README_zh.md

^{Lance: Unified Multimodal Modeling by Multi-Task Synergy}

Fengyi Fu^*, Mengqi Huang^*,✉, Shaojin Wu^*, Yunsheng Jiang^*, Yufei Huo, Jianzhu Guo^✉,§
Hao Li, Yinghang Song, Fei Ding, Qian He, Zheren Fu, Zhendong Mao, Yongdong Zhang
ByteDance
^* 共同一作    ^✉ 通讯作者    ^§ Project lead

English | 简体中文

🌟 亮点

Lance 是一个轻量级、原生统一的多模态模型，在单一框架下同时支持 图像与视频的理解、生成和编辑。

• 3B 规模高效强大。 仅使用 3B active parameters，Lance 即可在图像生成、图像编辑和视频生成等基准上取得强劲表现。
• 从零训练。 Lance 采用分阶段多任务训练配方，在 128 张 A100 GPU 的预算内从零完成训练。

🎨 演示

文生视频

视频编辑

多轮一致性编辑

智能视频生成

视频理解

问题： How many times did the person launch objects on the table? Options: (A) 3 (B) 2 (C) 4 回答： (A) 3	问题： The person makes sets of repeated actions. How many distinct repeated actions did the person do? Options: (A) 2 (B) 3 (C) 4 回答： (A) 2	问题： In which direction does the purple sphere move in the video? Options: (A) Down and to the right. (B) Up and to the left. (C) Up and to the right. (D) The object is stationary. 回答： (A) Down and to the right.
问题： What is the unrealistic phenomenon displayed in the video? Options: (A) The man can manipulate time via phone. (B) Man grabs an object through a phone screen. (C) Chocolate transforms into different objects. (D) Visible means of propulsion enables flight. 回答： (B) Man grabs an object through a phone screen.	问题： Offer a succinct account of the culinary process shown in this video. 回答： Add tomato puree and mix it well with chicken pieces.	问题： Provide a detailed description of the given video, capturing its key moments. 回答： In a sunlit meadow, a small tortoiseshell butterfly rests on a purple flower. A bee, with black and yellow stripes, lands on the same flower. The butterfly flaps its wings gently, while the bee busies itself, collecting nectar. The flower sways slightly in the breeze. The butterfly then takes off, and the bee follows, both heading to the next flower. The scene is a vivid display of insect interaction in a natural setting, with the colors of the butterfly and the bee contrasting against the green background of the meadow. The video captures this peaceful moment in a short 6-second duration.

文生图

图像编辑

图像理解

问题： Is the largest segment greater than sum of all the other segments? 回答： Yes, the largest segment in the pie chart is the blue one, which is the most populous segment.	问题： What percentage of respondents want better border security? 回答： 29%	问题： What is the license plate number of the car? 回答： The license plate number of the car is BX62 BFY.
问题： According to the data from the proprietary market research, how much amount was spent on the promotional meetings and events during 1998? 回答： According to the data from the proprietary market research, the total amount spent on the promotional meetings and events during 1998 was approximately $1.3 billion.	问题： What is the appearance of the Colosseum in Rome, Italy? 回答： The Colosseum in Rome, Italy is a magnificent ancient amphitheater made of stone and concrete. It is characterized by its large, arched openings and its semi-circular shape. The Colosseum has been partially restored, and its original color has faded, but the architectural details, like the columns and the semi-circular arches, are still visible. The lighting in the image highlights the grandeur and historical significance of the Colosseum, making it stand out against the backdrop of the evening sky.	问题： How does a total solar eclipse look like from Earth? 回答： A total solar eclipse looks like a dark circle with a bright, white outer edge. This is a result of the Earth's rotation and the Earth's shadow, which partially obscures the sun's light. The bright white edge of the eclipse is caused by the sun's high energy and its interaction with the Earth's atmosphere, while the dark part of the eclipse is due to the Earth's shadow and the surrounding air currents. The solar eclipse's shape, with its bright white edge and dark center, is similar to the shape of a full moon or a dark disk. It is a natural phenomenon that occurs in the atmosphere of the Earth and is an important part of the solar system.

🚀 安装

推荐环境

• 软件环境： Python 3.10+，CUDA 12.4+（必需）
• 硬件环境： 推理至少需要一张显存不低于 40GB 的 GPU

安装步骤

bash ./setup_env.sh

下载模型权重

请从 Lance-3B（HuggingFace 链接） 下载所需的全部模型权重，并放置到 downloads/ 目录下。

📚 使用方法

推理

Lance 为生成、编辑和理解任务提供了统一的命令行入口：

bash inference_lance.sh

• 运行前，请先在 inference_lance.sh 顶部配置推理参数。
• 支持任务： t2i、t2v、image_edit、video_edit、x2t_image 和 x2t_video。你也可以在 inference_lance.py 中修改 TASK_DEFAULT_CONFIGS，自定义每个任务默认使用的数据样例。

可用任务

任务名	说明	示例 JSON
t2v	文生视频	config/examples/t2v_example.json
t2i	文生图	config/examples/t2i_example.json
image_edit	图像编辑	config/examples/image_edit_example.json
video_edit	视频编辑	config/examples/video_edit_example.json
x2t_image	图像理解	config/examples/x2t_image_example.json
x2t_video	视频理解	config/examples/x2t_video_example.json

关于理解任务的示例文件：

• config/examples/x2t_image_example.json：用于图像理解示例，包括视觉问答和基于图像的推理。
• config/examples/x2t_video_example.json：用于视频理解示例，包括视频问答和视频描述。

参数说明

你可以在 inference_lance.sh 顶部配置以下超参数：

参数	默认值	说明
MODEL_PATH	"downloads/lance_3b"	下载后的 Lance 模型权重路径。
NUM_GPUS	1	用于推理的 GPU 数量。
VALIDATION_NUM_TIMESTEPS	30	去噪步数（例如 30 或 50）。
VALIDATION_TIMESTEP_SHIFT	3.5	Flow matching 调度中的 timestep shift 参数。
CFG_TEXT_SCALE	4.0	文本条件的 CFG（Classifier-Free Guidance）系数。
VALIDATION_DATA_SEED	42	用于复现实验的随机种子。
NUM_FRAMES	50	视频生成帧数（最大 121）。图像任务不使用该参数。
VIDEO_HEIGHT / VIDEO_WIDTH	768	空间分辨率。编辑任务不使用该参数（由输入图像/视频决定）。
RESOLUTION	"video_480p"	基础分辨率预设（如 image_768res 或 video_480p）。

Gradio

python lance_gradio_t2v_v2t.py --gpus 0 --server-port 7860

基准评测

DPG-Bench 评测

模型	# Params.	Global	Entity	Attribute	Relation	Other	Overall
仅生成模型
SDXL	3.5B	83.27	82.43	80.91	86.76	80.41	74.65
DALL-E 3	-	90.97	89.61	88.39	90.58	89.83	83.50
SD3-Medium	2B	87.90	91.01	88.83	80.70	88.68	84.08
FLUX.1-dev	12B	74.35	90.00	88.96	90.87	88.33	83.84
Qwen-Image	20B	91.32	91.56	92.02	94.31	92.73	88.32
统一模型
Janus-Pro-7B	7B	86.90	88.90	89.40	89.32	89.48	84.19
OmniGen2	4B	88.81	88.83	90.18	89.37	90.27	83.57
Show-o2	7B	89.00	91.78	89.96	91.81	91.64	86.14
BAGEL	7B	88.94	90.37	91.29	90.82	88.67	85.07
InternVL-U	1.7B	90.39	90.78	90.68	90.29	88.77	85.18
TUNA	7B	90.42	91.68	90.94	91.87	90.73	86.76
TUNA-2	7B	89.50	91.40	92.07	91.91	88.81	86.54
🌟 Lance (Ours)	3B	83.89	91.07	89.36	93.38	80.80	84.67

GenEval 评测

模型	# Params.	1-Obj.	2-Obj.	Count	Colors	Position	Attr.	Overall
仅生成模型
SDXL	3.5B	0.98	0.74	0.39	0.85	0.15	0.23	0.55
DALL-E 3	-	0.96	0.87	0.47	0.83	0.43	0.45	0.67
SD3-Medium	2B	0.99	0.94	0.72	0.89	0.33	0.60	0.74
FLUX.1-dev	12B	0.98	0.93	0.75	0.93	0.68	0.65	0.82
Qwen-Image	20B	0.99	0.92	0.89	0.88	0.76	0.77	0.87
统一模型
Janus-Pro-7B	7B	0.99	0.89	0.59	0.90	0.79	0.66	0.80
OmniGen2	4B	1.00	0.95	0.64	0.88	0.55	0.76	0.80
Show-o2	7B	1.00	0.87	0.58	0.92	0.52	0.62	0.76
BAGEL	7B	0.98	0.95	0.84	0.95	0.78	0.77	0.88
Mogao	7B	1.00	0.97	0.83	0.93	0.84	0.80	0.89
InternVL-U	1.7B	0.99	0.94	0.74	0.91	0.77	0.74	0.85
TUNA	7B	1.00	0.97	0.81	0.91	0.88	0.83	0.90
TUNA-2	7B	0.99	0.96	0.80	0.91	0.84	0.76	0.87
🌟 Lance (Ours)	3B	1.00	0.94	0.84	0.97	0.87	0.81	0.90

GEdit-Bench 评测

模型	# Params.	BC	CA	MM	MC	PB	ST	SA	SR	SRp	TM	TT	Avg/G_O
仅生成模型
Gemini 2.0	-	-	-	-	-	-	-	-	-	-	-	-	6.32
GPT Image 1	-	6.96	6.85	7.10	5.41	6.74	7.44	7.51	8.73	8.55	8.45	8.69	7.49
Qwen-Image-Edit	20B	8.23	8.30	7.33	8.05	7.49	6.74	8.57	8.09	8.29	8.48	8.50	8.01
统一模型
Lumina-DiMOO	8B	3.43	4.27	3.08	2.77	4.74	5.19	4.44	3.80	4.38	2.68	4.20	3.91
Ovis-U1	1.2B	7.49	6.88	6.21	4.79	5.98	6.46	7.49	7.25	7.27	4.48	6.31	6.42
BAGEL	7B	7.32	6.91	6.38	4.75	4.57	6.15	7.90	7.16	7.02	7.32	6.22	6.52
InternVL-U	1.7B	7.08	7.05	6.38	7.02	6.03	6.27	7.13	6.55	6.33	6.59	6.85	6.66
InternVL-U (w/ CoT)	1.7B	7.05	7.87	6.50	6.99	5.77	6.10	7.33	7.16	7.12	7.36	6.46	6.88
🌟 Lance (Ours)	3B	7.73	7.74	7.28	7.83	7.50	7.03	7.64	7.85	7.71	4.46	7.57	7.30

VBench 评测（视频生成）

类型	Model	# Params.	Total Score ↑
Gen. Only	ModelScope	1.7B	75.75
	LaVie	3B	77.08
	Show-1	6B	78.93
	AnimateDiff-V2	-	80.27
	VideoCrafter-2.0	-	80.44
	CogVideoX	5B	81.61
	Kling	-	81.85
	Open-Sora-2.0	-	81.71
	Gen-3	-	82.32
	Step-Video-T2V	30B	81.83
	Hunyuan Video	-	83.43
	Wan2.1-T2V	14B	83.69
Unified	HaproOmni	7B	78.10
	Emu3	8B	80.96
	VILA-U	7B	74.01
	Show-o2	2B	81.34
	TUNA	1.5B	84.06
	🌟 Lance (Ours)†	3B	85.11

运行基准评测

benchmarks/ 目录下提供了可直接运行的基准评测脚本：

基准	模态	脚本
GenEVAL（图像生成）	图像	benchmarks/image_gen/GenEVAL/sample_GenEVAL.sh
DPG（图像生成）	图像	benchmarks/image_gen/DPG/sample_DPG.sh
GEdit（图像编辑）	图像	benchmarks/image_gen/GEdit/sample_GEdit.sh
VBench（视频生成）	视频	benchmarks/video_gen/Vbench/sample_vbench.sh

📄 许可证

Copyright 2025 Bytedance Ltd. and/or its affiliates.

💖 引用

如果 Lance 对您的项目或研究有帮助，欢迎 🌟 本仓库，并使用以下 BibTeX 引用我们的工作：

@misc{lance2026,
  title  = {Lance: Unified Multimodal Modeling by Multi-Task Synergy},
  author = {Fengyi Fu and Mengqi Huang and Shaojin Wu and Yunsheng Jiang and Yufei Huo and Jianzhu Guo and Hao Li and Yinghang Song and Fei Ding and Qian He and Zheren Fu and Zhendong Mao and Yongdong Zhang},
  year   = {2026},
  note   = {Manuscript}
}

📞 联系方式

如有问题、反馈或合作需求，请联系 Mengqi Huang 和 Jianzhu Guo。