Appendix C Annotation Guidelines
Annotation interface. Our annotation workflow combines a question-based labeling interface with a lightweight temporal segmentation tool. Figure 5 shows example questions in our annotation framework. Annotators watch the full video, answer event-level questions, and revise previous selections whenever later observations reveal a better interpretation. In practice, this interface is used after temporal segmentation to annotate whether each event contains camera motion and semantic attributes, while the free-form camera motion description field is used to record finer details that are not explicitly enumerated in our discrete labels. This design helps reduce contradictory labels and keeps the annotation process consistent with the taxonomy in Section B .
Figure 5 : Example questions in our annotation framework. Temporal boundary annotation based on LosslessCut. For multi-event videos, we first annotate temporal boundaries with LosslessCut, as illustrated in Figure 6 . The workflow is: (1) import the original video into LosslessCut; (2) play the video and observe activity changes over time; (3) based on the event segmentation rules in Section 2.2 , place a cut marker whenever a new event begins; and (4) after all boundaries are marked, use the export function to automatically obtain all sub-videos together with their begin and end timestamps. After segmentation, annotators further label each event with camera motion and semantic attributes when applicable. This decomposition is practical because it separates temporal grounding from semantic labeling, making the overall process easier to review, correct, and scale.
Figure 6 : Annotating multi-event temporal boundaries based on LosslessCut. Human consensus and dataset triage. We employ three annotators to inspect videos from the curated pool and decide whether each sample should be kept as a multi-event video or filtered out as single-event or difficult. In practice, different annotators naturally have different sensitivity to small action changes, so disagreements often arise not because the video is ambiguous everywhere, but because one annotator prefers finer decomposition than another. To make the decision reproducible, we only keep videos whose event structure remains stable across annotators. For temporal boundaries, we treat two timestamps t 1 t_{1} t 1 t 2 t_{2} t 2 ∣ t 1 − t 2 ∣ ≤ δ |t_{1}-t_{2}|\leq\delta ∣ t 1 − t 2 ∣ ≤ δ δ = 0.5 \delta=0.5 δ = 0.5 This consensus rule is especially important because MEV favors action-driven atomicity rather than over-segmentation. When one annotator proposes more events but the additional cuts are not reproducible across the other annotators, we do not keep the finer split. Instead, we choose the more stable segmentation that can be consistently named and grounded by multiple people. In addition, a video is still treated as single-event when one coherent activity is accompanied by one consistent camera motion throughout the clip. By contrast, if the camera behavior itself changes in a semantically meaningful way and forms a new temporally grounded stage—for example, a static landscape shot followed by a pan-right reveal—we treat this transition as a new event.
Most candidate videos are eventually filtered out at this stage. Many clips are too short, contain only one coherent event, or lack consistently nameable stages and reproducible boundaries. Continuous sports or performance motions are assigned to the difficult category only when they fail these event-segmentation criteria. As a result, only a relatively small portion of the large curated pool is retained for detailed event-level annotation. Following the annotation split in Section 2.2 , the manual inspection stage covers 52.9% of the curated pool, and only samples retained as reliable multi-event videos receive full event-level timestamps, captions, camera descriptions, and semantic attributes.
Agreement statistics. Table 11 reports agreement at two levels, both aligned with the δ = 0.5 \delta=0.5 δ = 0.5 Statistic Ratio Denominator and meaning Unanimous single-event decision 51.2% Manually inspected candidates; all annotators judge no stable multi-event structure Difficult / ambiguous decision 27.7% Manually inspected candidates; event decomposition is unreliable or hard to name consistently Retained multi-event decision 21.1% Manually inspected candidates; clip passes reproducible event-structure triage Boundary reconciliation 97.7% Retained multi-event candidates; timestamps require consensus alignment before finalization Table 11 : Human triage and boundary agreement statistics.
Appendix D Data Sources, Release, Licensing, and Responsible Use
Source selection and scope. MEV is constructed from public video sources, primarily MixKit and a complementary set of YouTube-hosted videos. MixKit provides visually clean stock-style videos with stable imaging quality, while YouTube complements it with long-tail activity patterns that are less frequent in stock footage. For YouTube candidates, we use predefined activity-oriented categories rather than unrestricted crawling. These categories focus on visually clear long-take activities such as daily activities, cooking, fitness or sports, tool use or DIY, pet or animal interactions, and outdoor activities. We then apply the filtering and event-triage pipeline described in Sections 2.1 and 2.2 to retain only clips that satisfy MEV’s long-take multi-event requirements.
Dataset availability. The MEV dataset is publicly available at https://huggingface.co/datasets/MultiEventVideo/MEV , including the official release package and accompanying metadata.
Sensitive-content filtering and bias. MEV is intended for research on long-take multi-event generation. During collection and filtering, we exclude videos centered on sensitive or private-context content, including real-world violence, accidents, injuries, medical or hospital scenes, political or religious events, adult or explicit sexual content, surveillance or private security footage, hate or extremist content, news, and other highly sensitive public events. We also avoid annotating identity or sensitive personal attributes. These choices make MEV biased toward visually clear, action-rich, non-sensitive clips, which is appropriate for the target generation setting but should be considered when interpreting dataset coverage.
Licensing. MEV is released as an open research dataset. Author-created MEV annotations, labels, timestamps, prompts, and evaluation metadata are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0), unless a file explicitly states otherwise. Video files included in the official public release are provided together with MEV annotations as a directly usable research data package for evaluation and model training, subject to the applicable source license or platform terms. MEV does not impose additional restrictions beyond those source terms and responsible-use requirements, and users do not need to re-crawl source websites for videos included in the official package. Only a small subset of optional 4K source versions may be premium, paid, or otherwise restricted; these optional variants are not part of the public package, and MEV provides annotations, timestamps, prompts, and other metadata where permitted. Full release details are provided in the repository license notice.
Takedown policy. Users should comply with the applicable source terms. If a rights holder requests removal of a video file, video reference, annotation, or related metadata, or if legal, ethical, privacy, or security concerns arise, we may remove or modify the affected materials. Affected videos are removed from subsequent public packages when appropriate.
Appendix E Experiments on Annotation Model
VLM Details. To improve automatic event segmentation, we fine-tune an event-aware Qwen3-VL-32B model on SFT data constructed from the manually labeled videos described in Section 2.2 . In particular, the temporally grounded multi-event portion contains about 5K videos with human-annotated event boundaries, which serve as the core supervision for boundary prediction. We serialize the supervision into a small set of instruction-following video-text templates that mirror our annotation workflow. As summarized in Table 12 , the SFT data covers three practical forms: a status-only template for identifying videos that should be filtered out as single-event or difficult , a timestamp template for retained multi-event videos, and an explanation template for borderline cases where a short textual reason is useful for later human review. This design keeps the output easy to parse while remaining close to the way human annotators reason about event segmentation. For Qwen3-VL-32B [ 36 ] , we adopt an LLM-only supervised fine-tuning setup: the language model is trainable, while the vision tower and the multimodal projector remain frozen. We sample videos at 8 FPS and train with bf16, cosine learning-rate decay, a warmup ratio of 0.05, FlashAttention-2, and DeepSpeed ZeRO-3. Table 13 summarizes the configuration used for our 32B model.
Format Input Target Output Status-only Video + segmentation instruction single-event or difficult Timestamped multi-event Video + segmentation instruction multi-event with an ordered list of event begin/end timestamps Explanation-assisted Video + segmentation instruction Filtering decision or merge decision with a short textual explanation for later review Table 12 : SFT data formats used for event segmentation.
Hyperparameter Value tuning_mode LLM only per_device_train_batch_size 1 gradient_accumulation_steps 2 learning_rate 2.0e-6 num_train_epochs 6.0 lr_scheduler_type cosine warmup_ratio 0.05 tune_mm_llm true tune_mm_mlp false tune_mm_vision false video_fps 8.0 video_max_pixels 16384 image_max_pixels 262144 deepspeed zero3_config.json template qwen3_vl bf16 true flash_attn fa2 Table 13 : SFT hyperparameters for Qwen3-VL-32B.
Baselines. We compare our SFT model with three strong VLM baselines: Qwen3-VL-8B [ 36 ] , Qwen3-VL-32B [ 36 ] , and Gemini 2.5 Pro [ 37 ] . Qwen3-VL-8B and Qwen3-VL-32B represent open-source models with different parameter scales, while Gemini 2.5 Pro serves as a strong proprietary baseline.
Evaluation Metrics. Our task is a joint event counting and boundary prediction problem: given a video, the model predicts both the number of events and the temporal boundaries between adjacent events. We therefore report three complementary groups of metrics. First, for boundary detection , we compare the predicted internal boundary set B P = { e ^ 1 , … , e ^ N − 1 } B_{P}=\{\hat{e}_{1},\dots,\hat{e}_{N-1}\} B P = { e ^ 1 , … , e ^ N − 1 } B G = { e 1 , … , e M − 1 } B_{G}=\{e_{1},\dots,e_{M-1}\} B G = { e 1 , … , e M − 1 } δ \delta δ ∣ e ^ − e ∣ ≤ δ |\hat{e}-e|\leq\delta ∣ e ^ − e ∣ ≤ δ δ = 0.5 \delta=0.5 δ = 0.5 1.0 1.0 1.0 Second, for segment-level overlap , we compare the predicted segments P = { p j } j = 1 N P=\{p_{j}\}_{j=1}^{N} P = { p j } j = 1 N G = { g i = [ s i , e i ] } i = 1 M G=\{g_{i}=[s_{i},e_{i}]\}_{i=1}^{M} G = { g i = [ s i , e i ] } i = 1 M [ s p , e p ] [s_{p},e_{p}] [ s p , e p ] [ s g , e g ] [s_{g},e_{g}] [ s g , e g ] t I o U = ∣ [ s p , e p ] ∩ [ s g , e g ] ∣ ∣ [ s p , e p ] ∪ [ s g , e g ] ∣ . \mathrm{tIoU}=\frac{|[s_{p},e_{p}]\cap[s_{g},e_{g}]|}{|[s_{p},e_{p}]\cup[s_{g},e_{g}]|}. tIoU = ∣ [ s p , e p ] ∪ [ s g , e g ] ∣ ∣ [ s p , e p ] ∩ [ s g , e g ] ∣ . = 0.7 =0.7 = 0.7 Third, because the model also predicts the event count, we separately report counting error . Specifically, we use the absolute count error ∣ N − M ∣ |N-M| ∣ N − M ∣ ∣ N − M ∣ / M |N-M|/M ∣ N − M ∣/ M Experimental Results. We evaluate event counting and boundary prediction on videos from the remaining candidate pool, using human-verified annotations as references. Tables 14 and 15 show that the SFT model consistently improves event segmentation quality over all baselines. It achieves the lowest counting error and the best boundary and segment-level performance, indicating that supervised adaptation not only reduces over- or under-segmentation, but also produces a temporal partition that is closer to the full ground-truth event structure. Among the non-SFT baselines, Gemini 2.5 Pro and Qwen3-VL-32B are broadly comparable. Gemini shows slightly stronger recall-oriented behavior, leading to better boundary recall and slightly higher F1 under the looser matching setting, whereas Qwen3-VL-32B is slightly stronger in precision-oriented boundary detection and segment overlap quality. Qwen3-VL-8B is clearly weaker than the larger models across both counting and temporal localization metrics, confirming that model scale remains important for reliable long-video event segmentation.
Model Mean ∣ N − M ∣ |N-M| ∣ N − M ∣ Table 14 : Count error results for event segmentation. Lower is better. Model [email protected] [email protected] [email protected] B-F1@1s [email protected] LW-mtIoU Qwen3-VL-8B 40.2 37.4 38.7 50.1 35.4 0.46 Qwen3-VL-32B 50.4 45.6 47.9 60.8 46.5 0.58 Gemini 2.5 Pro 48.9 48.3 48.6 61.7 45.9 0.57 Qwen3-VL-32B (SFT) 56.8 54.9 55.8 68.1 53.4 0.64 Table 15 : Boundary and segment-level evaluation for event segmentation. Higher is better. Qualitative comparison of temporal boundary annotation. Figure 7 provides a qualitative comparison of human annotation, Qwen3-VL-32B (SFT), Qwen3-VL-32B, and Gemini 2.5 Pro on representative temporal segmentation cases. The SFT model is effective in recovering the coarse event structure and event count in many examples, matching the human annotation more closely than the non-SFT baselines. By contrast, native Qwen3-VL often fails to identify camera-behavior changes in long takes as valid event boundaries, and tends to aggregate different actions and camera-motion cues into a single broad event. However, the SFT model’s timestamp precision is still imperfect. In the landscape example, the transition from a cloud-focused aerial view to the later camera movement that reveals the mountain range and foreground ridge is visually salient, but the SFT boundaries still deviate from the human boundaries. This indicates that supervised VLM segmentation is useful for scaling annotation, while human verification remains necessary for precise temporal grounding.
The examples also illustrate why human annotation can differ from VLM outputs in event count and granularity. MEV follows the event definition in Section 2.2 : boundaries are not merely low-level motion changes, but correspond to stable changes in action state, scene state, interaction target, subject visibility, or camera behavior that can be named and grounded reproducibly. In the first example, the human boundaries separate a hand-centered static close-up, a camera tilt that moves the visual focus upward, a face-centered affectionate gaze, and a final phase in which both stop smiling and look ahead. The change between static and moving camera behavior therefore serves as a meaningful event cue rather than incidental visual motion, reducing ambiguity between physical interaction, camera transition, and social exchange. In the second example, human annotators isolate the window in which the boy runs around the wheelchair and disappears from the field of view, forming a complete event centered on subject disappearance. The non-SFT VLM captions describe the sustained family interaction but miss this visibility change, whereas the human-defined window makes the cue explicit and encourages more focused event captions. These atomic, modular event windows reduce boundary ambiguity and provide cleaner conditioning units for downstream generation, where explicit event structure helps the model follow localized actions instead of a coarse entangled description.
Figure 7 : Qualitative comparison of temporal boundary annotation. Human annotations provide more atomic and temporally focused event windows than non-SFT VLM outputs. Qwen3-VL-32B (SFT) improves event-count recovery, but human verification remains important for boundary timing, camera-transition boundaries, and subject disappearance from view. Appendix F Additional Analysis of MEV Dataset
Additional dataset statistics. Figure 9 provides four complementary views of the linguistic and temporal structure of MEV. Figure 9 (a) shows the top bigrams in event captions. The most frequent phrases are strongly human-centric and appearance-centric, such as man woman , her head , young man , her hands , and several clothing- or body-related phrases. This indicates that the event captions in MEV are not only action descriptions, but are also consistently grounded in visible subjects, body parts, and local visual details. Figure 9 (b) summarizes the most common action lemmas. High-frequency verbs such as look , wear , hold , sit , smile , and stand suggest that MEV covers a broad spectrum of event types, ranging from explicit physical actions to posture changes, gaze, and interaction-heavy moments that are common in long-take videos.
Figure 9 (c) shows video-level camera-motion coverage: 51.1% of videos contain at least one moving-camera event, and 59.7% contain at least one static event. These categories overlap because a video can contain both types of events. At the event level, moving-camera and static events account for 42.9% and 57.1%, respectively. The two categories are relatively balanced. Figure 9 (d) plots event duration for indices 1–17. Within this range, early events exhibit the largest spread and longest outliers, while later events are typically shorter and less variable. Intuitively, early events in long-take videos often serve as setup or scene establishment, whereas later events are more likely to be localized transitions or follow-up actions. The sparser points at later event indices also reflect the long-tailed distribution of event counts per video, with relatively few videos containing many events.
Quality-score granularity. Across 8,016 videos and 25,262 events, we measure the share of total event-score variation that occurs among events within the same video. For each quality metric, we compute R w i t h i n = ∑ v ∑ e ( x v e − x ˉ v ) 2 ∑ v ∑ e ( x v e − x ˉ ) 2 , R_{\mathrm{within}}=\frac{\sum_{v}\sum_{e}(x_{ve}-\bar{x}_{v})^{2}}{\sum_{v}\sum_{e}(x_{ve}-\bar{x})^{2}}, R within = ∑ v ∑ e ( x v e − x ˉ ) 2 ∑ v ∑ e ( x v e − x ˉ v ) 2 , x v e x_{ve} x v e e e e v v v x ˉ v \bar{x}_{v} x ˉ v x ˉ \bar{x} x ˉ 100 R w i t h i n 100R_{\mathrm{within}} 100 R within Metric Within-video share (%) Subject consistency 41.76 Background consistency 42.92 Temporal flickering 15.13 Motion smoothness 18.46 Imaging quality 11.39 Aesthetic quality 10.38 Table 16 : Within-video variation in event-level quality scores. The percentage of total event-score variation occurring within videos is reported for each metric.
Subject and background consistency show substantial variation within videos, supporting event-level scoring for localized analysis. Imaging and aesthetic quality vary mainly between videos, supporting their use for global screening, while their event-level scores can still identify localized degradation.
Multi-granular caption quality analysis. Figure 8 illustrates the impact of global captions at different granularities (short/middle/long) on video generation. As the granularity becomes finer, the generative model can leverage more information. For example, the Middle Caption (row 2) correctly captures the person’s orientation and preserves finer details ( e.g. , earrings and rings); the Long Caption (row 3) further specifies attributes such as the person’s skin tone. Kling (row 4) achieves the best results, exhibiting a more faithful profile and more accurate event execution; however, it still omits certain events ( e.g. , “turning off the lamp”). These observations suggest a potential limitation of current video generators: when handling multi-event video generation, some event-level information can be difficult to utilize effectively. Video–text retrieval is an important application of video–language alignment [ 61 , 62 , 27 , 63 ] . For retrieval, we fine-tune CLIP4Clip [ 61 ] , initialized from pretrained CLIP ViT-B/16, on MEV following its ActivityNet Captions [ 47 ] training configuration. We evaluate video–text retrieval using whole videos, with the same 80/20 split as the generation experiments (Appendix G ). For each setting, we concatenate the specified text fields from the same whole video into a single text input. Table 17 shows that enriching the short-caption baseline with global or event-level signals, including event captions, event camera descriptions, and global camera descriptions, consistently improves bidirectional retrieval performance.
Figure 8 : Comparison of video generation performance across different caption granularities and models. We evaluate the open-source Wan 2.1-14B [ 23 ] and the closed-source Kling 2.6 [ 64 ] by concatenating global captions with event-level captions in a single-pass generation manner. Results show that higher caption granularity improves alignment with the Ground Truth video and enhances coverage of sequential events. Wan 2.1 consistently fails to capture some events ( e.g. , turning off the lamp and lying down), and even the closed-source Kling 2.6 struggles to execute “turning off the lamp”, indicating potential limitations for multi-event video generation. Setting T2V V2T Base (Short) 81.8 84.1 + Event camera desc 84.8 85.2 + Event captions 87.0 89.5 + Global camera 83.8 87.4 + Global camera + Event captions 85.9 86.9 Table 17 : Bidirectional video–text retrieval across multi-granular captions (R@1).
Annotation model evaluation. We evaluate the original predictions of three VLM annotators against human annotations on event clips from 128 randomly sampled source videos (Table 18 ). Camera motion is measured by exact-set accuracy: a prediction is correct only when its complete set of camera-motion labels matches the human annotation. Missing, additional, or directionally incorrect labels make the prediction incorrect. Semantic attributes are evaluated using binary accuracy. The main takeaway is that camera understanding remains the weakest component: although the models often recognize that the viewpoint changes, they still struggle to assign the correct fine-grained motion label. By contrast, most semantic attributes are substantially more reliable overall. Multi-person, animal, and cooking cues are comparatively easy, while lighting-related judgments remain noticeably less stable. This gap between camera precision and semantic robustness is consistent with the qualitative results below.
Model Camera Appear. Disappear. Multi- person Env. Change LLaVA-Video-7B 6.2 81.1 80.1 86.5 58.6 Gemini 2.5 Pro 22.0 91.5 89.8 95.5 76.6 Qwen3-VL-32B 10.0 91.3 89.6 96.7 72.8 Model Lighting Animal Cooking Avg. LLaVA-Video-7B 55.6 89.4 90.5 68.5 Gemini 2.5 Pro 69.5 96.7 95.3 79.6 Qwen3-VL-32B 66.4 95.3 94.6 77.1 Table 18 : Accuracy (%) of original VLM predictions against human annotations on event clips from 128 randomly sampled source videos. Camera motion uses exact-set accuracy; semantic attributes use binary accuracy. Avg. is the unweighted mean of the eight displayed scores. Higher is better.
We additionally evaluate camera-specialized models from CameraBench [ 38 ] using its released Qwen2.5-VL-7B and Qwen2.5-VL-72B checkpoints on 330 available event clips from the same 128 human-annotated source videos. We map the reference annotations and model predictions to CameraBench’s 15 camera-motion outputs. Exact-set accuracy and F1 use a fixed Yes/No threshold of 0.5, without tuning on this subset. We compute mAP from continuous P ( Y e s ) P(\mathrm{Yes}) P ( Yes ) Model Exact-set Micro-F1 Macro-F1 mAP CameraBench-7B 10.91 35.72 22.12 27.88 CameraBench-72B 8.18 34.70 24.53 24.13 Table 19 : Evaluation of CameraBench models on MEV (%). Results use 330 event clips from 128 source videos after mapping predictions and reference annotations to CameraBench’s 15 camera-motion outputs. Higher is better.
Exact-set accuracy requires all motion labels in an event to be correct, whereas per-primitive AP evaluates each label independently. Among the 7B checkpoint’s incorrect predictions, 46.3% contain only omissions, with no additional incorrect labels. For 72B, the largest error category (46.2%) combines omitted and additional incorrect labels. These patterns illustrate the difficulty of recovering complete label sets for composite camera motions.
MEV also annotates focus shift, which is not included in CameraBench’s output taxonomy. For events with only a focus shift, the reference in the shared motion-label space is static; the focus change itself is not evaluated by these models. Overall, the specialized models still make substantial errors on MEV event clips, supporting the need for human verification in our annotation pipeline.
Human verification and correction. Human annotators reviewed the camera labels and semantic attributes of all 25,262 events and corrected errors before release. Across the dataset, 83.2% of Gemini-generated camera-label sets required at least one human correction. This rate counts events whose camera-label sets were changed, rather than individual label edits. It is computed over the full dataset, whereas the accuracy results in Table 18 use the 128-video evaluation subset.
Qualitative analysis and failure taxonomy of camera annotations. Figure 10 illustrates a representative failure mode of current VLMs on camera reasoning. In this example, all models capture that the framing changes as the girl lifts her head, but they fail to recover the correct motion decomposition. The ground-truth motion is a combination of truck-left and pedestal-up , whereas the models instead reduce the motion to tilt-up , or further confuse translation with rotation by predicting pan-left . This suggests that the challenge is not merely to notice motion, but to factorize it correctly into atomic primitives. Distinguishing physical camera translation from apparent framing rotation remains especially difficult.
Figure 11 further summarizes typical camera-label errors made by strong VLMs. Qwen3-VL is often ambiguous between rotation and translation, such as confusing pan with truck , or reducing vertical camera translation ( pedestal ) to viewpoint rotation ( tilt ). Gemini 2.5 Pro is less prone to some pan / truck confusions and often provides more detailed 3D motion rationales, but it can still assign the wrong motion axis or direction, such as predicting dolly-out for cases whose visual change is better explained by pedestal-down or pan-left . Overall, these examples suggest that existing VLMs can recognize that the view changes, but their 3D scene understanding is still insufficient for reliably decomposing camera motion into translation, rotation, depth motion, and direction. This motivates our manual verification of camera annotations and the release of both categorical camera labels and natural-language camera descriptions.
Qualitative analysis of semantic attributes. Figure 12 shows that semantic attributes are easier than camera labels, but still not uniformly solved. The focus-shift predictions shown here are binary ( true / false ) judgments of whether a focus shift occurs. All three models correctly recognize the salient multi-person interaction, which is visually strong and semantically explicit. However, the lighting change is missed by all of them, indicating that illumination-related reasoning is still weak even when the scene content itself is understood. In addition, LLaVA hallucinates an environmental change, while Qwen3-VL-32B hallucinates a focus shift. These errors suggest that subtle visual changes are often assigned to the wrong underlying factor, especially when lighting, scene state, and photographic effects may be confounded.
Qualitative analysis of global structured captions. Figure 13 compares Gemini 2.5 Pro and Qwen3-VL-32B under the same prompting format for global structured captioning and downstream video question answering. A clear pattern is that the two models emphasize different aspects of the same scene. Gemini is more concise and better at producing a compact description that covers most of the useful global information with little redundancy. Qwen, in contrast, tends to be richer in background details and object-level specification. This difference becomes visible in the QA examples: both models answer coarse scene-level questions well, but Qwen is better at detail-sensitive questions such as the centerpiece composition or the clothing and activity of the background customer. At the same time, both models fail on very fine-grained details such as the older woman’s eye color, which shows that even strong structured captions do not imply complete access to subtle visual attributes. Although Qwen captures more details relevant to the questions in this example, we select Gemini for MEV annotation based on its action-focused event descriptions and stronger overall performance in camera-motion and semantic-attribute annotation (Table 18 ).
These analyses suggest a consistent hierarchy of difficulty on MEV. Coarse global understanding and object- or interaction-centric semantic attributes are already fairly accessible to modern VLMs. Fine-grained scene details can also be recovered, but model behavior becomes style-dependent: Gemini tends to be compact and efficient, while Qwen often provides denser background grounding. In contrast, exact camera-motion factorization and subtle lighting-related reasoning remain notably challenging. Our additional manual verification confirms that these weak points are not merely anecdotal captioning errors, but recurring limitations of current VLMs on structured long-take video analysis. In this sense, MEV helps reveal where existing models remain unreliable, while its human-verified annotations also make it easier to connect these observations to downstream tasks such as video captioning, event understanding, and long-take video generation.
Figure 9 : Additional statistics and analysis of MEV Dataset. Figure 10 : Qualitative comparison of camera motion understanding for different VLMs. Figure 11 : Representative camera annotation failures. Qwen3-VL often confuses camera rotation with physical translation, such as pan versus truck or tilt versus pedestal . Gemini 2.5 Pro gives richer 3D motion rationales, but can still misidentify the correct axis or direction, indicating that fine-grained camera annotation remains challenging for current VLMs. Figure 12 : Qualitative comparison of semantic attribute prediction for different VLMs. Figure 13 : Global structured caption evaluation via video question-answering tasks. Color markers show answer coverage. Gemini produces a more compact description, while Qwen captures additional background details. Both omit some fine-grained attributes. Appendix G Implementation Details of MEV-Wan
Overview. MEV-Wan is built on top of Wan2.1-T2V-1.3B and follows the MinT-style design for multi-event video generation. The key idea is to preserve Wan’s pretrained single-video generation ability while adding explicit event-aware temporal conditioning. Concretely, we keep the original global text conditioning pathway for whole-video semantics and introduce an additional temporal conditioning pathway that aligns event captions with different time regions. The resulting model is trained with the same flow-matching objective as Wan, so the main change lies in the conditioning design rather than the denoising loss itself.
Relationship among compared methods. All four methods belong to the Wan2.1-1.3B lineage. Wan and LongLive are evaluated using their released checkpoints without additional fine-tuning by us. Ultra-Wan and MEV-Wan use the same Wan backbone and MinT-style multi-event adaptation, with different training resources. Table 20 summarizes these distinctions.
Method Fine-tuned by us? Fine-tuning data used in this work Multi-event mechanism Wan2.1-1.3B No – Original Wan conditioning LongLive No – Official autoregressive checkpoint with prompt switching Ultra-Wan Yes UltraVideo-Long subset MinT-style adaptation; inferred events and uniform boundaries MEV-Wan Yes MEV MinT-style adaptation; verified events and grounded boundaries Table 20 : Relationship among the four compared methods. “Fine-tuned by us” refers to additional training performed in this work. A dash indicates that no additional fine-tuning data are used by us.
Ultra-Wan data construction. We use temporal discourse connectives in UltraVideo-Long’s Summarized Description field, such as “then,” “after,” and “before,” to infer the number of events N N N N N N Dual-path conditioning architecture. Each Wan diffusion transformer block retains the original global text cross-attention branch and augments it with a temporal cross-attention branch. The global branch receives the video-level middle caption and is responsible for preserving scene layout, appearance consistency, and overall style. The temporal branch receives the ordered event captions of the sampled event group and injects local event semantics into temporally aligned frame regions. Following MinT, the temporal branch is merged back through a bounded residual gate with Tanh activation, and the gate is initialized near zero. This near-zero initialization is important in practice because it prevents the newly inserted branch from disrupting Wan’s pretrained denoising trajectory too aggressively at the beginning of training.
For temporal binding, we adopt ReRoPE as in MinT. Instead of hard-assigning each event to a fully isolated frame segment, ReRoPE introduces a soft temporal bias based on event identities, normalized timestamps, event boundaries, and validity masks. This design is better suited to MEV because neighboring events in long-take videos often transition smoothly and may visually overlap near event boundaries.
Dataset split. We partition the 8,016 original videos into training and test sets before constructing event groups. All event segments and contiguous groups derived from a given original video inherit its split assignment. Table 21 reports the numbers of original videos and annotated events, together with their total annotated event duration. The 80/20 ratio applies to original videos; event counts and durations need not follow exactly the same ratio.
Split Original videos Events Annotated event duration Train 6,413 20,241 29:07:24.974 Test 1,603 5,021 07:48:42.790 Total 8,016 25,262 36:56:07.764 Table 21 : MEV training and test split statistics. Duration is summed over annotated event intervals and is reported as HH:MM:SS.mmm. The split is performed at the original-video level before event-group construction.
Cross-split overlap audit. We extract DINOv2 features [ 65 ] from frames sampled at 1 fps to identify potential visual overlap between the training and test sets. A video pair is flagged when at least 80% of the sampled frames in the shorter video have temporally ordered matches in the longer video with cosine similarity ≥ 0.95 \geq 0.95 ≥ 0.95 Training sample construction. Suppose a video contains ordered events { e i } i = 1 N \{e_{i}\}_{i=1}^{N} { e i } i = 1 N After selecting an event group, we allocate training frames to events proportionally to their durations. If the selected durations are { d i } \{d_{i}\} { d i } T T T i i i T ⋅ d i / ∑ j d j T\cdot d_{i}/\sum_{j}d_{j} T ⋅ d i / ∑ j d j T T T Optimization and training procedure. MEV-Wan is optimized with the standard Wan flow-matching supervised fine-tuning objective. In each iteration, we sample a video clip, select a contiguous event group, construct the duration-aware temporal tensors, encode the clip into Wan latents, sample a flow-matching timestep, and train the denoiser to predict the target velocity under both global and temporal conditions. In this sense, MEV-Wan reuses the stable Wan training pipeline while extending the conditioning space to become explicitly multi-event aware.
Our implementation supports both full fine-tuning and LoRA fine-tuning under a unified code path. Consistent with our default setting, MEV-Wan uses LoRA-based adaptation for efficient multi-event training, while full fine-tuning of the Wan dit module is retained only as an optional implementation path. In the default LoRA setup, we use rank 32 and adapt the standard Wan attention and feed-forward projections, namely q,k,v,o,ffn.0,ffn.2 .
Implementation choices. Several engineering choices are important for stable multi-event adaptation. First, we cap the number of events per training sample at 5 and use explicit padding masks to support variable-length event groups without changing the architecture. Second, we keep gradient checkpointing enabled and use memory-oriented runtime settings to reduce fragmentation and peak memory overhead for long video clips. Finally, the combination of near-zero temporal gate initialization and default LoRA-based adaptation helps reduce catastrophic forgetting of Wan’s original generation ability.
Category Hyperparameter Value Video setup Backbone Wan2.1-T2V-1.3B Video setup Resolution 480 × 832 480\times 832 480 × 832 1 × 10 − 5 1\times 10^{-5} 1 × 1 0 − 5 L L L Table 22 : Default training hyperparameters for MEV-Wan on MEV. We list the parameters that directly determine the model configuration or the multi-event adaptation behavior. Compute resources. The reported experiments are conducted on AMD MI210 GPUs. Across the MEV-Wan training runs, annotation-model training, generation, and evaluation used for the reported results, the total compute budget is approximately 2,800 GPU-hours. Unless otherwise stated, training uses a per-GPU batch size of 1 and an effective total batch size of 128. For generation evaluation, each 128-prompt test set is randomly sampled five times from the MEV test split, and we report the average over the five draws.
Appendix H MEV-Eval Protocol for Multi-Event Video Generation
Protocol overview. MEV-Eval is an event-centric evaluation protocol designed to complement general video-quality metrics such as VBench. Given a generated video and an input event sequence E = { e i } i = 1 N E=\{e_{i}\}_{i=1}^{N} E = { e i } i = 1 N { τ i } i = 1 N \{\tau_{i}\}_{i=1}^{N} { τ i } i = 1 N Event Execution Accuracy (EEA). For each event e i e_{i} e i x i ∈ { 0 , 1 } x_{i}\in\{0,1\} x i ∈ { 0 , 1 } E E A = 1 N ∑ i = 1 N x i . \mathrm{EEA}=\frac{1}{N}\sum_{i=1}^{N}x_{i}. EEA = N 1 ∑ i = 1 N x i . Event Order Accuracy (EOA). The judge identifies the observed order of executed events. We then compute pairwise order accuracy over all event pairs: E O A = 1 ( N 2 ) ∑ i < j 1 [ e i ≺ e j ] , \mathrm{EOA}=\frac{1}{\binom{N}{2}}\sum_{i<j}\mathbf{1}\left[e_{i}\prec e_{j}\right], EOA = ( 2 N ) 1 ∑ i < j 1 [ e i ≺ e j ] , e i ≺ e j e_{i}\prec e_{j} e i ≺ e j e i e_{i} e i e j e_{j} e j Temporal Grounding Accuracy (TGA). The generated video is evaluated against the same duration-normalized temporal assignment used by the input event sequence. For each event region τ i \tau_{i} τ i g i ∈ { 0 , 1 } g_{i}\in\{0,1\} g i ∈ { 0 , 1 } e i e_{i} e i T G A = 1 N ∑ i = 1 N g i . \mathrm{TGA}=\frac{1}{N}\sum_{i=1}^{N}g_{i}. TGA = N 1 ∑ i = 1 N g i . Boundary Transition Coherence (BTC). For each adjacent event boundary, the judge rates the transition on a 1–5 scale based on semantic continuity, subject identity stability, object-state consistency, scene/camera continuity, and the absence of severe drift, stagnation, or object splitting. A score of 1 indicates a broken or unrelated transition, 3 indicates a partially coherent transition with visible inconsistency, and 5 indicates a smooth and logically coherent transition. We normalize the score to [ 0 , 1 ] [0,1] [ 0 , 1 ] B T C = 1 5 ( N − 1 ) ∑ i = 1 N − 1 b i , \mathrm{BTC}=\frac{1}{5(N-1)}\sum_{i=1}^{N-1}b_{i}, BTC = 5 ( N − 1 ) 1 ∑ i = 1 N − 1 b i , b i b_{i} b i e i e_{i} e i e i + 1 e_{i+1} e i + 1 Human validation. We use a fixed random sample of 128 test prompts. Each of the four methods generates one video per prompt, yielding 512 videos. Three annotators independently score these videos using the same EEA, EOA, TGA, and BTC criteria, with method identities hidden. We use these judgments to assess agreement with Gemini and to compare generation methods under human evaluation. For EEA and TGA, the decision unit is an event-level binary label. For EOA, the decision unit is a pairwise event-order label. For BTC, the decision unit is an adjacent-boundary 1–5 coherence score. For EEA, EOA, and TGA, we report agreement between the VLM label and the majority-vote human consensus, together with mean pairwise human inter-rater agreement. For BTC, we report Spearman rank correlation between the VLM score and the mean human score, together with within-one rating agreement among humans.
Metric Decision unit VLM–human validation Human inter-rater validation EEA Event execution label Agreement w/ consensus: 0.82 Pairwise agreement: 0.86 EOA Event-pair order label Agreement w/ consensus: 0.76 Pairwise agreement: 0.81 TGA Event-region match label Agreement w/ consensus: 0.73 Pairwise agreement: 0.78 BTC Boundary coherence score Spearman ρ \rho ρ Table 23 : Human validation of MEV-Eval judging. These statistics validate the VLM-assisted judge and are not additional generation metrics. For EEA, EOA, and TGA, agreement is computed on binary decision units; for BTC, validation compares 1–5 boundary-transition ratings. Using the same videos and human judgments, we also compare Gemini and human scores for each method in Table 24 . Human EEA, EOA, and TGA use the majority decision of the three annotators for each decision unit. For BTC, we average their 1–5 ratings for each boundary and apply the same normalization as above. We compute scores per prompt using the metric definitions above and then average over the 128 prompts. These results use one fixed subset, while Table 2 reports averages over five random draws.
Method Judge EEA ↑ \uparrow ↑ ↑ \uparrow ↑ ↑ \uparrow ↑ ↑ \uparrow ↑ Table 24 : Human and Gemini evaluation on the same generated videos. Four methods are evaluated on 128 fixed test prompts, yielding 512 videos. The human judgments are the same as those used in Table 23 . Best scores for each judge are in bold. Both human and Gemini evaluations rank MEV-Wan first on all four metrics. The relative ordering of LongLive and Ultra-Wan changes for EOA and TGA, showing that judge differences affect some comparisons while preserving MEV-Wan’s advantage on this subset.
Appendix I Additional Qualitative Results for Multi-Event Generation
Challenges under flat text-only conditioning. Figure 14 provides qualitative examples from Wan2.1-1.3B under flat text-only conditioning. The prompts contain clear action information and temporal cues such as “then”, but the generated videos do not consistently organize all instructed actions into a stable temporal progression. These cases suggest that multi-event video generation can be difficult when rich action descriptions are not paired with structured event-level conditioning. MEV addresses this setting by providing event captions and temporal boundaries that make the intended event sequence explicit during training.
Figure 14 : Qualitative examples under flat text-only conditioning. These Wan2.1-1.3B examples suggest that flat text-only conditioning can struggle to organize rich action cues into a coherent multi-event progression without structured event-level supervision. Generation comparison. Figure 15 shows an additional qualitative comparison on a three-event generation example. Compared with the vanilla Wan2.1 baseline, both MEV-Wan and Ultra-Wan exhibit a clearer event structure and stronger instruction-following ability. The generated sequence more faithfully reflects the intended progression from turning to look at the camera, to leaning the body toward the motorcycle, and then to touching the hair while facing the camera. In contrast, the original Wan baseline preserves the overall scene style but is weaker at expressing the staged temporal structure specified by the prompt. Among the compared methods, MEV-Wan provides the best alignment between event order and visual evolution. The pose change, body orientation, and interaction with the motorcycle remain more consistent with the three event descriptions across the full sequence. Ultra-Wan also improves over vanilla Wan, suggesting that explicit multi-event conditioning is helpful even in a relatively short three-event setting, although its cross-event continuity and instruction following are still weaker than MEV-Wan. LongLive still suffers from severe drifting even when the prompt contains only three events. In particular, a bright patch appears on the right side of the last two frames, forming an obvious artifact that is unrelated to the target scene. This failure is accompanied by less stable subject control and weaker temporal consistency. Overall, this example suggests that event-aware conditioning improves both event structure and instruction following, while autoregressive long-video generation remains vulnerable to drift accumulation even in comparatively simple cases.
Figure 15 : Additional qualitative comparison on multi-event video generation. Appendix J Broader Use, Limitations, and Future Directions
Beyond the attributes used in the current release, MEV also contains cases that are useful for studying more specialized long-take generation challenges. Figure 16 shows three representative directions. First, reappearance requires a model to preserve identity, spatial context, and object state when a subject temporarily disappears because of occlusion, framing, or scene layout and later returns to the visible region. This is challenging for current video generation models because the subject is no longer continuously observed, yet the later event must remain consistent with earlier context rather than introducing a new identity or duplicating the subject. Second, mirror scenes require reasoning about real and reflected spaces simultaneously. A model must keep body pose, object interaction, directionality, and reflection geometry consistent, while avoiding the common error of treating the mirror image as an independent actor or producing physically inconsistent reflected motion. Third, multi-view cases require stable event understanding across viewpoint changes, foreground occlusion, and perspective-dependent appearance. The model must preserve the same subjects and scene state while the camera or view configuration changes, which is difficult when generation is driven only by coarse text descriptions.
MEV’s structured annotations support broader applications in video generation and understanding. Its event-level camera descriptions and semantic attributes support research on controllable video generation [ 66 , 67 , 68 ] . The paired event captions and temporal boundaries support language-guided temporal grounding [ 69 , 70 ] , while the ordered event sequences and structured annotations support studies of event-to-event relations and causal discovery over visual state changes. Compared with video reasoning datasets such as MECD [ 20 ] , MEV provides a larger-scale video collection with richer generation-oriented conditioning signals and more explicit temporal changes, making it complementary for investigating how event transitions, visual state changes, and controllable conditioning interact in long videos.
Broader impacts. MEV can support positive research progress in controllable long-video generation, event-level video understanding, and transparent dataset auditing by providing temporally grounded supervision rather than only coarse video–text pairs. At the same time, stronger long-video generation models may be misused to create misleading synthetic content or imitate recognizable public scenes, and public-video datasets can inherit source and content biases. We mitigate these risks by filtering sensitive and private-context videos, avoiding sensitive personal-attribute annotation, documenting data sources and licensing constraints, and supporting removal or modification of affected materials when rights, privacy, or safety concerns arise.
Limitations and future directions. Extending the evaluation of MEV fine-tuning to other model scales and video domains remains future work. Although we introduce event-centric criteria for multi-event video generation, such as event execution, event order, temporal grounding, and transition coherence, scalable evaluation still relies heavily on VLM-based judging. This remains imperfect because current VLMs can struggle with fine-grained temporal localization, subtle state changes, camera motion, and reflection or multi-view reasoning, as also indicated by our annotation-model analysis. Future work should therefore improve human validation, judge calibration, and uncertainty estimates for event-level video generation evaluation. In addition, our automatic annotation pipeline still depends on manual verification and correction for fields such as event captions, semantic attributes, and camera descriptions. More accurate annotation models, uncertainty-aware filtering, model-ensemble checks, and task-specific annotators for camera motion or hard attributes could reduce this correction burden while preserving the reliability required for temporally grounded multi-event supervision.
Figure 16 : Broader future use of MEV. MEV contains challenging long-take cases involving reappearance, mirror reasoning, and multi-view consistency, which can support future attribute-guided studies and event-level analysis beyond the current generation experiments.