PaperlayerView exact revision on arXiv

Source-linked AI summary

Seedance 2.0: Advancing Video Generation for World Complexity

Team Seedance, De Chen, Liyang Chen, Xin Chen, Ying Chen, Zhuo Chen, Zhuowei Chen, Feng Cheng, Tianheng Cheng, Yufeng Cheng, Mojie Chi, Xuyan Chi, Jian Cong, Qinpeng Cui, Fei Ding, Qide Dong, Yujiao Du, Haojie Duanmu, Junliang Fan, Jiarui Fang, Jing Fang, Zetao Fang, Chengjian Feng, Yu Gao, Diandian Gu, Dong Guo, Hanzhong Guo, Qiushan Guo, Boyang Hao, Hongxiang Hao, Haoxun He, Jiaao He, Qian He, Tuyen Hoang, Heng Hu, Ruoqing Hu, Yuxiang Hu, Jiancheng Huang, Weilin Huang, Zhaoyang Huang, Zhongyi Huang, Jishuo Jin, Ming Jing, Ashley Kim, Shanshan Lao, Yichong Leng, Bingchuan Li, Gen Li, Haifeng Li, Huixia Li, Jiashi Li, Ming Li, Xiaojie Li, Xingxing Li, Yameng Li, Yiying Li, Yu Li, Yueyan Li, Chao Liang, Han Liang, Jianzhong Liang, Ying Liang, Wang Liao, J. H. Lien, Shanchuan Lin, Xi Lin, Feng Ling, Yue Ling, Fangfang Liu, Jiawei Liu, Jihao Liu, Jingtuo Liu, Shu Liu, Sichao Liu, Wei Liu, Xue Liu, Zuxi Liu, Ruijie Lu, Lecheng Lyu, Jingting Ma, Tianxiang Ma, Xiaonan Nie, Jingzhe Ning, Junjie Pan, Xitong Pan, Ronggui Peng, Xueqiong Qu, Yuxi Ren, Yuchen Shen, Guang Shi, Lei Shi, Yinglong Song, Fan Sun, Li Sun, Renfei Sun, Wenjing Tang, Boyang Tao, Zirui Tao, Dongliang Wang, Feng Wang, Hulin Wang, Ke Wang, Qingyi Wang, Rui Wang, Shuai Wang, Shulei Wang, Weichen Wang, Xuanda Wang, Yanhui Wang, Yue Wang, Yuping Wang, Yuxuan Wang, Zijie Wang, Ziyu Wang, Guoqiang Wei, Meng Wei, Di Wu, Guohong Wu, Hanjie Wu, Huachao Wu, Jian Wu, Jie Wu, Ruolan Wu, Shaojin Wu, Xiaohu Wu, Xinglong Wu, Yonghui Wu, Ruiqi Xia, Xin Xia, Xuefeng Xiao, Shuang Xu, Bangbang Yang, Jiaqi Yang, Runkai Yang, Tao Yang, Yihang Yang, Zhixian Yang, Ziyan Yang, Fulong Ye, Bingqian Yi, Xing Yin, Yongbin You, Linxiao Yuan, Weihong Zeng, Xuejiao Zeng, Yan Zeng, Siyu Zhai, Zhonghua Zhai, Bowen Zhang, Chenlin Zhang, Heng Zhang, Jun Zhang, Manlin Zhang, Peiyuan Zhang, Shuo Zhang, Xiaohe Zhang, Xiaoying Zhang, Xinyan Zhang, Xinyi Zhang, Yichi Zhang, Zixiang Zhang, Haiyu Zhao, Huating Zhao, Liming Zhao, Yian Zhao, Guangcong Zheng, Jianbin Zheng, Xiaozheng Zheng, Zerong Zheng, Kuan Zhu, Feilong Zuo

arXiv:2604.14148v1cs.CV

TL;DR

Video generation has been moving from short clips with limited controllability toward synthesis guided by diverse control signals. Seedance 2.0 adopts multimodal audio-video generation, and reports more accurate instruction following and more stable motion.

  • Problem

    Video generation has been moving from short clips with limited controllability toward synthesis guided by diverse control signals.

    Source: §1 · p. 1

  • Method

    Seedance 2.0 supports multimodal reference inputs and controllable video generation guided by text, images, video, and audio.

    Source: §1 · p. 2

  • Results

    The model reports more accurate compliance with complex instructions, more stable motion, and competitive usability in real-world production settings.

    Source: §1 · p. 1

  • Takeaways & Limitations

    The model is applicable to diverse creative production scenarios and more sophisticated production workflows.

    Source: §1 · p. 1

    Areas for improvement include deformation artifacts, motion plausibility in edge cases, visual and audio noise, audio distortion, and lip-sync errors in multi-speaker scenes.

    Source: §2.1 · p. 4

Abstract

from arXiv · show

Seedance 2.0 is a new native multi-modal audio-video generation model, officially released in China in early February 2026. Compared with its predecessors, Seedance 1.0 and 1.5 Pro, Seedance 2.0 adopts a unified, highly efficient, and large-scale architecture for multi-modal audio-video joint generation. This allows it to support four input modalities: text, image, audio, and video, by integrating one of the most comprehensive suites of multi-modal content reference and editing capabilities available in the industry to date. It delivers substantial, well-rounded improvements across all key sub-dimensions of video and audio generation. In both expert evaluations and public user tests, the model has demonstrated performance on par with the leading levels in the field. Seedance 2.0 supports direct generation of audio-video content with durations ranging from 4 to 15 seconds, with native output resolutions of 480p and 720p. For multi-modal inputs as reference, its current open platform supports up to 3 video clips, 9 images, and 3 audio clips. In addition, we provide Seedance 2.0 Fast version, an accelerated variant of Seedance 2.0 designed to boost generation speed for low-latency scenarios. Seedance 2.0 has delivered significant improvements to its foundational generation capabilities and multi-modal generation performance, bringing an enhanced creative experience for end users.

1 Introduction

Seedance 2.0 shifts toward highly controllable video synthesis using diverse control signals, with multimodal reference and editing capabilities across video, audio, and creative workflows. The introduction highlights improvements in motion realism, audio-video synchronization, and production usability while noting that the model remains imperfect.

Introduction

  • Seedance 2.0 shifts from short, limited-control clips to highly controllable synthesis, with multimodal reference and editing for subject, motion, style, effects, content, and video extension.It improves complex-instruction compliance, motion stability, multi-shot narrative, cross-frame consistency, and usability in production workflows.Source: §1 · p. 1
  • Seedance 2.0 improves human-motion naturalness, temporal coherence, and physical plausibility, with realistic close-ups and higher usability in complex motion and multi-subject interactions.The model is described as generating temporally precise interactions while following real-world motion laws and reducing common artifacts.Source: §1 · pp. 1–2
  • Seedance 2.0 accepts text, image, video, and audio references, supports storyboard-based generation, and adds controllable editing and prompt-aligned video continuation.It can preserve subject identity across complex scripts and autonomously plan shot sequences and visual presentation.Source: §1 · p. 2
  • Seedance 2.0 generates synchronized binaural audio-video, including background audio, ambient effects, and narration aligned with visual actions.The introduction describes natural sound design and precise temporal alignment as supporting immersive, professional-grade content.Source: §1 · p. 2
  • Seedance 2.0 targets advertising, film and television effects, game animation, and commentary, with AI generation intended to reduce production costs and cycle time.The introduction attributes this potential to cross-scene adaptability and replacing some complex effects and live-action workflows.Source: §1 · p. 2
  • Seedance 2.0 extends the unified architecture behind Seedance 1.5 [16], acknowledges remaining output imperfections, and uses structured safety assessment throughout model iteration.The stated safety aim is to evaluate and mitigate risks in support of responsible, compliant, and ethically aligned development.Source: §1 · p. 2

2 Evaluation · 2.1 Overview

The evaluation benchmark covers audio-video generation, reference-based generation, and video editing. Seedance 2.0 ranks first across video and audio dimensions in T2V, I2V, and R2V, with strengths in video quality, audio-visual alignment, and reference consistency alongside remaining limitations.

2.1 Overview

  • The team established an expert-informed benchmark covering audio-video generation, reference-based generation, and video editing, assessed across six core capability dimensions.The dimensions include reference-based generation, complex audio-video instruction following, motion stability, cinematographic expression, audio and video expressiveness, and audio-visual alignment.Source: §2.1 · p. 3
  • Seedance 2.0 achieves the highest scores across every evaluated dimension in T2V, I2V, and R2V, ranking first across video and audio dimensions on all three tasks.Figure 1 compares overall performance across the three generation tasks.Source: Figure 1 · p. 3, §2.1 · pp. 3–4
  • Reference-based generation supports multimodal references, editing, and continuation, with strengths in instruction following, photorealism, and consistency of appearance, voice, action logic, effects, and narrative.The overview identifies multi-subject consistency, text restoration, and complex editing as areas needing optimization.Source: §2.1 · p. 4
  • Video results highlight stable complex motion, strong instruction following, cinematic realism, and professional camera movement and narrative pacing.The model handles long-script and open-ended prompts, high-amplitude actions and subtle expressions, and preserves reference-image art styles in I2V.Source: §2.1 · pp. 3–4
  • Audio results show detailed dual-channel expressiveness, contextual melodies, and tighter audio-visual synchronization, with improved performance on Chinese dialects, opera, and singing.The reported alignment includes lip movements, dialogue, sound effects, background audio, and beat-matching.Source: §2.1 · pp. 3–4
  • Remaining weaknesses include minor deformation artifacts, implausible motion in edge cases, visual noise, audio distortion, and lip-sync errors in multi-speaker scenes.The overview notes room to improve text restoration and complex editing performance as well as multi-subject consistency.Source: §2.1 · p. 4

2.2 Evaluation Framework

SeedVideoBench 2.0 expands evaluation to multimodal generation, narrative quality, and multilingual coverage, while Arena.AI results show Seedance 2.0 leading both video-generation leaderboards at 720p. The framework combines objective metrics with expert and user judgments to assess generation quality and human preference.

2.2 Evaluation Framework

  • SeedVideoBench 2.0 adds multimodal generation, narrative quality, and multilingual coverage, refines audio-expressiveness assessment, and separates automated objective metrics from blind expert review of subjective qualities.Expert evaluators from advertising and game production rate narrative and aesthetic quality; a separate realism study tests whether evaluators can distinguish generated outputs from real clips, informing aesthetic tuning.Source: §2.2 · p. 4, §2.2.1 · p. 4

2.2.1 SeedVideoBench 2.0

  • Multimodal Task Following evaluates reference, editing, and extension scenarios across fine-grained task types, with task groups spanning reference, editing, timeline extension, and paired combinations.The framework defines task following and generation consistency while also evaluating baseline prompt-following and motion quality in multimodal settings; task coverage includes subject, motion, style, scene, visual effects, and audio.Source: §2.2.1 · pp. 4–5
  • The benchmark measures reference alignment and editing consistency, and assesses narrative quality through cinematography, plot design, and stylistic aesthetics alongside vividness and aesthetics.Narrative review considers story coherence, emotional impact, shot logic and pacing, whether brief prompts yield engaging plots, and visual choices such as lighting, framing, composition, and color grading.Source: §2.2.1 · p. 5

2.2.2 Arena.AI Results

  • 1450 (±15) and 1449 (±11) Elo scores rank Dreamina Seedance 2.0 720p #1 on Text-to-Video and Image-to-Video, leading by 79 and 29 points, respectively.The leaderboards reflect pairwise human preferences; Rank Spread is 1↔1 on both.Source: Figure 2 · p. 6, §2.2.2 · p. 6

2.3 Text-to-Video Evaluation on SeedVideoBench 2.0

On SeedVideoBench 2.0 text-to-video evaluation, Seedance 2.0 leads overall and ranks first across most fine-grained motion, prompt-following, aesthetics, audio, and synchronization categories. Its reported advantage is especially pronounced in audio satisfaction and quality.

2.3.1 Overall Results

  • Seedance 2.0 ranks first on all six dimensions, exceeds 3.4 on each, and improves over Seedance 1.5 [16] by 0.86 points on average, with a 1.36-point motion-quality gain.It scores 3.75 on both motion quality and audio-visual sync, at least 0.65 points ahead of the runner-up; competitors generally struggle on audio, while Seedance 2.0 exceeds 3.5 on all three audio dimensions.Source: Table 1 · p. 7, §2.3.1 · p. 7
  • Seedance 2.0 is the only model above 83% usability on every dimension and exceeds 51% satisfaction on each, while no competitor exceeds 44% on any dimension.Audio quality satisfaction is 62.05% versus below 10% for every competitor, and audio-visual sync satisfaction is 68.30% versus a competitor high of 25.45%. Only Seedance 2.0 has delight-rated audio quality outputs (6.70%); its audio prompt-following delight rate is 26.92%, versus 11.68% for Sora 2 Pro.Source: Table 2 · p. 7, §2.3.1 · p. 7

2.3.2 Motion Quality

  • Seedance 2.0 ranks first in 29 of 30 motion categories, scoring 3.29–4.43, and improves over Seedance 1.5 by more than 1.5 points on three cited weak categories.It ties with Kling 3.0 on group coordinated motion; its multi-entity feature match score is 4.43, while physical feedback, natural phenomena, and intense sports motion each improve by over 1.5 points.Source: Table 3 · p. 9, §2.3.2 · p. 9

2.3.3 Video Prompt Following

  • Seedance 2.0 ranks first in 27 of 30 video prompt-following categories, scoring 2.71–4.29, with its largest gains over Seedance 1.5 on text-related categories.Creative text improves from 1.86 to 3.43, short text from 2.00 to 3.57, and text overlay from 2.15 to 3.31; counter-reality instructions score 4.29. Sora 2 Pro leads on abstract challenges and framing/composition.Source: §2.3.3 · pp. 9–10, Table 4 · p. 10

2.3.4 Video Aesthetics

  • Seedance 2.0 ranks first or tied for first in 28 of 30 aesthetics categories, scoring 2.79–4.14, and leads on visual effects, scene design, lighting and color, and realistic detail.It does not lead on consumer visual effects or surreal motion, and ties on anthropomorphic motion and advanced camera movement.Source: §2.3.4 · p. 10, Table 5 · p. 11

2.3.5 Audio Quality

  • Seedance 2.0 ranks first in all 17 audio-quality categories, scoring 2.82–4.17, with its strongest results in English, voice-plus-action interaction, minority languages, and ambient/background sound.Its largest gains over Seedance 1.5 are Chinese opera (2.50 to 3.75), English (3.00 to 4.17), and singing/rap (2.71 to 3.71); no competitor exceeds 3.2 except Sora 2 Pro on singing/rap (3.67).Source: Table 6 · p. 11, §2.3.5 · pp. 11–12

2.3.6 Audio-Visual Sync

  • Seedance 2.0 ranks first in 16 of 17 audio-visual synchronization categories, scoring 2.86–4.17, while competitors commonly show lip-speech mismatches and action-sound offsets.It ties with Seedance 1.5 on off-screen voice (2.86); its largest gains include Chinese multi-person dialogue (2.36 to 3.86), object interaction sound (2.65 to 3.82), and animal sound (2.79 to 3.93).Source: Table 7 · p. 12, §2.3.6 · p. 12

2.3.7 Audio Prompt Following

  • Seedance 2.0 ranks first in 16 of 17 audio prompt-following categories, scoring 2.91–4.25, and is strong on complex multilingual, dialect-specific, and diverse sound instructions.It ties with Kling 3.0 on off-screen voice (3.14); its largest gains over Seedance 1.5 include Chinese opera (1.75 to 3.50), singing/rap (2.14 to 3.71), and animal sound (2.50 to 3.86). Sora 2 Pro is the strongest competitor overall, while Chinese dialect is below 2.0 for five of six models.Source: §2.3.7 · pp. 12–13, Table 8 · p. 13

2.4 Image-to-Video Evaluation on SeedVideoBench 2.0

Seedance 2.0 ranks first across the reported I2V video and audio dimensions, with detailed evaluations showing strengths across visual and audio tasks. Difficult areas remain, including advanced camera movement, group motion, and physical simulations.

2.4.1 Overall Results

  • Seedance 2.0 ranks first on all six I2V dimensions, scoring 3.31–3.70 while no competitor exceeds 3.18.Its video scores are 3.35 for motion quality, 3.46 for video prompt following, and 3.31 for image preservation; its audio scores are 3.61, 3.54, and 3.70.Source: Table 9 · p. 13, §2.4.1 · p. 13
  • Seedance 2.0 is the only model with usability above 87% on all six dimensions, and its audio usability and satisfaction substantially exceed competitors’.Human evaluation also notes dynamic motion, expressive faces, varied camera work, coherent special art styles, nuanced dialogue, and well-layered audio.Source: §2.4.1 · pp. 13–14, Table 10 · p. 14

2.4.2 Detailed Visual Evaluation Results

  • Seedance 2.0 leads on all three script-controlled prompt-abstraction metrics and on motion quality and image preservation for UGC prompts.For script-controlled generation, video prompt following is 3.87 versus Kling 3.0’s 3.00 and Wan 2.6’s 2.50; on compound multi-instruction tasks, it scores 4.00 on motion quality and 3.75 on video prompt following, over 1 point ahead of Kling 3.0 on motion quality.Source: Table 11 · p. 14, §2.4.2 · p. 15, Table 12 · p. 15
  • Seedance 2.0 leads on motion quality in all three complex-camera subcategories, while advanced camera movement is hardest; reported complex-motion scores include sports MQ 3.73 and combat MQ 3.63.On advanced-camera motion quality, Seedance 2.0 and Kling 3.0 tie at 2.71, and no model exceeds 3.14 on any metric. Seedance 2.0’s sports VPF is 3.93; its combat MQ is 3.63 versus 2.25 for Kling 3.0 and Seedance 1.5 Pro.Source: §2.4.2 · p. 15
  • Seedance 2.0’s strongest interaction results are same-type interaction and cross-type video prompt following, and it leads on motion quality in all four creative subcategories.Group motion remains hard for all models; Seedance 2.0 scores 3.00/3.00/2.88, while its creative visual-effects scores are 3.50/3.38/3.50.Source: Table 15 · p. 16, §2.4.2 · p. 16, Table 16 · p. 16
  • Seedance 2.0 leads on motion quality across physical-law tasks and on motion quality and video prompt following for both complex-reference tasks.Physical simulations remain challenging for every model, and Kling 3.0 exceeds Seedance 2.0 on natural-phenomena image preservation (3.56 vs. 3.44); Seedance 2.0’s visual averages are 3.35 for motion quality, 3.31 for image preservation, and 3.46 for video prompt following.Source: Table 17 · p. 17, §2.4.2 · p. 17, Table 18 · p. 17

2.4.3 Detailed Audio Evaluation Results

  • Seedance 2.0 leads on audio quality across all Chinese voice subcategories and scores at least 3.50 on audio quality for all six non-Chinese languages.Chinese conversation is its strongest Chinese voice task (audio quality 3.92, audio prompt following 4.08); non-Chinese peaks include Spanish audio quality and audio-visual sync at 4.14, and English audio prompt following at 4.20.Source: §2.4.3 · pp. 17–18
  • Seedance 2.0 leads on all four composite-voice tasks, audio-visual sync for all five sound-effects tasks, and audio quality for four of five sound-effects tasks.It also scores 3.80 on instruments-and-audio prompt following and 3.53 on dual-channel sync; its audio averages are 3.61/3.54/3.70, ahead of Seedance 1.5 Pro at 3.07/2.95/3.10.Source: §2.4.3 · pp. 18–19

2.5 Reference-to-Video Evaluation on SeedVideoBench 2.0

Seedance 2.0 ranks first across all five overall R2V dimensions and supports more reference-task types than the evaluated competitors. Its strongest advantages include motion and reference alignment, while video extension remains a weakness relative to Veo 3.1.

2.5.1 Quantitative Results

  • Seedance 2.0 ranks first on all five R2V dimensions, scoring 2.50 on multimodal task following, 2.52 on prompt following, 3.54 on editing consistency, 3.03 on reference alignment, and 3.24 on motion quality.Its largest overall leads are in motion quality and reference alignment; editing consistency is the closest comparison, with Kling 3.0 trailing by 0.17.Source: Table 24 · p. 19, §2.5.1 · p. 20
  • Seedance 2.0 supports 20 of 22 R2V tasks, the broadest coverage among evaluated models, including seven visual-effects, creative-reference, continuation, and extension tasks competitors cannot handle.All models lack the same two tasks: subject audio-visual plus audio reference, and video audio editing.Source: Table 25 · p. 20, §2.5.1 · p. 20
  • Seedance 2.0 leads on image- and video-based subject-reference alignment, scoring 3.18 and 3.35, respectively, while its image-based task-following score is 2.80.For video-based subject reference, it scores 2.95 on task following, with 95% of outputs reaching 3 points. First-video task following is led by Sora 2, while combined image-and-audio reference remains difficult: Seedance 2.0 scores 2.29 versus Kling 3 Omni’s 2.11.Source: §2.5.1 · pp. 20–21, Table 26 · p. 21
  • Seedance 2.0 leads motion-reference task following at 2.60 and alignment at 2.64, and leads style-reference task following at 2.57 and alignment at 2.37.Kling 3 Omni leads first-frame preservation at 4.31 versus Seedance 2.0’s 2.71; Seedance 2.0 also performs better on combined style-and-subject reference, where competitors often misinterpret the task or produce artifacts.Source: Table 27 · p. 21, §2.5.1 · p. 21
  • Video editing is the closest R2V competition: Kling O1 leads task following at 2.29 versus Seedance 2.0’s 2.20, but Seedance 2.0 leads alignment at 3.79 and editing consistency at 3.75.Seedance 2.0 handles long-text and multi-edit instructions more completely, though all models share failures involving unresponsive edits and unintended changes outside edited regions.Source: §2.5.1 · p. 21, Table 28 · p. 22
  • Video continuation is supported only by Seedance 2.0, which scores 2.88 on task following and 3.18 on reference alignment.It handles complex narratives and long-text continuation prompts, but can have color-consistency problems, omit multiple subjects, or duplicate subjects.Source: §2.5.1 · p. 21, Table 28 · p. 22
  • Video extension is Seedance 2.0’s weakest R2V task: Veo 3.1 scores 2.78 on task following versus 1.93, and 3.44 on reference alignment versus 3.28.Seedance 2.0 accepts arbitrary uploaded videos and can combine extension with subject-image input, while Veo 3.1 extends only videos it generated itself.Source: §2.5.1 · p. 22, Table 28 · p. 22

2.6 Visualization Results

Seedance 2.0’s visualizations depict physically coherent, complex motion and adaptable generation across varied content scenarios. The paper also claims AI generation can reduce professional production costs and cycles.

Visualization Results

  • Seedance 2.0 depicts complex motion that adheres to real-world physical laws, including synchronized skating take-offs, rotations, and landings.Figure 4 shows text-to-video and image-to-video generation examples.Source: §2.6 · p. 22, Figure 4 · p. 22
  • Across advertising, television visual effects, video games, and explainer videos, Seedance 2.0 demonstrates scenario adaptability and high-quality generation.The passage also characterizes its multimodal task performance as robust.Source: §2.6 · p. 22
  • AI-driven generation can replace complex visual-effects pipelines and live-action workflows, reducing professional audio-video production costs and cycles.The paper says this can help creators and enterprises realize their creative visions.Source: §2.6 · p. 22

3 Contributions and Acknowledgments … Jiashi Li

The authors of Seedance 2.0 are listed alphabetically by last name.

3 Contributions and Acknowledgments

  • All authors are listed in alphabetical order by last name.Source: §3 · p. 25

De Chen … Xingxing Li

The supplied passage under “Ming Li” is a list of names, from Liyang Chen through Weilin Huang.

Ming Li

  • The “Ming Li” subsection lists names from Liyang Chen through Weilin Huang.Source: Ming Li · p. 25

Yameng Li … Yueyan Li

The supplied Yueyan Li subsection consists of a roster of names, presented across six consecutive passages.

Yueyan Li

  • The list runs from Chao Liang through Wenjing Tang.It includes Chao Han, Jianzhong Liang, Ying Liang, Jiawei Liu, Jingting Ma, and Renfei Sun, among others.Source: Yueyan Li · p. 25
  • The next passage lists names from Zhaoyang Huang through Gen Li.It includes Zhongyi Huang, Ashley Kim, and Shanshan Lao.Source: Yueyan Li · p. 25
  • The list continues from Dongliang Wang through Zhonghua Zhai.It includes Rui Wang, Meng Wei, Di Wu, Shuang Xu, and Jiaqi Yang, among many others.Source: Yueyan Li · p. 25
  • The next passage lists Bowen Zhang, Chenlin Zhang, Heng Zhang, Jun Zhang, Manlin Zhang, Peiyuan Zhang, Shuo Zhang, and Xiaohe Zhang.Source: Yueyan Li · p. 26
  • The list continues from Xiaoying Zhang through Liming Zhao.It includes Xinyan Zhang, Xinyi Zhang, Yichi Zhang, Zixiang Zhang, Haiyu Zhao, and Huating Zhao.Source: Yueyan Li · p. 26
  • The final passage lists names from Yian Zhao through Feilong Zuo.It includes Guangcong Zheng, Jianbin Zheng, Xiaozheng Zheng, Zerong Zheng, and Kuan Zhu.Source: Yueyan Li · p. 26

Related papers

Loading 2604.14148v1…