yiyexy commited on
Commit
cfd2b25
·
verified ·
1 Parent(s): 54be389

Model card: shorter quickstart, settings as superscripts in the video table

Browse files
Files changed (2) hide show
  1. README.md +33 -85
  2. README_CN.md +33 -85
README.md CHANGED
@@ -44,35 +44,35 @@ LLaVA-OneVision-2.5-4B-Instruct is a 4B vision-language model for single-image,
44
  | Benchmark | Qwen3.5-4B | OV2.5-4B (frames) | OV2.5-4B (codec) |
45
  |:---|:---:|:---:|:---:|
46
  | **General video understanding** | | | |
47
- | MotionBench | 60.1 | 62.4 | **64.2** |
48
- | TOMATO \* | 34.9 | 33.6 | **38.6** |
49
- | MVBench | 64.0 | 72.0 | **75.5** |
50
- | NextQA MC test | 80.7 | 86.1 | **86.4** |
51
- | TempCompass MC | 73.4 | 75.8 | **79.7** |
52
- | MMVU val | **52.6** | 44.1 | -- |
53
- | Video-MME-v2 | 15.5 | 16.4 | **17.2** |
54
- | EgoSchema subset | 68.2 | **80.0** | -- |
55
- | Video-Index 512f | 19.6 | **20.0** | 19.8 |
56
- | MMOU test-mini | -- | 46.0 | -- |
57
  | **VideoMME** | | | |
58
- | short, no subtitles | **81.7** | 79.7 | 81.1 |
59
- | medium, no subtitles | 70.3 | 70.6 | **74.1** |
60
- | long, no subtitles | **63.0** | 59.0 | 60.6 |
61
- | short, subtitles | -- | 82.4 | **82.7** |
62
- | medium, subtitles | -- | 75.4 | **75.6** |
63
- | long, subtitles | -- | 65.8 | **65.9** |
64
  | **Video spatial** | | | |
65
- | VSI-Bench | 59.2 | **68.6** | 67.7 |
66
- | ReVSI 64f | 47.1 | **55.9** | 55.5 |
67
  | **Long video** | | | |
68
- | LVBench | 56.5 | **62.0** | 61.6 |
69
- | MLVU dev | 73.8 | **75.3** | -- |
70
- | LongVideoBench val | 62.7 | 69.0 | **71.1** |
71
- | VideoEval-Pro | 62.8 | 64.4 | **64.6** |
72
  | **Temporal grounding (mIoU)** | | | |
73
- | TimeLens ActivityNet | 53.8 | 58.8 | **59.1** |
74
- | TimeLens Charades | 48.1 | 60.6 | **61.3** |
75
- | TimeLens QVHighlights | 67.5 | 68.4 | **70.3** |
76
 
77
  ### Image
78
 
@@ -87,7 +87,7 @@ LLaVA-OneVision-2.5-4B-Instruct is a 4B vision-language model for single-image,
87
  | OCRBench | **85.2** | 79.8 | EmbSpatial | 74.7 | **83.0** |
88
  | DocVQA val | **95.2** | 92.9 | MMSI-Bench | 31.4 | **33.3** |
89
 
90
- <sub>Compared with Qwen3.5-4B, thinking disabled. **Bold** is the best score in each row; `--` means not available. Each side reports its best score over the settings tried, from single runs. `*` On TOMATO the two sides use different frame settings. The image table shows selected benchmarks; full results, settings and protocol will be given in the technical report.</sub>
91
 
92
  ## Model Overview
93
 
@@ -140,9 +140,7 @@ with torch.inference_mode():
140
  print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
141
  ```
142
 
143
- For several images, put one `{"type": "image"}` per image in `content` and pass `images=` in the same order. The per-image budget defaults to `min_pixels=3136` / `max_pixels=4000000`.
144
-
145
- ### Video: frame sampling
146
 
147
  ```python
148
  messages = [{"role": "user", "content": [
@@ -159,16 +157,7 @@ inputs = processor(
159
  ).to(model.device)
160
  ```
161
 
162
- `target_fps` and `num_frames` are also accepted, and `max_frames` caps at 768. Each frame is prefixed with its real sample time, so temporal questions work directly:
163
-
164
- ```
165
- <0.0 seconds><|vision_start|><|image_pad|>×N<|vision_end|>
166
- <1.0 seconds><|vision_start|><|image_pad|>×N<|vision_end|>
167
- ```
168
-
169
- ### Video: codec mode
170
-
171
- Codec mode packs the informative regions of the video into a few canvases. Patches on one canvas may come from different source frames, and each keeps its own `(source frame, row, column)` position. Preprocessing comes from `codec-video-prep` (Linux x86_64, CPython 3.10/3.11; H.264, HEVC, VP9 and AV1 files):
172
 
173
  ```bash
174
  pip install -i https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple \
@@ -181,57 +170,16 @@ inputs = processor(text=text, videos=["example.mp4"],
181
  video_backend="codec", return_tensors="pt").to(model.device)
182
  ```
183
 
184
- Defaults live in the `codec` field of `preprocessor_config.json` (512 candidate frames, 64 canvases, 313600 pixels per canvas) and can be overridden per call with `codec_config={"target_canvas": 32, ...}`.
185
-
186
  ### Streaming
187
 
188
- Frames arrive one per user turn, and each turn the model emits a state token.
189
-
190
- | Mode | Call | Default system |
191
- |---|---|---|
192
- | Offline (default) | `apply_chat_template(messages, ...)` | `You are a helpful assistant.` |
193
- | Streaming | `apply_chat_template(messages, ..., streaming=True)` | `You are a real-time streaming video assistant.` |
194
-
195
- | Token | Id | Meaning |
196
- |---|---|---|
197
- | `<silence>` | 248077 | this turn: say nothing |
198
- | `<response>` | 248078 | this turn: speak |
199
- | `<question_memory>` / `</question_memory>` | 248079 / 248080 | a new question, written once when it is asked |
200
- | `<answer_memory>` / `</answer_memory>` | 248081 / 248082 | an earlier answer, re-injected once it is due |
201
- | `<captioning>` | 248083 | this turn: speak, as event-level captioning |
202
- | `<caption_begin>` | 248084 | user-side marker that this stream is a captioning task |
203
-
204
- Turns strictly alternate between user and assistant, with at most one frame per user turn. Give each frame its `timestamp`:
205
-
206
- ```python
207
- messages = [{"role": "user", "content": [
208
- {"type": "text", "text": '<question_memory>{"timestep": "00:00:00", "content": "What is he doing next?"}</question_memory>\n'},
209
- {"type": "image", "timestamp": 0.0},
210
- {"type": "text", "text": "\n"},
211
- ]}]
212
- text = processor.apply_chat_template(
213
- messages, tokenize=False, add_generation_prompt=True, streaming=True)
214
- ```
215
-
216
- ```
217
- <|im_start|>user
218
- <question_memory>{"timestep": "00:00:00", "content": "..."}</question_memory>
219
- <0.0 seconds><|vision_start|><|image_pad|>×N<|vision_end|>
220
- <|im_end|>
221
- <|im_start|>assistant
222
- <silence><|im_end|>
223
- ```
224
-
225
- The eight tokens are ordinary added tokens, so they survive `skip_special_tokens=True`: the state token is the output you read. The real-time loop around the model (frame scheduling, stopping, memory re-injection) is not part of this repository.
226
 
227
  ## Usage Tips
228
 
229
- - **Pass video paths, not pre-decoded frames.** Only a path lets the processor recover real timestamps. If you decode yourself, pass `videos=[{"frames": ..., "timestamps": ..., "frame_indices": ...}]`.
230
- - **Tune frames and pixels together.** Cost is roughly their product: 64–128 frames at 153664–321489 px for short video, 256 at 136900 for medium, 640–768 at 72900–108900 for long.
231
- - **Forward `patch_positions` in codec mode.** It comes back in the processor output and must reach `model.forward(...)`.
232
- - **Decode greedily and leave room.** Use `do_sample=False`, and a generous `max_new_tokens` on tasks where the model reasons before answering.
233
- - **Ask for the format you will parse.** For multiple choice, add `Answer with the option's letter from the given choices directly.` For grounding, state the coordinate range, e.g. integers 0–1000.
234
- - **Warm up before going multi-process.** Load the model once in a single process so the remote code is compiled before workers start.
235
 
236
  ## Citation
237
 
 
44
  | Benchmark | Qwen3.5-4B | OV2.5-4B (frames) | OV2.5-4B (codec) |
45
  |:---|:---:|:---:|:---:|
46
  | **General video understanding** | | | |
47
+ | MotionBench | 60.1 <sup>128f×153,664</sup> | 62.4 <sup>128f×153,664</sup> | **64.2** <sup>TC128×313,600</sup> |
48
+ | TOMATO \* | 34.9 <sup>16f×153,664, fps=2</sup> | 33.6 <sup>model default</sup> | **38.6** <sup>TC128×313,600</sup> |
49
+ | MVBench | 64.0 <sup>64f×153,664</sup> | 72.0 <sup>64f×153,664</sup> | **75.5** <sup>TC64×313,600</sup> |
50
+ | NextQA MC test | 80.7 <sup>64f×153,664</sup> | 86.1 <sup>64f×153,664</sup> | **86.4** <sup>TC64×313,600</sup> |
51
+ | TempCompass MC | 73.4 <sup>64f×153,664</sup> | 75.8 <sup>64f×153,664</sup> | **79.7** <sup>TC64×313,600</sup> |
52
+ | MMVU val | **52.6** <sup>64f×153,664</sup> | 44.1 <sup>64f×153,664</sup> | -- |
53
+ | Video-MME-v2 | 15.5 <sup>64f×330,000</sup> | 16.4 <sup>64f×330,000</sup> | **17.2** <sup>TC64×313,600</sup> |
54
+ | EgoSchema subset | 68.2 <sup>64f×153,664</sup> | **80.0** <sup>64f×153,664</sup> | -- |
55
+ | Video-Index 512f | 19.6 <sup>task frames, ≤262,144/f</sup> | **20.0** <sup>task frames, ≤262,144/f</sup> | 19.8 <sup>TC≤256×313,600</sup> |
56
+ | MMOU test-mini | -- | 46.0 <sup>model default</sup> | -- |
57
  | **VideoMME** | | | |
58
+ | short, no subtitles | **81.7** <sup>128f×321,489</sup> | 79.7 <sup>128f×321,489</sup> | 81.1 <sup>TC256×313,600</sup> |
59
+ | medium, no subtitles | 70.3 <sup>256f×136,900</sup> | 70.6 <sup>256f×136,900</sup> | **74.1** <sup>TC512×313,600</sup> |
60
+ | long, no subtitles | **63.0** <sup>1450f×120,000</sup> | 59.0 <sup>750f×250,000</sup> | 60.6 <sup>TC512×313,600</sup> |
61
+ | short, subtitles | -- | 82.4 <sup>128f×321,489</sup> | **82.7** <sup>TC256×313,600</sup> |
62
+ | medium, subtitles | -- | 75.4 <sup>256f×136,900</sup> | **75.6** <sup>TC256×313,600</sup> |
63
+ | long, subtitles | -- | 65.8 <sup>750f×250,000</sup> | **65.9** <sup>TC256×313,600</sup> |
64
  | **Video spatial** | | | |
65
+ | VSI-Bench | 59.2 <sup>128f×153,664</sup> | **68.6** <sup>128f×153,664</sup> | 67.7 <sup>TC128×313,600</sup> |
66
+ | ReVSI 64f | 47.1 <sup>64f×153,664</sup> | **55.9** <sup>64f×153,664</sup> | 55.5 <sup>TC64×313,600</sup> |
67
  | **Long video** | | | |
68
+ | LVBench | 56.5 <sup>2048f×120,000</sup> | **62.0** <sup>950f×200,000</sup> | 61.6 <sup>TC1024×156,800</sup> |
69
+ | MLVU dev | 73.8 <sup>512f×72,900</sup> | **75.3** <sup>512f×72,900</sup> | -- |
70
+ | LongVideoBench val | 62.7 <sup>700f×300,000</sup> | 69.0 <sup>1280f×80,000</sup> | **71.1** <sup>TC768×235,200</sup> |
71
+ | VideoEval-Pro | 62.8 <sup>2048f×100,000</sup> | 64.4 <sup>950f×200,000</sup> | **64.6** <sup>TC768×235,200</sup> |
72
  | **Temporal grounding (mIoU)** | | | |
73
+ | TimeLens ActivityNet | 53.8 <sup>2048f×409,600, fps=4</sup> | 58.8 <sup>2048f×409,600, fps=4</sup> | **59.1** <sup>TC128×313,600</sup> |
74
+ | TimeLens Charades | 48.1 <sup>2048f×409,600, fps=4</sup> | 60.6 <sup>2048f×409,600, fps=4</sup> | **61.3** <sup>TC128×313,600</sup> |
75
+ | TimeLens QVHighlights | 67.5 <sup>2048f×409,600, fps=4</sup> | 68.4 <sup>2048f×409,600, fps=4</sup> | **70.3** <sup>TC128×313,600</sup> |
76
 
77
  ### Image
78
 
 
87
  | OCRBench | **85.2** | 79.8 | EmbSpatial | 74.7 | **83.0** |
88
  | DocVQA val | **95.2** | 92.9 | MMSI-Bench | 31.4 | **33.3** |
89
 
90
+ <sub>Compared with Qwen3.5-4B, thinking disabled. **Bold** is the best score in each row; `--` means not available. The superscript on each video score is the setting that produced it: frames × pixels per frame, or for codec, canvases (TC) × pixels per canvas. Each side reports its best score over the settings tried, from single runs. `*` On TOMATO the two sides use different frame settings. The image table shows selected benchmarks; full results, settings and protocol will be given in the technical report.</sub>
91
 
92
  ## Model Overview
93
 
 
140
  print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
141
  ```
142
 
143
+ ### Video
 
 
144
 
145
  ```python
146
  messages = [{"role": "user", "content": [
 
157
  ).to(model.device)
158
  ```
159
 
160
+ For codec mode, install `codec-video-prep` and pass `video_backend="codec"`:
 
 
 
 
 
 
 
 
 
161
 
162
  ```bash
163
  pip install -i https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple \
 
170
  video_backend="codec", return_tensors="pt").to(model.device)
171
  ```
172
 
 
 
173
  ### Streaming
174
 
175
+ The checkpoint already supports streaming: pass `streaming=True` to `apply_chat_template`. The runtime library for real-time online inference is coming soon.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
176
 
177
  ## Usage Tips
178
 
179
+ - Pass video paths rather than pre-decoded frames, so the processor can recover real timestamps.
180
+ - Tune frame count and per-frame pixels together: cost is roughly their product.
181
+ - Decode greedily (`do_sample=False`) and leave a generous `max_new_tokens`.
182
+ - For multiple choice or grounding, state the answer format you expect in the prompt.
 
 
183
 
184
  ## Citation
185
 
README_CN.md CHANGED
@@ -44,35 +44,35 @@ LLaVA-OneVision-2.5-4B-Instruct 是一个 4B 视觉语言模型,支持单图
44
  | Benchmark | Qwen3.5-4B | OV2.5-4B(抽帧) | OV2.5-4B(codec) |
45
  |:---|:---:|:---:|:---:|
46
  | **通用视频理解** | | | |
47
- | MotionBench | 60.1 | 62.4 | **64.2** |
48
- | TOMATO \* | 34.9 | 33.6 | **38.6** |
49
- | MVBench | 64.0 | 72.0 | **75.5** |
50
- | NextQA MC test | 80.7 | 86.1 | **86.4** |
51
- | TempCompass MC | 73.4 | 75.8 | **79.7** |
52
- | MMVU val | **52.6** | 44.1 | -- |
53
- | Video-MME-v2 | 15.5 | 16.4 | **17.2** |
54
- | EgoSchema subset | 68.2 | **80.0** | -- |
55
- | Video-Index 512f | 19.6 | **20.0** | 19.8 |
56
- | MMOU test-mini | -- | 46.0 | -- |
57
  | **VideoMME** | | | |
58
- | short,无字幕 | **81.7** | 79.7 | 81.1 |
59
- | medium,无字幕 | 70.3 | 70.6 | **74.1** |
60
- | long,无字幕 | **63.0** | 59.0 | 60.6 |
61
- | short,有字幕 | -- | 82.4 | **82.7** |
62
- | medium,有字幕 | -- | 75.4 | **75.6** |
63
- | long,有字幕 | -- | 65.8 | **65.9** |
64
  | **视频空间** | | | |
65
- | VSI-Bench | 59.2 | **68.6** | 67.7 |
66
- | ReVSI 64f | 47.1 | **55.9** | 55.5 |
67
  | **长视频** | | | |
68
- | LVBench | 56.5 | **62.0** | 61.6 |
69
- | MLVU dev | 73.8 | **75.3** | -- |
70
- | LongVideoBench val | 62.7 | 69.0 | **71.1** |
71
- | VideoEval-Pro | 62.8 | 64.4 | **64.6** |
72
  | **时间定位(mIoU)** | | | |
73
- | TimeLens ActivityNet | 53.8 | 58.8 | **59.1** |
74
- | TimeLens Charades | 48.1 | 60.6 | **61.3** |
75
- | TimeLens QVHighlights | 67.5 | 68.4 | **70.3** |
76
 
77
  ### 图像
78
 
@@ -87,7 +87,7 @@ LLaVA-OneVision-2.5-4B-Instruct 是一个 4B 视觉语言模型,支持单图
87
  | OCRBench | **85.2** | 79.8 | EmbSpatial | 74.7 | **83.0** |
88
  | DocVQA val | **95.2** | 92.9 | MMSI-Bench | 31.4 | **33.3** |
89
 
90
- <sub>对比对象为关闭 thinking 的 Qwen3.5-4B。**加粗**为该行最高分,`--` 表示暂无结果。两边各取已测设置中的最高分,均为单次运行。`*` TOMATO 两边帧数设置不同。图像表为代表性 bench 的节选,完整结果、设置与评测口径见后续技术报告。</sub>
91
 
92
  ## 模型概览
93
 
@@ -140,9 +140,7 @@ with torch.inference_mode():
140
  print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
141
  ```
142
 
143
- 多图时每张图放一个 `{"type": "image"}`,`images=` 按同样顺序传入。单图像素预算默认 `min_pixels=3136` / `max_pixels=4000000`。
144
-
145
- ### 视频:抽帧
146
 
147
  ```python
148
  messages = [{"role": "user", "content": [
@@ -159,16 +157,7 @@ inputs = processor(
159
  ).to(model.device)
160
  ```
161
 
162
- 也接受 `target_fps` 和 `num_frames`,`max_frames` 上限 768。每帧前面带该帧的真实采样时刻,可以直接问时间类问题:
163
-
164
- ```
165
- <0.0 seconds><|vision_start|><|image_pad|>×N<|vision_end|>
166
- <1.0 seconds><|vision_start|><|image_pad|>×N<|vision_end|>
167
- ```
168
-
169
- ### 视频:codec 模式
170
-
171
- codec 模式把视频中信息量大的区域打包进少量画布。同一张画布上的 patch 可以来自不同源帧,各自保留 `(源帧, 行, 列)` 位置。预处理来自 `codec-video-prep`(Linux x86_64,CPython 3.10/3.11;支持 H.264、HEVC、VP9、AV1 文件):
172
 
173
  ```bash
174
  pip install -i https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple \
@@ -181,57 +170,16 @@ inputs = processor(text=text, videos=["example.mp4"],
181
  video_backend="codec", return_tensors="pt").to(model.device)
182
  ```
183
 
184
- 默认配置在 `preprocessor_config.json` 的 `codec` 字段(512 候选帧、64 画布、每画布 313600 像素),单次调用可用 `codec_config={"target_canvas": 32, ...}` 覆盖。
185
-
186
  ### 流式
187
 
188
- 画面每个 user 轮进来一帧,模型每轮输出一个状态 token。
189
-
190
- | 模式 | 调用 | 默认 system |
191
- |---|---|---|
192
- | 离线(默认) | `apply_chat_template(messages, ...)` | `You are a helpful assistant.` |
193
- | 流式 | `apply_chat_template(messages, ..., streaming=True)` | `You are a real-time streaming video assistant.` |
194
-
195
- | Token | Id | 含义 |
196
- |---|---|---|
197
- | `<silence>` | 248077 | 当前轮不发言 |
198
- | `<response>` | 248078 | 当前轮发言 |
199
- | `<question_memory>` / `</question_memory>` | 248079 / 248080 | 新问题,在提问时刻写入一次 |
200
- | `<answer_memory>` / `</answer_memory>` | 248081 / 248082 | 此前回答到期后回插一次 |
201
- | `<captioning>` | 248083 | 当前轮发言,且是事件级 caption |
202
- | `<caption_begin>` | 248084 | user 侧标记,说明这条流是 caption 任务 |
203
-
204
- user 与 assistant 严格交替,每个 user 轮最多一帧。每帧给上 `timestamp`:
205
-
206
- ```python
207
- messages = [{"role": "user", "content": [
208
- {"type": "text", "text": '<question_memory>{"timestep": "00:00:00", "content": "接下来他在做什么?"}</question_memory>\n'},
209
- {"type": "image", "timestamp": 0.0},
210
- {"type": "text", "text": "\n"},
211
- ]}]
212
- text = processor.apply_chat_template(
213
- messages, tokenize=False, add_generation_prompt=True, streaming=True)
214
- ```
215
-
216
- ```
217
- <|im_start|>user
218
- <question_memory>{"timestep": "00:00:00", "content": "..."}</question_memory>
219
- <0.0 seconds><|vision_start|><|image_pad|>×N<|vision_end|>
220
- <|im_end|>
221
- <|im_start|>assistant
222
- <silence><|im_end|>
223
- ```
224
-
225
- 八个 token 都是普通 added token,`skip_special_tokens=True` 不会过滤它们:状态 token 就是要读取的输出。围绕模型的实时循环(取帧调度、停止生成、记忆回插)不在本仓库内。
226
 
227
  ## 使用建议
228
 
229
- - **传视频路径,不要传自己解码好的帧。** 只有路径才能让 processor 拿到真实时间戳。必须自己解码时,传 `videos=[{"frames": ..., "timestamps": ..., "frame_indices": ...}]`。
230
- - **帧数与像素一起调。** 开销约等于两者之积:短视频 64~128 帧 × 153664~321489 px,中等 256 帧 × 136900,长视频 640~768 帧 × 72900~108900。
231
- - **codec 模式下转发 `patch_positions`。** 它在 processor 输出里,必须传给 `model.forward(...)`。
232
- - **贪心解码并留足长度。** 用 `do_sample=False`;需要先推理再作答的任务给足 `max_new_tokens`。
233
- - **要什么格式就在 prompt 里说。** 选择题加上 `Answer with the option's letter from the given choices directly.`;grounding 任务写明坐标范围,例如 0~1000 的整数。
234
- - **多进程前先预热。** 先在单进程里完整加载一次模型,让 remote code 编译完成后再启动多个 worker。
235
 
236
  ## 引用
237
 
 
44
  | Benchmark | Qwen3.5-4B | OV2.5-4B(抽帧) | OV2.5-4B(codec) |
45
  |:---|:---:|:---:|:---:|
46
  | **通用视频理解** | | | |
47
+ | MotionBench | 60.1 <sup>128f×153,664</sup> | 62.4 <sup>128f×153,664</sup> | **64.2** <sup>TC128×313,600</sup> |
48
+ | TOMATO \* | 34.9 <sup>16f×153,664, fps=2</sup> | 33.6 <sup>model default</sup> | **38.6** <sup>TC128×313,600</sup> |
49
+ | MVBench | 64.0 <sup>64f×153,664</sup> | 72.0 <sup>64f×153,664</sup> | **75.5** <sup>TC64×313,600</sup> |
50
+ | NextQA MC test | 80.7 <sup>64f×153,664</sup> | 86.1 <sup>64f×153,664</sup> | **86.4** <sup>TC64×313,600</sup> |
51
+ | TempCompass MC | 73.4 <sup>64f×153,664</sup> | 75.8 <sup>64f×153,664</sup> | **79.7** <sup>TC64×313,600</sup> |
52
+ | MMVU val | **52.6** <sup>64f×153,664</sup> | 44.1 <sup>64f×153,664</sup> | -- |
53
+ | Video-MME-v2 | 15.5 <sup>64f×330,000</sup> | 16.4 <sup>64f×330,000</sup> | **17.2** <sup>TC64×313,600</sup> |
54
+ | EgoSchema subset | 68.2 <sup>64f×153,664</sup> | **80.0** <sup>64f×153,664</sup> | -- |
55
+ | Video-Index 512f | 19.6 <sup>task frames, ≤262,144/f</sup> | **20.0** <sup>task frames, ≤262,144/f</sup> | 19.8 <sup>TC≤256×313,600</sup> |
56
+ | MMOU test-mini | -- | 46.0 <sup>model default</sup> | -- |
57
  | **VideoMME** | | | |
58
+ | short,无字幕 | **81.7** <sup>128f×321,489</sup> | 79.7 <sup>128f×321,489</sup> | 81.1 <sup>TC256×313,600</sup> |
59
+ | medium,无字幕 | 70.3 <sup>256f×136,900</sup> | 70.6 <sup>256f×136,900</sup> | **74.1** <sup>TC512×313,600</sup> |
60
+ | long,无字幕 | **63.0** <sup>1450f×120,000</sup> | 59.0 <sup>750f×250,000</sup> | 60.6 <sup>TC512×313,600</sup> |
61
+ | short,有字幕 | -- | 82.4 <sup>128f×321,489</sup> | **82.7** <sup>TC256×313,600</sup> |
62
+ | medium,有字幕 | -- | 75.4 <sup>256f×136,900</sup> | **75.6** <sup>TC256×313,600</sup> |
63
+ | long,有字幕 | -- | 65.8 <sup>750f×250,000</sup> | **65.9** <sup>TC256×313,600</sup> |
64
  | **视频空间** | | | |
65
+ | VSI-Bench | 59.2 <sup>128f×153,664</sup> | **68.6** <sup>128f×153,664</sup> | 67.7 <sup>TC128×313,600</sup> |
66
+ | ReVSI 64f | 47.1 <sup>64f×153,664</sup> | **55.9** <sup>64f×153,664</sup> | 55.5 <sup>TC64×313,600</sup> |
67
  | **长视频** | | | |
68
+ | LVBench | 56.5 <sup>2048f×120,000</sup> | **62.0** <sup>950f×200,000</sup> | 61.6 <sup>TC1024×156,800</sup> |
69
+ | MLVU dev | 73.8 <sup>512f×72,900</sup> | **75.3** <sup>512f×72,900</sup> | -- |
70
+ | LongVideoBench val | 62.7 <sup>700f×300,000</sup> | 69.0 <sup>1280f×80,000</sup> | **71.1** <sup>TC768×235,200</sup> |
71
+ | VideoEval-Pro | 62.8 <sup>2048f×100,000</sup> | 64.4 <sup>950f×200,000</sup> | **64.6** <sup>TC768×235,200</sup> |
72
  | **时间定位(mIoU)** | | | |
73
+ | TimeLens ActivityNet | 53.8 <sup>2048f×409,600, fps=4</sup> | 58.8 <sup>2048f×409,600, fps=4</sup> | **59.1** <sup>TC128×313,600</sup> |
74
+ | TimeLens Charades | 48.1 <sup>2048f×409,600, fps=4</sup> | 60.6 <sup>2048f×409,600, fps=4</sup> | **61.3** <sup>TC128×313,600</sup> |
75
+ | TimeLens QVHighlights | 67.5 <sup>2048f×409,600, fps=4</sup> | 68.4 <sup>2048f×409,600, fps=4</sup> | **70.3** <sup>TC128×313,600</sup> |
76
 
77
  ### 图像
78
 
 
87
  | OCRBench | **85.2** | 79.8 | EmbSpatial | 74.7 | **83.0** |
88
  | DocVQA val | **95.2** | 92.9 | MMSI-Bench | 31.4 | **33.3** |
89
 
90
+ <sub>对比对象为关闭 thinking 的 Qwen3.5-4B。**加粗**为该行最高分,`--` 表示暂无结果。视频分数的上标是取得该分数时的设置:帧数 × 每帧像素;codec 为画布数(TC)× 每画布像素。两边各取��测设置中的最高分,均为单次运行。`*` TOMATO 两边帧数设置不同。图像表为代表性 bench 的节选,完整结果、设置与评测口径见后续技术报告。</sub>
91
 
92
  ## 模型概览
93
 
 
140
  print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
141
  ```
142
 
143
+ ### 视频
 
 
144
 
145
  ```python
146
  messages = [{"role": "user", "content": [
 
157
  ).to(model.device)
158
  ```
159
 
160
+ 使用 codec 模式时,安装 `codec-video-prep` 并传入 `video_backend="codec"`:
 
 
 
 
 
 
 
 
 
161
 
162
  ```bash
163
  pip install -i https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple \
 
170
  video_backend="codec", return_tensors="pt").to(model.device)
171
  ```
172
 
 
 
173
  ### 流式
174
 
175
+ 权重本身已支持流式:给 `apply_chat_template` 传 `streaming=True` 即可。用于实时在线推理的运行库即将发布。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
176
 
177
  ## 使用建议
178
 
179
+ - 传视频路径,不要传自己解码好的帧,这样 processor 才能拿到真实时间戳。
180
+ - 帧数与逐帧像素一起调:开销约等于两者之积。
181
+ - 使用贪心解码(`do_sample=False`),并给足 `max_new_tokens`。
182
+ - 选择题或 grounding 任务,在 prompt 里写明期望的回答格式。
 
 
183
 
184
  ## 引用
185