIGHT RAIN 31 FRAMES SOURCE-PAIRED GT CGI SCENE
Input · rainyrfc-1…31
Target · clean GTgtc-1…31

Prompt

Remove the rain streaks and restore a clean, temporally stable video while preserving the scene, motion, color, framing, and timing.

Metadata

Dataset / split
RainSynLight25 · test
Mapping
rfc-N.jpg → gtc-N.jpg
Source geometry
889×500 · 31 paired JPEGs
Preview
890×500 · 10 fps · 3.10 s · H.264/yuv420p
Target origin
Publisher-provided clean ground truth
Review note
右侧仅补 1 像素供浏览器播放;无 crop / resize / retime。
查看 metadata.json 形式
{
  "task_id": 118,
  "sample_ref": "rainsyn-light25-test-17",
  "dataset": "RainSynLight25",
  "subset": "frames_light_test_JPEG",
  "source_sequence": "17",
  "input_members": "rfc-1.jpg ... rfc-31.jpg",
  "target_members": "gtc-1.jpg ... gtc-31.jpg",
  "frame_count": 31,
  "source_resolution": "889x500",
  "preview_resolution": "890x500",
  "preview_fps": 10,
  "target_origin": "source_ground_truth",
  "workflow_status": "gate_a_draft"
}
02

RainSynComplex25 · sequence 25

Reference rainsyn-complex25-test-25

COMPLEX RAIN 31 FRAMES SOURCE-PAIRED GT REAL-SCENE PLATE
Input · rainyrfc-1…31
Target · clean GTgtc-1…31

Prompt

Remove the rain streaks and restore a clean, temporally stable video while preserving the scene, motion, color, framing, and timing.

Metadata

Dataset / split
RainSynComplex25 · test
Mapping
rfc-N.jpg → gtc-N.jpg
Source geometry
889×500 · 31 paired JPEGs
Preview
890×500 · 10 fps · 3.10 s · H.264/yuv420p
Target origin
Publisher-provided clean ground truth
Review note
蜜蜂与向日葵实景底片;合成 complex rain,仅右侧补 1 像素。
查看 metadata.json 形式
{
  "task_id": 118,
  "sample_ref": "rainsyn-complex25-test-25",
  "dataset": "RainSynComplex25",
  "subset": "frames_heavy_test_JPEG",
  "source_sequence": "25",
  "input_members": "rfc-1.jpg ... rfc-31.jpg",
  "target_members": "gtc-1.jpg ... gtc-31.jpg",
  "frame_count": 31,
  "source_resolution": "889x500",
  "preview_resolution": "890x500",
  "preview_fps": 10,
  "target_origin": "source_ground_truth",
  "workflow_status": "gate_a_draft"
}
Task 121 · indoor video dehazing

Indoor dehazing · REVIDE

用最短 Train 序列和 held-out Test 序列覆盖边界与 split。hazy 与 clear GT 是同一静态室内 scene 与机器人轨迹的两次采集,不是同曝光同步拍摄。

Gate B approved · PR merged
01

REVIDE Train · E007

Reference revide-train-E007 · shortest sequence

TRAIN 11 FRAMES SHORTEST BOUNDARY SOURCE GT
Input · hazycomplete sequence
Target · clear GTcomplete sequence

Prompt

Remove the visible haze from this indoor video while preserving the original scene, camera motion, frame timing, framing, colors, textures, and all scene content.

Metadata

Dataset / split
REVIDE Indoor · Train
Source sequence
E007
Mapping
hazy/E007 → gt/E007
Preview
2708×1800 · 11 frames · 8 fps · 1.375 s
Target origin
Publisher-provided haze-free pass
Review note
完整最短序列;检查 room 与 robot trajectory 对应。
查看 metadata.json 形式
{
  "task_id": 121,
  "sample_ref": "revide-train-E007",
  "dataset": "REVIDE_Indoor",
  "split": "Train",
  "source_sequence": "E007",
  "input_mapping": "hazy/E007",
  "target_mapping": "gt/E007",
  "frame_count": 11,
  "preview_resolution": "2708x1800",
  "preview_fps": 8,
  "target_origin": "source_ground_truth_separate_clear_pass",
  "coverage_role": "shortest_sequence"
}
02

REVIDE Test · C005

Reference revide-test-C005 · held-out split

TEST 43 FRAMES HELD-OUT SOURCE GT
Input · hazycomplete sequence
Target · clear GTcomplete sequence

Prompt

Remove the visible haze from this indoor video while preserving the original scene, camera motion, frame timing, framing, colors, textures, and all scene content.

Metadata

Dataset / split
REVIDE Indoor · Test
Source sequence
C005
Mapping
hazy/C005 → gt/C005
Preview
2708×1800 · 43 frames · 8 fps · 5.375 s
Target origin
Publisher-provided haze-free pass
Review note
确认左右未反转;两次采集的 scene 与轨迹应对应。
查看 metadata.json 形式
{
  "task_id": 121,
  "sample_ref": "revide-test-C005",
  "dataset": "REVIDE_Indoor",
  "split": "Test",
  "source_sequence": "C005",
  "input_mapping": "hazy/C005",
  "target_mapping": "gt/C005",
  "frame_count": 43,
  "preview_resolution": "2708x1800",
  "preview_fps": 8,
  "target_origin": "source_ground_truth_separate_clear_pass",
  "coverage_role": "held_out_split"
}
Task 130 · video motion deblurring

Motion deblurring · GoPro + BSD

一组 GoPro synthetic blur 与一组 BSD real beam-splitter pair,均保留发布方定义的 完整自然序列。历史 25 帧切片不在本页使用。

Gate A revised draft
01

GoPro Test · GOPR0384_11_00

Reference gopro-test-GOPR0384_11_00

SYNTHETIC BLUR TEST 100 FRAMES PRIVACY-SENSITIVE
Input · blur_gammacomplete sequence
Target · sharpcomplete sequence

Prompt

Remove motion blur from this video and restore a sharp video. Preserve the original scene content, camera trajectory, subject motion, frame order, timing, framing, resolution, colors, textures, and illumination; do not stabilize, crop, recolor, retime, add, or remove anything.

Metadata

Dataset / split
GoPro · test
Source sequence
GOPR0384_11_00
Mapping
blur_gamma → sharp
Preview
1280×720 · 100 frames · 15 fps · 6.67 s
Blur type
Publisher-provided synthetic gamma blur
Review note
街景、人群;检查人物、车辆、camera path 与隐私内容。
查看 metadata.json 形式
{
  "task_id": 130,
  "sample_ref": "gopro-test-GOPR0384_11_00",
  "dataset": "GoPro",
  "split": "test",
  "source_sequence": "GOPR0384_11_00",
  "input_mapping": "blur_gamma",
  "target_mapping": "sharp",
  "frame_range": "000001.png ... 000100.png",
  "frame_count": 100,
  "preview_resolution": "1280x720",
  "preview_fps": 15,
  "target_origin": "publisher_paired_sharp",
  "privacy_disposition": "review_required"
}
02

BSD 2ms–16ms Test · 015

Reference bsd-2ms16ms-test-015

REAL PAIRED BLUR TEST 150 FRAMES FAST PAN / DENSE LINES
Input · Blur/RGBcomplete sequence
Target · Sharp/RGBcomplete sequence

Prompt

Remove motion blur from this video and restore a sharp video. Preserve the original scene content, camera trajectory, subject motion, frame order, timing, framing, resolution, colors, textures, and illumination; do not stabilize, crop, recolor, retime, add, or remove anything.

Metadata

Dataset / split
BSD 2ms–16ms · test
Source sequence
015
Mapping
Blur/RGB → Sharp/RGB
Preview
640×480 · 150 frames · 15 fps · 10.00 s
Blur type
Real beam-splitter paired exposure
Review note
检查 fast pan、栅栏/线条,并区分光学对齐差异与去模糊。
查看 metadata.json 形式
{
  "task_id": 130,
  "sample_ref": "bsd-2ms16ms-test-015",
  "dataset": "BSD_2ms16ms",
  "split": "test",
  "source_sequence": "015",
  "input_mapping": "Blur/RGB",
  "target_mapping": "Sharp/RGB",
  "frame_range": "00000000.png ... 00000149.png",
  "frame_count": 150,
  "preview_resolution": "640x480",
  "preview_fps": 15,
  "target_origin": "publisher_paired_sharp",
  "pairing_note": "real_beam_splitter_pair"
}
Built locally on 2026-07-26 from Huayi's verified Task 118 / 121 / 130 review materials. Local summary only — no S3, GitHub, workflow decision, or public deployment.