Seedance 1.0: Exploring the Boundaries of Video Generation Models
1. Paper overview
In one sentence: Seedance combines shared text/image-conditioned video generation with a separate resolution refiner and extensive post-training, yielding strong overall preference results while leaving individual component gains difficult to isolate. identitygoalsarchitectureconditioning-refinerpipelinefeedbackdistillationevaluation-protocol
| At a glance | What to know |
|---|---|
| Research problem | Source description The target is simultaneous prompt adherence, plausible motion, visual fidelity, and affordable generation latency. Chinese and English prompts, optional reference images, and cinematic shot transitions make this broader than single-image synthesis. The experiments judge generated video content rather than interaction with a physical environment. goalsarchitectureevaluation-protocol |
| Core mechanism | Source description The system integrates shot-aware data curation, dense bilingual captions, factorized multimodal attention, and a shared conditional interface. Multi-shot support follows from interleaving visual sequences with shot-specific captions and positional encoding. architectureconditioning-refinerdatacaptioning |
| A key reported result | Image-to-video public preference: Seedance 1.0: 1365, −13/+13; 7,280 appearances. Arena Elo; displayed 95% confidence offsets. Artificial Analysis Arena snapshot reproduced in Figure 9; ranking contextualized as 10 June 2025 by Figure 1, not a current leaderboard check. Veo 3 Preview: 1240, −10/+11; Kling 2.0: 1195, −8/+10. The displayed scores imply leads of 125 and 170 Elo points, respectively. These are preference-rating differences, not success-rate gains. arena-i2voverview-caveat |
| Reading caution | Reader analysis Rising reward curves and selected pretraining/CT/SFT/RLHF frames are diagnostics, not controlled component ablations. The claimed advantage over DPO/PPO/GRPO lacks displayed comparative measurements, and reward improvement alone does not establish independent human preference improvement. reward-diagnosticsfeedback |
Core contributions
- Source description
The system integrates shot-aware data curation, dense bilingual captions, factorized multimodal attention, and a shared conditional interface. Multi-shot support follows from interleaving visual sequences with shot-specific captions and positional encoding. architectureconditioning-refinerdatacaptioning
- Author claim
The authors attribute better quality and efficiency to continued training, curated SFT and model merging, separate base/refiner preference alignment, and multistage distillation. These are integrated system claims; the supplied experiments do not quantify each component's independent contribution. trainingfeedbackdistillationreward-diagnostics
Figure 2. Language and vision interact spatially; visual information then propagates across time. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the two input streams on the left: the VAE produces visual tokens and the LLM produces caption tokens. Shot 1 through Shot N arrange these together before the repeated spatial/temporal block pair. The enlarged spatial block uses MM-RoPE and then separates vision and text MLP paths. The temporal block shows only visual tokens and a vision MLP. Its attention is labelled Spatial-temporal Attention; Section 2.2 specifies spatial windows within frames with a receptive field across time. That label therefore should not be read as evidence for unrestricted attention across every space-time token. The text stream participates only in the spatial blocks. vaearchitectureconditioning-refiner
What it supports. The architecture provides a concrete route for sharing text-to-video and image-to-video generation while accommodating shot-specific descriptions. Temporal processing can connect visual content across frames without repeating the text branch in every block. The paper supports this architectural description, but supplies no isolated performance gain for the factorization.
Where the evidence stops. The diagram is an architectural view, not a complete inference graph: noisy inputs, conditioning masks and the refiner are explained elsewhere. Causal VAE compression does not establish a causally masked or autoregressive DiT.
2. Motivation
2.1 The problem and the proposed response
The target is simultaneous prompt adherence, plausible motion, visual fidelity, and affordable generation latency. Chinese and English prompts, optional reference images, and cinematic shot transitions make this broader than single-image synthesis. The experiments judge generated video content rather than interaction with a physical environment. goalsarchitectureevaluation-protocol
2.2 What this reading follows
A convincing generated video has to satisfy several constraints at once: depict the requested subjects, move them plausibly, preserve their appearance, and deliver useful resolution at acceptable latency. Seedance 1.0 approaches this through an entire production pipeline. Dense captions and prompt rewriting establish a language interface; alternating spatial and temporal transformer blocks generate video; a learned refiner supplies detail; preference learning and distillation shape the final output. These figures explain that information flow and distinguish architectural descriptions, training diagnostics, public preference ratings, and selected cinematic examples. The reviewed artifact is the June 2025 v2 technical report, whose experiments evaluate video generation rather than executed actions. identitygoalsarchitectureconditioning-refinerpipelinefeedbackdistillationevaluation-protocol
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is Components of WAMs. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational video-generation-backbone classification is supported. A shared T2V/I2V generator predicts video, while a separate refiner improves resolution. Neither branch predicts motor actions or implements inverse dynamics. Therefore the world-action-model architecture, prediction paradigm and quadrant remain not applicable; joint video-task training is insufficient for a One Model action-system designation. architectureconditioning-refinerpipeline
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
r= rac{C T H W}{3T'H'W'}= rac{C}{3r_t r_h r_w}5. Method in detail
5.1 Trace two kinds of conditioning through the generator
Begin with a user request and, for image-to-video, a reference frame. Prompt engineering rewrites the request into the dense caption format used in training; a text encoder then supplies language tokens. Visual conditioning follows a different route: the unified input formulation concatenates noisy inputs with clean or zero-padded frames and identifies supplied frames with a binary mask. The spatial blocks integrate the language and visual streams, while temporal blocks connect visual information across frames. MM-RoPE additionally supports ordered shot/caption sequences. A separate conditional refiner enhances the base video's resolution. This chain explains both shared task support and the multiple components involved. Its outputs remain generated pixels: the architecture specifies no motor-action head, inverse-dynamics decoder, or environment feedback controller. architectureconditioning-refinerprompt-engineeringpipeline
Figure 5. The shared generator and resolution refiner have distinct training histories. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the upper box as parameter preparation. The blue base-model path progresses through pretraining, continued training, supervised fine-tuning and feedback alignment. The downward arrow from base pretraining initializes the yellow refiner branch, consistent with Section 2.3. Read the lower box as the deployed data flow: a user prompt is rewritten, encoded, passed to the base generator, and then processed by the refiner. Section 2.3 specifies 480p base generation followed by 720p or 1080p refinement. The feedback-alignment boxes belong to training; the depicted inference chain contains no reward-model search or action-execution loop. Image conditioning is specified in the method text rather than drawn here. pipelineconditioning-refinertrainingfeedbackdistillation
What it supports. Sharing one generator across input tasks still leaves several learned components in the full system. The refiner is trained to condition on low-resolution video and receives its own preference alignment. This explains why strong final output quality cannot automatically be attributed entirely to the base transformer.
Where the evidence stops. Distillation is omitted from this overview. Section 4.4.4 explicitly applies feedback learning to the accelerated refiner, while Section 5.1 describes acceleration. The drawing alone cannot recover their complete scheduling or frozen/trainable parameter choices.
5.2 Separate the training curriculum from evidence of causality
The curriculum changes more than resolution. Pretraining introduces images and videos progressively, while continued training increases the image-to-video fraction from 20% to 40%. The latter stage also selects data with stronger aesthetics and motion and introduces captions that omit static details already visible in the first frame. SFT then changes data again, trains specialized models and merges them. Reader interpretation: motion-only captions may make language concentrate on what should change while the reference image supplies what should persist. That is a plausible explanation of the design, not an isolated experimental finding. Figure 6's selected stage comparisons cannot tell which of task frequency, captions, selection or merging caused a gain. A controlled comparison must vary those choices separately before attributing improvements to any single mechanism. trainingreward-diagnosticsconditioning-refiner
5.3 Keep optimization rewards, preferences and latency separate
The reward curves answer whether training increases the displayed reward-model scores. The Arena table answers which complete system public participants prefer, while the internal GSB plots divide expert judgments into motion, prompt following, visual quality and overall preference. These are distinct measurements. Reader interpretation: a rise in aesthetics reward could coexist with unchanged narrative adherence, and an overall preference advantage can coexist with a weaker component score, as the Kling comparison illustrates. Acceleration introduces another axis: the reported 41.4-second generation time concerns a five-second 1080p output on L20, but the complete device and timing configuration is not given. Evaluate quality at the deployed sampling budget, preserve ties and uncertainty in preference reporting, and resist treating either reward growth or a speedup claim as a complete reproduction result. feedbackreward-diagnosticsarena-i2vevaluation-protocolexpert-resultslatencydistillation
5.4 Training and inference
During training
Data processing segments videos into clips of at most 12 seconds, removes overlays and unsuitable content, deduplicates semantic clusters, and rebalances categories. Tarsier2-based captioning freezes its visual encoder while fully tuning the language model on bilingual annotations. Dataset size and exact filtering thresholds are not supplied. datacaptioning
VAE training combines L1, KL, LPIPS and adversarial losses. DiT uses flow matching with velocity prediction, logit-normal timesteps and a resolution/duration-aware noise shift. Training progresses from 256px images to image/video mixtures, then 640px and finally 24-fps video; early video training uses 12 fps and 3–12-second clips. vaetraining
Continued training raises the image-to-video share from 20% to 40%, selects stronger aesthetic/motion data, and mixes full captions with motion-only captions. SFT trains on curated subsets, uses smaller learning rates and early stopping, and merges specialized models. The merge coefficients and exact training schedules are unspecified. training
Foundational, motion and aesthetic reward models supervise generated videos; aesthetics uses keyframes. Base feedback learning directly predicts clean video x_0 when rewards can assess quality and maximizes composite rewards, alternating generator/reward-model learning across rounds. The accelerated refiner separately maximizes a linear reward combination. Reward weights and detailed update rules are absent. feedback
During inference
The deployed path rewrites the prompt, encodes text, generates base video, refines it if required, and decodes pixels. Reward models are described as training supervision, not an inference-time planner. Conditional video synthesis does not extract or execute physical actions. pipelineconditioning-refinerfeedbackarchitecture
Acceleration combines trajectory-segmented consistency distillation, score distillation and adversarial preference supervision. A thinner VAE decoder is retrained with its encoder fixed. Kernel fusion, quantization, sparse attention, parallelism and offloading further optimize deployment; exact final sampling steps are not disclosed. distillationinfrastructure
5.5 Implementation flow
- Compress and condition
A temporally causal convolutional VAE compresses video with temporal/spatial downsampling (4,16,16) and 48 latent channels; DiT patchification is removed. Noisy inputs are channel-concatenated with clean or zero-padded conditioning frames, and binary masks identify supplied frames. vaeconditioning-refiner
- Alternate spatial and temporal processing
VAE visual tokens and decoder-only-LLM text tokens enter MMDiT-style spatial blocks with separate modality weights. Text interacts with vision only there. Temporal blocks operate on visual tokens using spatial window partitions spanning time. Q and K normalization stabilizes attention. architecture
- Organize shots and refine resolution
MM-RoPE combines visual 3D positions and extra text positions, supporting temporally ordered shots with individual captions. The separate diffusion refiner initializes from the pretrained base; it conditions on upsampled low-resolution video concatenated with diffusion noise to synthesize higher-resolution detail. architectureconditioning-refiner
- Align user language with training captions
A Qwen2.5-14B-based prompt rewriter receives full SFT on prompt/caption pairs and then LoRA DPO on correct/incorrect rewrites. It converts user requests into dense caption format before text encoding. This component is distinct from the video captioner used to prepare training data. prompt-engineeringcaptioning
6. Experiments & results
Seedance 1.0 combines a bilingual video diffusion backbone with dense captioning, a learned resolution refiner, human preference alignment, and inference acceleration. Its spatial blocks mix language and vision; temporal blocks propagate visual information across frames. A common conditioning interface supports text-to-video and image-to-video, while interleaved shot captions support multi-shot sequences. The reported preference results favor the overall system, but do not isolate its components or establish an action-executing world model.
The source provides no controlled quantitative component-ablation table: Figures 6 and 8 are selected stage comparisons, and Figure 7 contains reward-training diagnostics. The diagnostic fills the ablation section with that limitation explicit. Figure 9 contains the original quantitative leaderboard tables. Multi-shot and style claims are illustrated by sampled frames without aggregate success rates; the PDF cannot supply playback evidence. reward-diagnosticsfeedbackarena-i2varena-t2vmulti-shotstyles
6.1 Read the original evidence
Figure 9, right-hand image-to-video leaderboard table. The paper's public-preference snapshot places Seedance first in image-to-video. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First check the selected Image to Video tab; the adjacent original table belongs to a separate text-to-video track. The Arena ELO column summarizes public pairwise preferences. The next column displays confidence-interval offsets around each score, and the last gives appearances rather than a shared fixed prompt count. In the first row, Seedance has 1365 Elo with offsets of −13/+13 and 7,280 appearances. Veo 3 Preview and Kling 2.0 occupy the next rows. Section 7.1 explains the public preference setting, while Figure 1 contextualizes the ranking claim as June 10, 2025. This is the snapshot reproduced in the paper. arena-i2voverview-caveatevaluation-protocol
What it supports. Seedance's displayed Elo exceeds Veo 3 Preview by 125 points and Kling 2.0 by 170 points. The original rows retain the reported uncertainty: Veo 3 Preview is 1240 with −10/+11, and Kling 2.0 is 1195 with −8/+10. These support the paper's historical preference lead in this track.
Where the evidence stops. Elo differences are not percentage-point accuracy improvements. This is not a current leaderboard check or a matched-compute experiment. The table uses Kling 2.0; the internal expert benchmark instead evaluates Kling 2.1 Master.
Figure 11. Overall preference gains coexist with component-level tradeoffs. Original paper, p. 17 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the legend before comparing lengths: green is a Seedance preference, red favors the other model, and gray records Same. Each horizontal bar occupies the full comparison distribution. Compare the green and red portions of a single row; keep ties visible rather than silently treating them as wins or discarding them. The four panels belong to the internal T2V task, for which SeedVideoBench supplies 300 prompts. In the Kling panel, prompt following favors Seedance, while motion and visual quality have longer red portions. The Overall row therefore captures a different judgment from any one component. The absolute-score plots elsewhere use a separate five-point protocol. evaluation-protocolexpert-results
What it supports. Seedance receives more overall favorable than unfavorable judgments against all four displayed competitors. Its prompt-following advantage over Kling is particularly visible, but Kling wins more motion and visual-quality comparisons. This directly limits an interpretation that the leading overall model is best on every evaluated dimension.
Where the evidence stops. The bars have no printed numerical values or confidence intervals, so exact percentages are not reported here. Rater counts, agreement and complete generation settings are missing; these internal preferences should not be combined numerically with Arena Elo.
Figure 14. A selected cinematic prompt probes event ordering and changes of view. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The retained caption identifies rows: Seedance above, Kling in the middle, Veo below. Read frames from left to right within a row. The Chinese prompt requests warm cinematic imagery of a short-haired woman crying outside a bar, a cut to cigarette ends and empty bottles near her feet, then a cut to a suited man approaching and patting her shoulder. Compare which requested views appear and whether the woman, clothing and location remain recognizable. Seedance's central insert visibly changes the framing to the ground-level objects. The final interaction requires particular care: selected stills cannot establish every action or the smoothness of the transition between shots. architecturemulti-shot
What it supports. The montage illustrates the intended multi-shot interface: one prompt specifies an ordered narrative that can include close inserts and returning views. It gives a concrete example of view changes in generated content, complementing the interleaved shot-caption architecture. It supplies no frequency estimate for successful narrative completion.
Where the evidence stops. These are selected frames, not full videos or repeated trials. The comparison does not isolate MM-RoPE, and a missing action between sampled frames cannot establish a failure. Broader continuity claims remain author claims without an aggregate multi-shot metric.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Image-to-video public preference Artificial Analysis Arena snapshot reproduced in Figure 9; ranking contextualized as 10 June 2025 by Figure 1, not a current leaderboard check. | Seedance 1.0: 1365, −13/+13; 7,280 appearances. Arena Elo; displayed 95% confidence offsets | Veo 3 Preview: 1240, −10/+11; Kling 2.0: 1195, −8/+10. The displayed scores imply leads of 125 and 170 Elo points, respectively. These are preference-rating differences, not success-rate gains. arena-i2voverview-caveat |
| Text-to-video public preference Separate text-to-video track of the reproduced Arena snapshot. | Seedance 1.0: 1314, −12/+12; 6,337 appearances. Arena Elo; displayed 95% confidence offsets | Veo 3 Preview: 1253, −9/+9; 8,902 appearances. The displayed lead is 61 Elo points. These results concern a different track and participant pool from the image-to-video table. arena-t2v |
| Expert video-quality comparison Internal SeedVideoBench 1.0: 300 prompts per T2V/I2V task; five-point absolute ratings and pairwise Good-Same-Bad comparisons. | Overall GSB comparisons favor Seedance against every displayed competitor in both tasks; component rankings vary. Absolute ratings and GSB preference distributions | T2V comparisons include Kling 2.1 Master, Veo 3, Wan 2.1 and Sora; I2V substitutes Runway Gen4 for Sora. Kling leads I2V absolute preservation; Veo 3 and Kling exceed Seedance in T2V absolute aesthetics. Overall preference does not imply leadership in every quality dimension. Unlabelled bar heights are not transcribed as exact scores; internal uncertainty and rater counts are not reported. evaluation-protocolexpert-results |
| 1080p video-generation latency Author-reported five-second output on NVIDIA L20; device count, batching and timing boundary are unspecified. | 41.4 seconds; introduction claims over 10× acceleration. Generation latency and claimed end-to-end acceleration | Section 5.1 separately claims 4× DiT acceleration with TSCD and 2× VAE decoding acceleration; no complete matched timing table is supplied. A reported deployment result, not a measured reproduction. Component speedups cannot be multiplied to reconstruct end-to-end latency. latencydistillation |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 7 — training diagnostic. The optimized reward signals rise for both the base model and the refiner. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across a row to compare the three kinds of feedback, then move down to the separate refiner optimization. Blue is the foundational reward, green is motion, and orange is aesthetics. Each panel has its own vertical range, and the two rows span different iteration ranges, so compare trends within panels rather than numerical heights across models. The foundational reward targets alignment and structural stability; the aesthetic reward evaluates video keyframes. Section 4.4 explains direct reward maximization and iterative generator/reward-model learning. Although the caption describes a stable upward trend, the plotted curves visibly fluctuate; the supported observation is overall increase rather than monotonic improvement. feedbackreward-diagnostics
What it supports. Both generation stages respond to their optimization signals, and improvement is not confined to a single displayed reward dimension. This is useful evidence that the feedback-training machinery changes measured outputs. It does not tell us which reward contributes most, or how much an independent viewer would prefer the resulting videos.
Where the evidence stops. This is a training diagnostic, not a controlled ablation. There are no component-removal curves, independent human test scores, or uncertainty bands. The text's claimed superiority over DPO/PPO/GRPO has no numerical comparison in this figure.
7. Analysis & limitations
7.1 What the evidence leaves open
Rising reward curves and selected pretraining/CT/SFT/RLHF frames are diagnostics, not controlled component ablations. The claimed advantage over DPO/PPO/GRPO lacks displayed comparative measurements, and reward improvement alone does not establish independent human preference improvement. reward-diagnosticsfeedback
Internal evaluation does not disclose exact rater counts, inter-rater agreement, generation settings, or confidence intervals. Its Kling 2.1 comparison must remain separate from the Arena's Kling 2.0; Figure 1 explicitly substitutes Kling 2.0 Elo when plotting Kling 2.1. evaluation-protocolexpert-resultsoverview-caveat
Multi-shot and style figures show selected outputs without aggregate continuity or style-generalization success rates. Human ratings of physical plausibility do not measure action-conditioned dynamics, closed-loop control, or robot execution. multi-shotstylesevaluation-protocolarchitecture
7.2 Questions for discussion
- How much of continued training's I2V benefit comes from task frequency versus motion-only captions and data selection?
- Do improvements in training rewards survive blinded human evaluation after aggressive distillation?
- How often do multi-shot outputs preserve subject identity and all requested events across unseen prompts?
8. Reproducibility audit
8.1 Requirements and known gaps
A faithful implementation needs the VAE, DiT, refiner, captioner, prompt rewriter, reward models and curated data. The paper omits base-model parameter count, corpus totals, complete optimizer settings, reward coefficients, merge procedure and final sampling schedule; these prevent exact reconstruction from this report alone. vaearchitecturedatatrainingfeedbackdistillation
Infrastructure describes training on thousands of GPUs using sharded data/sequence parallelism, activation offloading and workload balancing, plus dynamic post-training memory management. Exact GPU allocation, total training time and software versions are absent; the appendix supplies contributor credits rather than configuration tables. infrastructureidentity
Proposed minimal checks: separate the changed I2V sampling ratio from changed caption/data quality, and compare reward-aligned versus unaligned accelerated refiners on identical base latents with blinded human ratings. Neither check has been run here. trainingfeedbackevaluation-protocol
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Factor task frequency from motion-only captioning
Reader-proposed experiment, conditional on access to compatible checkpoints and data: start four runs from the same pretrained base in a 2×2 comparison of I2V sampling share (20% versus 40%) and caption policy (full captions versus the reported full/motion-only mixture). Hold the selected data pool, update budget, optimizer, resolution and evaluation prompts fixed, and explicitly record the mixture ratio chosen because the paper omits it. Use blinded motion, prompt-following and preservation ratings on held-out I2V cases, with uncertainty and repeated seeds. If the sampling change gives no consistent benefit at either caption policy, a task-frequency explanation of the CT gain is weakened. This isolates two factors without claiming to reproduce the full CT recipe. trainingevaluation-protocol
Check 2: Test whether refiner reward gains survive human evaluation
Reader-proposed experiment: compare the same accelerated refiner immediately before and after its RLHF stage, using identical held-out low-resolution base latents, prompts and matched random seeds. Keep output resolution, NFE and deployment settings fixed; do not update the evaluation reward models during the comparison. Measure the three reward signals and obtain blinded pairwise human judgments of motion, detail, prompt adherence and source-image preservation. Include a base-resolution condition to identify artifacts introduced during refinement, and report latency with device count and batching. If reward scores increase while human preferences are unchanged or worsen, the curves support optimization of the proxies without demonstrating a user-visible benefit. conditioning-refinerfeedbackreward-diagnosticsdistillationevaluation-protocollatency
8.3 Reading coverage
Visual audit: Visually inspected the title/version block and contributor pages, all sixteen original figures, both embedded leaderboard tables, Eq. (1), and all method, training, infrastructure and evaluation pages supporting retained claims or proposed checks. All six final crops were viewed; the preference-plot crop was widened and viewed again to preserve complete row labels. Figure 2's temporal attention label was cross-checked against Section 2.2's spatial-window description, and Figure 5's refiner-initialization arrow against Section 2.3. Figure 7 has overall upward trends with visible local decreases. The short Figure 14 caption and original prompt are retained because they identify rows and requested events; their interpretation is explained in English. Contents and references were read in the complete text but were not visually inspected. Linked full-motion examples, supplements and external leaderboard updates remain outside this pass.
PDF pages inspected for this edition: 1, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 25, 26. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract and title/version block (p. 1)
- Contents (p. 2)
- 1 Introduction (p. 3)
- 2 Model Design, including 2.1–2.4 (pp. 4–6)
- 3 Data, including 3.1–3.3 (pp. 6–8)
- 4 Model Training, including 4.1–4.4.4 (pp. 8–12)
- 5 Inference Optimizations, including 5.1–5.2 (pp. 12–14)
- 6 Training Infrastructure, including 6.1–6.2 (pp. 14–15)
- 7 Model Performance, including 7.1–7.5 and 7.2.1–7.2.3 (pp. 15–22)
- 8 Conclusion (p. 22)
- References (pp. 23–24)
- Appendix A Contributions and Acknowledgments (pp. 25–26)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Pages with little extracted text require visual inspection: 10, 26
- The extraction-related visual gaps above were addressed by inspecting all sixteen original figures, the printed equation, the embedded leaderboard tables, and pages 10 and 26. All seven supplied text chunks were read individually. Contents and reference pages were read as text; no external reference was opened.
- Identity/version note: the exact catalog title matches the inspected PDF, arXiv:2506.09113v2 [cs.CV], dated 28 June 2025. This is later than the catalog submission date of 10 June 2025. The title-page byline is ByteDance Seed; all 44 catalog individuals appear under Core Contributors on page 25. Additional Contributors on pages 25–26 remain separately credited. No earlier PDF revision was supplied, so changes from v1 cannot be assessed.
- The acquisition manifest describes the previous HTML selection as the same v2 with a different artifact hash. That HTML was not supplied for this reading; this report cites only the supplied PDF and does not assert cross-format content identity.
- Code, weights, training data, online services, the live leaderboard and linked example videos were not inspected. No experiment was reproduced. Static PDF frames cannot establish playback smoothness.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title/byline and arXiv margin; pp. 25–26, Appendix A
Exact title and ByteDance Seed byline; arXiv:2506.09113v2 [cs.CV], 28 June 2025. Page 25 lists all 44 catalog names under Core Contributors; the additional Contributors list continues onto page 26. The appendix contains credits, not experimental configurations.
Go to primary source ↓goalsPDF p. 3, Section 1
Introduces native Chinese/English T2V and image-guided video generation, balancing instruction following, motion and visual quality with efficient inference and multi-shot support.
Go to primary source ↓vaePDF p. 4, Section 2.1, Eq. (1), Temporally-Causal Compression and VAE Training
Causal encoder/decoder; pixel and latent shapes include an extra initial frame; Eq. (1) uses T rather than T+1. Ratios (4,16,16), C=48, no DiT patchification; L1, KL, LPIPS and adversarial losses; single-image special case T=T'=0.
Go to primary source ↓architecturePDF pp. 4–5, Section 2.2 and Figure 2
Spatial MMDiT blocks combine text/vision with separate modality parameters and Q/K normalization; temporal blocks contain only visual tokens and spatially partitioned attention across time. MM-RoPE interleaves captions and ordered video shots. Figure 2 labels the temporal attention module Spatial-temporal Attention; Section 2.2 clarifies its windowing.
Go to primary source ↓conditioning-refinerPDF p. 5, Unified Task Formulation and Section 2.3
Noisy inputs, clean/zero-padded conditioning frames and binary masks unify T2I/T2V/I2V. Base output is 480p; a base-initialized conditional diffusion refiner upsamples to 720p/1080p with upsampled LR input concatenated to noise.
Go to primary source ↓prompt-engineeringPDF pp. 5–6, Section 2.4
Qwen2.5-14B initializes the prompt rewriter; full SFT precedes LoRA DPO on correct versus incorrect rewrites, motivated by semantic hallucination after SFT. User prompts are rewritten into dense caption format.
Go to primary source ↓dataPDF pp. 6–7, Figure 3 and Section 3.1
Video pipeline uses shot-aware segmentation capped at 12 seconds, overlay cropping, quality/safety filtering, semantic deduplication retaining the highest-quality cluster member and category rebalancing. Source types are described broadly; corpus totals and exact filtering thresholds are not supplied.
Go to primary source ↓captioningPDF p. 7, Section 3.2
Dense captions describe dynamic actions/camera movements and static subject/style properties. Tarsier2 caption training freezes the visual encoder and fully fine-tunes the language model on Chinese and English annotations.
Go to primary source ↓pipelinePDF p. 8, Section 4 introduction and Figure 5
Training diagram separates base pretraining, CT, SFT and RLHF from a refiner branch initialized from base pretraining; inference runs prompt engineering, text encoder, base DiT and refiner to outputs.
Go to primary source ↓trainingPDF p. 9, Sections 4.1–4.3
Velocity flow matching, logit-normal sampling and resolution-aware timestep shift; progressive 256px/640px and 12/24-fps training, 3–12-second clips. CT changes I2V share 20% to 40%, data selection and caption types; SFT uses curated subsets, lower learning rates, early stopping and model merging without exact settings.
Go to primary source ↓feedbackPDF pp. 11–12, Sections 4.4.1–4.4.4
Human preferences train foundational, motion and aesthetic rewards. Base learning predicts clean x_0 when rewards can assess video, maximizes composite rewards and iterates generator/reward-model rounds. Refiner learning maximizes a linear reward mixture directly on an accelerated model. Claimed DPO/PPO/GRPO superiority has no displayed comparison values here.
Go to primary source ↓reward-diagnosticsPDF p. 10, Figure 6; p. 11, Figure 7; p. 13, Figure 8
Selected sequences compare training stages and refiner alignment. Six reward plots trend upward overall, with visible fluctuations, across foundational/motion/aesthetic rewards for base and refiner. They show neither independent human test scores nor controlled reward-component ablations.
Go to primary source ↓distillationPDF p. 12, Section 5.1
TSCD is associated with 4× acceleration, followed by score distillation and multistep adversarial preference supervision. Authors claim comparable expert quality after distillation; narrowing later VAE decoder channels with a fixed encoder is associated with 2× decoding speedup. Final NFE and full timing comparisons are absent.
Go to primary source ↓infrastructurePDF pp. 12–15, Section 5.2 and Sections 6.1–6.2
Inference uses fused kernels, mixed precision, sparsity, parallelism and offloading. Training infrastructure targets thousands of GPUs with HSDP/Ulysses, multilevel activation checkpointing and workload balancing; post-training adds CPU offloading, recomputation and dynamic memory management. No exact training allocation, total runtime or software versions are given.
Go to primary source ↓arena-i2vPDF p. 16, Figure 9 right table, Image to Video, Seedance/Veo 3 Preview/Kling 2.0 rows; Section 7.1
Seedance 1365 Elo, 95% CI offsets −13/+13, 7,280 appearances; Veo 3 Preview 1240, −10/+11, 9,298; Kling 2.0 1195, −8/+10, 10,538. Section 7.1 identifies these as public preference comparisons.
Go to primary source ↓arena-t2vPDF p. 16, Figure 9 left table, Text to Video, Seedance and Veo 3 Preview rows
Seedance 1314 Elo, 95% CI offsets −12/+12, 6,337 appearances; Veo 3 Preview 1253, −9/+9, 8,902 appearances.
Go to primary source ↓overview-caveatPDF p. 1, Figure 1 caption
Dates the dual Arena ranking claim to 10 June 2025, explicitly substitutes Kling 2.0 Elo for Kling 2.1 because public data are unavailable, and defines API speed as inverse average generation time per second of video.
Go to primary source ↓evaluation-protocolPDF pp. 16–18, Sections 7.2–7.2.3; p. 19, opening continuation of Section 7.2.3
SeedVideoBench has 300 prompts per task across subject, action, camera and aesthetic categories, with additional first-frame labels for I2V. Professional metrics cover motion, prompt following, aesthetics and I2V preservation; five-point absolute scores and pairwise GSB comparisons. T2V baselines include Kling 2.1 Master/Veo 3/Wan 2.1/Sora; I2V replaces Sora with Runway Gen4. Rater counts and full generation protocol are not supplied.
Go to primary source ↓expert-resultsPDF p. 17, Figures 10–11; p. 19, Figures 12–13 and Section 7.2.3; p. 20, Section 7.2.3 continuation
Overall GSB green segments exceed red in every displayed pair. In T2V GSB, Kling's motion and visual-quality preferences exceed Seedance's, while prompt following favors Seedance; absolute plots put Kling/Veo above Seedance on T2V aesthetics and Kling above Seedance on I2V preservation. No exact bar labels or internal confidence intervals are displayed.
Go to primary source ↓multi-shotPDF p. 20, Figures 14–15 and Figure 14 prompt/caption; p. 21, Section 7.3
Figure 14 orders Seedance, Kling 2.1 and Veo 3 from top to bottom for a prompt requesting a crying woman, a cut to bottles/cigarette ends, and an approaching man. Figure 15 gives further selected multi-shot examples; Section 7.3 claims cinematic continuity, without aggregate success rates.
Go to primary source ↓stylesPDF p. 21, Figure 16 and Section 7.4; p. 22, Section 7.5
Selected pixel-style, miniature/felt-style and animated animal-news frames illustrate stylistic variety. The text claims broader style adaptability and references further online examples, without a quantitative style-generalization evaluation.
Go to primary source ↓latencyPDF p. 1, Abstract; p. 3, Inference Acceleration and Ultra-Fast Generation Experience
Reports 41.4 seconds to generate a five-second 1080p video on NVIDIA L20 and over 10× end-to-end acceleration in the introduction. Device count, batching, uncertainty and exact timing boundaries are not specified.
Go to primary source ↓8.5 Primary sources
Seedance 1.0: Exploring the Boundaries of Video Generation Models ↗
PDF · 10,775 extracted words
Source fingerprint
e4ec66efd7e354bc7b3d25478fc1cb21709d0dc907f70cee7000ad2bf036fd13