Wan: Open and Advanced Large-Scale Video Generative Models
1. Paper overview
In one sentence: Wan makes video diffusion practical through causal latent compression, progressive image-video training and reusable conditioning, while its evaluation establishes media-generation quality rather than an executed-action model. e01e04e07e08e09e14e15e16e18e23e24e25
| At a glance | What to know |
|---|---|
| Research problem | Source description The paper targets open video models that must improve visual quality, motion, prompt adherence and accessibility together. Its central challenge is making long visual sequences tractable while preserving enough spatial and temporal information for controllable generation. e01e04e07 |
| Core mechanism | |
| A key reported result | Text-to-video generation on VBench: Wan14B: 86.22% / 86.67% / 84.44%; Wan1.3B total: 83.96%. Total / quality / semantic scores (higher is better). Paper-reported leaderboard snapshot; Wan14B row dated 2025-02-24; comparator versions/settings retained in Table 4. Sora total 84.28%; open-source Hunyuan total 83.24%. Wan14B leads this reported snapshot. These percentages are benchmark scores, not task-success rates; uncertainty and uniform per-model generation settings are not supplied in the table. e15 |
| Reading caution | Source description The authors identify lost fine details during large motion, expensive 14B inference and limited domain-specific expertise. Their roughly 30-minute unoptimized single-GPU statement lacks a fixed GPU model, duration and resolution, so it is not a portable latency benchmark. e27 |
Core contributions
Figure 5. Compression reduces the sequence that the generative model must learn. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the arrows from the RGB frame stack through the encoder to the central latent tensor, then back through the decoder. The blue-green block reduces spatial dimensions; the two orange blocks each reduce both space and time. Together they yield eightfold spatial and fourfold temporal compression. Read the tensor labels carefully: the input contains 1+T frames, whereas the latent contains 1+T/4 positions. The isolated first frame receives only spatial compression, allowing the same design to handle images. The text fixes the latent channel count C at 16 and explains that causal convolutions and RMSNorm enable processing successive chunks. e04e05e06e07
What it supports. The diagram explains how Wan reduces visual sequence size before applying its transformer. The companion reconstruction experiment uses 200 videos of 25 frames at 720×720 and reports 2.5× reconstruction speed over HunyuanVideo. That is evidence about the VAE component, with a comparison protocol distinct from the text-to-video leaderboards.
Where the evidence stops. Temporal causality here belongs to the VAE; the base DiT uses noncausal attention. Compression ratios alone do not guarantee faithful reconstruction, and the reported speed comparison does not isolate the contribution of feature caching.
2. Motivation
2.1 The problem and the proposed response
The paper targets open video models that must improve visual quality, motion, prompt adherence and accessibility together. Its central challenge is making long visual sequences tractable while preserving enough spatial and temporal information for controllable generation. e01e04e07
2.2 What this reading follows
Wan is easiest to understand as a sequence of engineering choices around a video diffusion transformer. First compress pixels into a manageable latent sequence; then learn a text-conditioned velocity field; finally add the conditions and acceleration needed by particular applications. This reading follows the April 2025 v2 report from those choices to its actual evidence. The tables support strong aggregate video-generation performance, but dimension-level weaknesses, image-stage ablations and unresolved reporting inconsistencies qualify that conclusion. The six original excerpts below connect the method to its measurements and explain why image conditioning, camera control and streaming remain distinct from physical action execution. e01e04e07e08e09e14e15e16e18e23e24e25
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-backbone category is supported. Wan predicts media latents, and its camera/mask conditions steer synthesis; the paper supplies neither joint future/action outputs nor an inverse-dynamics action extractor or executed feedback-control evaluation. Thus the recorded architecture, prediction-paradigm and quadrant 'Not applicable' values fit the world-action taxonomy. A unified editing model is not evidence for a unified world/action model. e07e08e18e21e23e24
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
5. Method in detail
5.1 Separate compression from generation
The VAE and DiT solve different learning problems. Wan first trains a video representation that can reconstruct pixels with spatial and temporal compression. Its first-frame exception makes image and video inputs compatible, while causal feature caches keep successive encoding or decoding chunks connected. The DiT then learns a distribution over these representations. Equations (1)–(3) construct an intermediate latent between Gaussian noise and a clean training latent and ask the network to predict their difference as a velocity. At inference, repeated model evaluations move a noise sample toward a generated latent before decoding. This distinction explains two otherwise confusing features of Figure 9: a training video is shown even though text-to-video inference needs no reference video, and the VAE's temporal causality does not imply causal attention inside the video DiT. e04e05e07e08e11
Figure 9. The model learns to transform noisy video latents under text conditioning. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the horizontal path as the representation pipeline: a training video is encoded, its latent is combined with noise, the DiT processes it, and the decoder returns pixels. The lower path carries the caption through umT5 into cross-attention; the upper path supplies diffusion time. Figure 10 and Equations (1)–(3) specify what this overview abbreviates: video tokens interact through self-attention, text enters separately, and the model predicts velocity along a noise-to-data interpolation. For text-to-video inference, begin from noise and the prompt. The left-hand training video is not an additional input that a user must provide. e07e08e09e11
What it supports. The backbone separates visual compression, language representation and learned latent dynamics. During foundation-model training, only the DiT is optimized; the text and VAE encoders stay frozen. This distinction helps identify where the generative capacity is being trained and why the diagram should not be read as a single end-to-end reconstruction objective.
Where the evidence stops. The arrows are a pipeline overview, not a complete sampler specification. The output is generated imagery: neither an action head nor an inverse-dynamics module appears, and iterative denoising is not environment feedback.
5.2 Read the training curriculum alongside the ablation
The training curriculum tackles the cost of learning semantics and temporal detail at once. Wan14B starts with low-resolution images, then introduces short videos and progressively increases resolution. The caption pipeline gives those examples detailed language supervision, while inference-time prompt rewriting tries to reduce the mismatch with brief user requests. Within the DiT, sharing the timestep-modulation MLP saves parameters that can support more layers. Figure 16 tests this allocation during the image stage: its strongest parameter-matched comparison trades half sharing at 30 layers for full sharing at 35 layers. Reader interpretation: that is evidence for a useful early optimization choice, not a demonstration that sharing itself improves motion. A full causal test would continue comparable checkpoints through video training and measure the same quality dimensions used for final evaluation. e02e03e07e09e12e16
Figure 18. Image conditioning supplies both spatial anchors and global visual context. Original paper, p. 27 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the red first-frame bracket. The upper branch pads that image with zero frames and encodes the resulting sequence. Its latent joins the noisy input, drawn as X_t and called z_t in the text, and the rearranged binary mask before entering DiT. Section 5.1.1 defines this mask explicitly: 1 marks preserved frames and 0 marks frames to generate. The lower image branch uses CLIP and a projection MLP to supply global context, alongside umT5 text through decoupled cross-attention. The training description matters: the image-encoder branch is omitted during initial pre-training and introduced during supervised fine-tuning. e18e19e20e21
What it supports. The same mask-guided formulation accommodates initial-frame animation, continuation, first-last-frame transformation and interpolation before task-specific refinement. Its two image pathways have different roles: latent conditioning anchors supplied visual content, while CLIP supplies global context. This is a reusable media-conditioning mechanism, not evidence that the model extracts or executes actions.
Where the evidence stops. Do not transfer this mask convention to VACE: Section 5.2 defines 1 as an editing region there. The figure alone does not demonstrate exact pixel preservation or quantify the benefit of CLIP; the reported human-comparison terminology also remains ambiguous.
5.3 Understand what the control signals actually control
Image conditioning changes the information available to video synthesis. An I2V mask distinguishes supplied frames from unknown frames, and the later CLIP branch adds image-level context. VACE generalizes editing through context frames and masks, but reverses the meaning of mask value 1; an implementation must preserve that task-specific convention. Personalization supplies reference faces in latent space, and the camera extension injects features derived from a specified camera trajectory. Streamer changes temporal computation again, using a sliding queue and cached completed tokens rather than unlimited attention over the past. Reader interpretation: these are several ways to steer or extend generated media. None supplies the action-output mechanism or executed feedback experiment required to turn this paper into a world-action policy result. This supports retaining the catalog's foundational-video-backbone classification. e18e19e21e22e23e24
5.4 Training and inference
During training
Train the VAE first: 2D image learning, inflation to 3D on five-frame 128×128 videos, then high-quality video refinement with a 3D adversarial loss. Initial video training weights L1, KL and LPIPS losses by 3, 3×10⁻⁶ and 3. e05
For 14B, begin with 256px images; jointly train 256px images and five-second 192px/16fps videos, then progress to 480px and 720px. Use bf16 AdamW, weight decay 10⁻³ and initial learning rate 10⁻⁴, reduced on FID/CLIP plateaus. Post-training uses curated 480px/720px data. Only DiT is optimized; text and VAE encoders remain frozen. e09
FSDP and Ring/Ulysses context parallelism distribute long sequences; activation offloading and selective checkpointing manage memory. The illustrated 128-GPU arrangement is an example, not the total compute expenditure. e10
During inference
Rewrite user prompts with Qwen2.5-Plus, encode text, then evolve noise toward video latents using the learned velocity and classifier-free guidance, typically over about 50 sampling steps. Decode with Wan-VAE. Training videos in the architecture diagram are not required inputs for text-only generation. e07e08e11e12
Optional distributed execution, attention/guidance caching and reduced-precision kernels improve latency. Their validation-selected settings belong to the inference configuration; they should not be conflated with the separately distilled streaming model. e11e24
5.5 Implementation flow
- Curate and describe the visual distribution
Deduplicate and filter source media, score visual quality within clusters, and grade motion. Separate text-rendering data combine synthetic Chinese characters with OCR-assisted real examples. A ViT–MLP–Qwen captioner supplies dense descriptions; curated post-training media emphasize visual and motion quality. e02e03
- Compress pixels before generative modeling
Wan-VAE converts (1+T) RGB frames into (1+T/4) latent frames with eightfold spatial downsampling and 16 channels. The first frame has spatial compression only. Causal convolutions and RMSNorm permit chunk-wise feature caching; the VAE contains 127M parameters. e04e05
- Predict latent velocity with a DiT
A (1,2,2) patchifying convolution produces video tokens. Noncausal spatio-temporal self-attention models their relationships, cross-attention supplies umT5 text, and a shared timestep MLP supplies modulation with block-specific biases. The prediction target is the flow velocity, followed by latent decoding into pixels. e07e08
- Add task-specific visual conditions
I2V concatenates noisy latents, encoded reference-plus-zero frames and a mask; 1 means preserved. CLIP features enter through decoupled cross-attention during SFT. VACE instead uses 1 for regions to edit and offers full fine-tuning or a context adapter. Personalization prepends latent reference faces; camera adapters inject Plücker features. These mechanisms condition generated media. e18e19e21e22e23
- Extend temporal and output scope
Streamer replaces fixed-length generation with a sliding denoising queue, discards warmup tokens and reuses completed tokens at zero noise; consistency distillation uses four steps. Joint training also supports still images. A separate waveform-VAE/DiT produces video-conditioned ambient audio and music, excluding speech. e24e26e28
6. Experiments & results
Wan is a video-generation foundation-model family combining a causal video VAE, a text-conditioned diffusion transformer and large-scale curated image/video training. Its strongest evidence is comparative generation quality; image, editing, identity, camera, streaming and audio extensions broaden media creation without establishing an executed-action policy.
6.1 Read the original evidence
Table 2. An aggregate lead coexists with clear dimension-level weaknesses. Original paper, p. 22 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First locate the bottom weighted-score row, then return to individual rows rather than treating the aggregate as universal superiority. Wan14B scores 0.724, Sora 0.700 and Hunyuan 0.673 on 1,035 samples per candidate. The preceding benchmark section explains that the rows mix optical-flow measurements, detector outputs, feature similarities and Qwen2-VL judgments. Their contributions are weighted using correlations with human preferences, rather than simply averaged. Physical plausibility is therefore a model-judged visual property. Large motion, human artifacts, stability and stylization should each be checked separately when choosing a generator for a specific use case. e13e14
What it supports. Wan14B has the highest weighted score in this comparison and physical-plausibility scores of 0.939 versus Sora's 0.933. Yet its large-motion score is 0.415 versus Sora's 0.482, and its stylization score is 0.328 versus CNTopB's 0.623. The evidence supports aggregate strength with an uneven capability profile.
Where the evidence stops. Wan-Bench is constructed by the authors, includes learned evaluators and uses preference-derived weights. No uncertainty intervals accompany this table, and anonymized commercial columns limit exact replication. A high physical-plausibility score does not validate a physics simulator or control policy.
Table 4. The external benchmark snapshot complements the internal Wan-Bench comparison. Original paper, p. 24 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row's full model label before comparing its three scores. Wan14B is dated 2025-02-24; Gen-3 and Kling include older release settings, and CogVideoX is explicitly marked as a five-second, SAT, prompt-optimized configuration. The rightmost column reports total score; the two preceding columns separate quality from semantics. Wan14B reaches 86.22% total, with 86.67% quality and 84.44% semantic score. Wan1.3B reaches 83.96% total. This table is the paper's reported VBench snapshot, so its numerical claims belong to those listed entries and cannot automatically be applied to later versions of any system. e15
What it supports. Wan14B's 86.22% total exceeds Sora's 84.28% and open-source Hunyuan's 83.24% in the reported table. Wan1.3B is also competitive, though below Sora. This gives a second benchmark supporting the family's generation quality, while keeping the compact and large model results separate.
Where the evidence stops. These percentages are benchmark scores, not fractions of successfully executed tasks. The table provides no uncertainty estimates or full common generation recipe. It does not show a matched-data or matched-compute test of architecture alone.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Text-to-video generation on VBench Paper-reported leaderboard snapshot; Wan14B row dated 2025-02-24; comparator versions/settings retained in Table 4. | Wan14B: 86.22% / 86.67% / 84.44%; Wan1.3B total: 83.96%. Total / quality / semantic scores (higher is better) | Sora total 84.28%; open-source Hunyuan total 83.24%. Wan14B leads this reported snapshot. These percentages are benchmark scores, not task-success rates; uncertainty and uniform per-model generation settings are not supplied in the table. e15 |
| Text-to-video generation on Wan-Bench 1,035 samples per candidate; 14 dimensions weighted using human-feedback correlations. | Wan14B 0.724; Wan1.3B 0.689. Weighted score (higher is better) | Sora 0.700; Hunyuan 0.673; Mochi 0.639. The aggregate favors 14B, but individual rankings differ: large motion is 0.415 versus Sora 0.482; stylization is 0.328 versus CNTopB 0.623. Internal evaluators and preference weights influence the conclusion. e13e14 |
| Adaptive-normalization parameter allocation Text-to-image pre-training from scratch for 200,000 steps, global batch 1536; four sharing/depth configurations. | The fully shared, 35-layer 1.5B configuration has the lowest endpoint curve; no precise endpoint table is supplied. Training loss (lower is better) | Half-shared 30-layer 1.5B and non-shared 30-layer 1.7B end higher. Supports allocating parameters to depth in this image-stage experiment. It does not isolate sharing from depth, match FLOPs, or establish final video quality. e16 |
| Video reconstruction efficiency 200 videos, 25 frames, 720×720; same hardware stated, GPU model unspecified. | Author-reported 2.5× versus HunyuanVideo. Relative reconstruction speed | Figure 7 also compares PSNR and parameter size; several other VAEs use different compression settings. A component-level efficiency result, not end-to-end video generation latency or an isolated cache ablation. e06 |
| Reference-face video personalization Unseen evaluation set; ArcFace similarities averaged over faces sampled at 1 fps; evaluation-set size unspecified. | Wan 0.5526. ArcFace similarity (higher is better) | CN-TopA 0.5655; CN-TopB 0.5197; CN-TopC 0.4998. Competitive identity matching, with one comparator higher. Identity similarity alone does not measure motion quality or overall preference. e22 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 16. Sharing timestep modulation frees parameters that can instead increase depth. Original paper, p. 25 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend to distinguish the blue fully shared 1.3B model, orange fully shared 1.5B model, green half-shared 1.5B model and red non-shared 1.7B model. The horizontal axis is training step and the vertical axis is training loss. Section 4.7.2 supplies the essential control: the orange model has 35 layers, while the other configurations have 30. All are trained from scratch on text-to-image for 200,000 steps with global batch 1536. The closest parameter-matched comparison is orange versus green. Follow their late-training separation, rather than attributing every difference among all four curves solely to normalization sharing. e16
What it supports. The fully shared, deeper 1.5B model ends with lower training loss than the half-shared 1.5B and non-shared 1.7B alternatives. The authors use this result to motivate shared modulation. It supports a parameter-allocation choice during image pre-training, with a measurable proxy objective rather than a final video-quality measurement.
Where the evidence stops. Depth and sharing change together in the parameter-matched comparison, and FLOPs are not held equal. The source prose says the deeper model is consistently lowest, but early curves cross or overlap; the clear claim is the lower endpoint. No seed variability or endpoint values are tabulated.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors identify lost fine details during large motion, expensive 14B inference and limited domain-specific expertise. Their roughly 30-minute unoptimized single-GPU statement lacks a fixed GPU model, duration and resolution, so it is not a portable latency benchmark. e27
Wan-Bench physical plausibility is a Qwen2-VL judgment, not measured physical dynamics. Aggregate wins do not establish reliable control, and headline tables provide no uncertainty estimates. Several competitors remain anonymized. e13e14e15e20
Reporting ambiguities remain unresolved: Tables 3/7 mix 'win rate gap' with preference-proportion descriptions, including −4.2% in Table 7; Table 5 says 10k/15k steps while prose says 100k/150k. Figure 31 says 20 FPS on one RTX 4090, whereas text says 8 FPS. These are not reconciled measurements. e17e20e25
Streaming assumes bounded temporal dependencies; unlimited queue operation does not prove unlimited semantic consistency. TensorRT can introduce flicker/artifacts. Audio evaluation is qualitative, and speech/vocal generation is explicitly limited. e24e25e26
7.2 Questions for discussion
- Does the deeper shared-modulation model retain its advantage after video training and equal-compute comparison?
- How sensitive are model rankings to prompt rewriting, evaluator choice and human-preference weights?
- How far can streaming preserve identity and scene state beyond its local temporal window?
8. Reproducibility audit
8.1 Requirements and known gaps
A faithful training reproduction needs the curated media, grading/captioning models, stage mixtures, checkpoint transitions and long-sequence infrastructure. The paper leaves many filtering thresholds and stage budgets unspecified; I2V explicitly mentions an unreported SigLIP cutoff. The release claim alone does not establish a reproducible data pipeline. e01e02e03e09e10e19
Reader-proposed checks: repeat the 1.5B depth-versus-sharing comparison with both token and compute controls, then compare cached and uncached inference on identical seeds/prompts with blinded motion/detail ratings. Report variability and full runtime configurations rather than assuming training loss or cache similarity guarantees visual equivalence. e11e16e27
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: Does the depth allocation survive fair budgets and video training?
Reader-proposed experiment, not performed here: reproduce the fully shared 35-layer 1.5B and half-shared 30-layer 1.5B image-stage configurations with the same data order, text encoder, VAE and multiple initialization seeds. First match training tokens and global batch as in the paper; then run a separate comparison matched by measured training FLOPs. Continue both through identical video-training stages. Record training/held-out flow loss and blinded video quality, motion and prompt adherence. A lower image loss without a reliable video-quality advantage would falsify the proposed transfer of the ablation's benefit to final generation; an advantage only under token matching would identify a compute tradeoff. e07e08e09e16
Check 2: Does diffusion caching preserve difficult video details?
Reader-proposed experiment, not performed here: compare uncached and cached inference from the same checkpoint on identical original and rewritten prompts, initial latent seeds, resolution, frame count, solver schedule and guidance setting. Select cache steps on a disjoint validation set and freeze them before testing. Keep precision and hardware fixed so quantization does not confound the comparison. Include large-motion and visible-text cases, measure end-to-end latency including decoding, and obtain blinded ratings for fine-detail retention, artifacts, temporal stability and prompt adherence with uncertainty across seeds. A repeatable quality deficit would contradict lossless caching for that test distribution even if average latency improves. e04e11e12e27
8.3 Reading coverage
Visual audit: The title/version/byline on p. 1 and full contributor roster on p. 49 were visually checked. Original architecture, data, evaluation, ablation and extension pages listed here were rendered and viewed, including low-text image pages 36 and 39. All six final original crops were individually viewed; labels, legends, arrows, table rows and axes are retained. Figure 5 was checked against Sections 4.1.1–4.1.3; Figure 9 against Figure 10 and Equations (1)–(3); Figure 18 against Sections 5.1.1–5.1.3, including the mask convention and staged CLIP branch. Figure 16's endpoint is supported, while its early crossings qualify the prose's 'consistently' claim. Uncropped pages supply the remaining retained training, evaluation, hardware, limitation and reproduction details. Figures 2, 15, 19, 20 and 33 were not visually inspected; their text was read. No video playback, audio listening, external artifact inspection or experimental reproduction is claimed. There is no separate appendix in the supplied PDF.
PDF pages inspected for this edition: 1, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 24, 25, 26, 27, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 47, 48, 49. Appendix coverage: not present.
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 identity, p. 1; contents, p. 2
- Sections 1–2: introduction and related work, pp. 3–5
- Section 3: pre-/post-training data and dense captioning, pp. 5–10
- Sections 4.1–4.2: VAE, DiT, objectives and training, pp. 10–15
- Sections 4.3–4.5: scaling, inference and prompt alignment, pp. 15–20
- Sections 4.6–4.7: benchmarks, results and ablations, pp. 20–26
- Sections 5.1–5.5: I2V, editing, image generation, personalization and camera control, pp. 26–40
- Sections 5.6–5.7: streaming, distillation and audio generation, pp. 38–47
- Sections 6–7: limitations, conclusion and contributors, pp. 47–49
- References, pp. 50–60; all 15 supplied text chunks individually read
Outside the original text pass
- Version scope: the inspected PDF is arXiv:2503.20314v2 dated 19 April 2025, not the catalog submission date of 26 March 2025. Its title matches. The observed byline Wan Team corresponds to catalog Team Wan; all 61 individual contributors match, including two separately marked Wei Wang entries.
- The acquisition snapshot identifies its earlier HTML selection as the same v2 revision. Only the supplied PDF is a scientific source for this report; the earlier HTML and original submission were not independently compared.
- 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: 36, 39
- The extraction cautions were addressed by inspecting original PDF pages and all six final crops, including low-text pages 36 and 39. Visual inspection covered the pages declared in the illustrated edition; Figures 2, 15, 19, 20 and 33 were read through their supplied text but not visually inspected.
- No separate appendix is present in this PDF. Linked supplements, repositories, weights, datasets and live leaderboards were not accessed. Videos were not played, audio was not heard, and experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, arXiv margin and Abstract; p. 49, Section 7 Contributors
The title matches the catalog. The byline is Wan Team, Alibaba Group; the margin identifies arXiv:2503.20314v2, 19 April 2025. Page 49 lists the 61 individual contributors, including two separately marked Wei Wang entries. The abstract introduces 1.3B and 14B video foundation models.
Go to primary source ↓e02PDF pp. 6–8, Sections 3.1–3.2; p. 7, Figure 3
Pre-training data come from internal copyrighted and publicly accessible sources, followed by deduplication, basic filtering, cluster-based quality selection and motion grading. Separate synthetic Chinese-character and OCR-assisted real-image branches support text rendering. Post-training selects curated images and videos.
Go to primary source ↓e03PDF pp. 8–10, Sections 3.3.1–3.3.4; p. 9, Figure 4
The dense captioner uses a ViT, a two-layer projection and Qwen. It is trained through projection alignment, all-parameter training and high-quality refinement. Caption evaluation covers ten dimensions; the internal model and Gemini 1.5 Pro each lead different dimensions.
Go to primary source ↓e04PDF p. 10, Figure 5 and Section 4.1.1; p. 11, Section 4.1.1 continuation
Wan-VAE maps (1+T)×H×W RGB frames to (1+T/4)×(H/8)×(W/8) latents with 16 channels. The first frame is only spatially compressed. Causal convolutions and RMSNorm preserve temporal causality; the VAE has 127M parameters.
Go to primary source ↓e05PDF p. 11, Sections 4.1.2–4.1.3 and Figure 6; p. 12, Section 4.1.3 continuation
VAE training proceeds from a 2D image VAE to inflated 3D training on 128×128 five-frame clips and high-quality refinement with a 3D GAN loss. Initial video losses weight L1, KL and LPIPS by 3, 3e−6 and 3. Chunk processing caches two prior features for kernel-size-three convolutions, or one for noninitial stride-two temporal downsampling.
Go to primary source ↓e06PDF p. 12, Section 4.1.4 and Figure 7
The VAE comparison uses 200 videos of 25 frames at 720×720. The authors report 2.5× reconstruction speed relative to HunyuanVideo on the same unspecified hardware. Figure 7 plots reconstruction PSNR, efficiency and parameter size. Compression and channel counts differ for several comparators.
Go to primary source ↓e07PDF p. 13, Figure 9 and Section 4.2; p. 14, Section 4.2.1 and Figure 10; p. 15, Section 4.3.1
Wan combines a VAE, DiT and umT5. A (1,2,2) convolution patchifies latents; DiT uses video self-attention, text cross-attention and shared timestep modulation with block-specific biases. Attention is noncausal in the DiT. Only DiT parameters are optimized in foundation-model training.
Go to primary source ↓e08PDF pp. 14–15, Section 4.2.2, Equations (1)–(3)
Flow matching linearly interpolates Gaussian noise x0 and clean latent x1 at logit-normal timesteps. The target velocity is x1−x0. Squared-error training predicts that velocity conditioned on a 512-token umT5 sequence and timestep.
Go to primary source ↓e09PDF p. 15, Sections 4.2.2–4.2.3 and 4.3.1
The 14B curriculum begins with 256px image pre-training, then 256px images plus five-second 192px videos at 16 fps, then 480px and 720px image-video stages. It uses bf16, AdamW, weight decay 1e−3 and initial learning rate 1e−4 reduced on FID/CLIP plateaus. Post-training retains architecture and optimizer; text and VAE encoders remain frozen while DiT is optimized.
Go to primary source ↓e10PDF pp. 16–17, Sections 4.3.2–4.3.4 and Figure 11
Training combines FSDP, data parallelism and two-dimensional context parallelism with Ring Attention outside Ulysses. Activation offloading is combined with selective gradient checkpointing. Figure 11 is explicitly an illustrative 128-GPU configuration, not a reported total training budget.
Go to primary source ↓e11PDF pp. 17–19, Sections 4.4.1–4.4.3; p. 17, Figure 12
Inference typically uses about 50 sampling steps with classifier-free guidance. The acceleration methods include FSDP/context parallelism, cached attention and guidance computations, FP8 GEMMs and mixed INT8/FP8 attention with FP32 cross-block accumulation. Cache steps are selected on validation data.
Go to primary source ↓e12PDF pp. 19–20, Section 4.5 and Table 1
Multiple caption lengths/styles are used during training. Qwen2.5-Plus rewrites inference prompts to resemble dense training captions, adding descriptive and motion detail while being instructed to preserve meaning. Table 1 provides translated examples, not a numerical rewriting ablation.
Go to primary source ↓e13PDF pp. 20–21, Section 4.6 and Figures 13–14
Wan-Bench has 14 fine-grained dimensions across dynamic quality, image quality and instruction following. It combines detectors, optical flow, representation similarity and Qwen2-VL judgments. More than 5,000 pairwise comparisons inform Pearson-correlation-based dimension weights. Physical plausibility is evaluated with Qwen2-VL video questions.
Go to primary source ↓e14PDF p. 22, Table 2, Weighted Score and selected dimension rows; Section 4.7.1
Wan-Bench evaluates 1,035 samples per candidate. Weighted scores are Wan14B 0.724, Sora 0.700, Hunyuan 0.673, Mochi 0.639 and Wan1.3B 0.689. Wan14B physical plausibility is 0.939 versus Sora 0.933, but large motion is 0.415 versus 0.482 and stylization 0.328 versus CNTopB 0.623.
Go to primary source ↓e15PDF p. 24, Table 4; p. 22, Section 4.7.1, Wan in public leaderboard
The reported VBench snapshot gives Wan14B (2025-02-24) quality 86.67%, semantic 84.44%, total 86.22%; Sora total 84.28%; Wan1.3B total 83.96%; and open-source Hunyuan total 83.24%. Rows retain different dated/versioned settings, including prompt-optimized CogVideoX.
Go to primary source ↓e16PDF p. 24, Section 4.7.2, adaptive-normalization ablation; p. 25, Figure 16 and adjacent interpretation
Four models are trained from scratch on text-to-image for 200,000 steps at global batch 1536. Full sharing uses 30 layers at 1.3B, or 35 at 1.5B; half sharing uses 30 layers at 1.5B; no sharing uses 30 layers at 1.7B. The deeper, fully shared 1.5B curve ends below the others. No numerical endpoint table or uncertainty interval is provided.
Go to primary source ↓e17PDF p. 25, Tables 5–6 and Section 4.7.2; p. 26, Figure 17 and autoencoder-ablation continuation
umT5 has lower training loss than the tested Qwen2.5 and GLM alternatives. Table 6 reports FID 43.01 for umT5 and 42.91 for second-last-layer Qwen-VL features. Table 5 labels its VAE/VAE-D rows 10k and 15k steps, whereas adjacent prose specifies 100,000 and 150,000.
Go to primary source ↓e18PDF pp. 26–27, Section 5.1.1 and Figure 18
I2V concatenates noisy latents, encoded first-frame-plus-zero-padding conditions and a rearranged mask. Here mask 1 means preserved and 0 means generated. Additional input projection starts at zero. CLIP image features, projected by an MLP, supply global context through decoupled cross-attention.
Go to primary source ↓e19PDF p. 27, Sections 5.1.2–5.1.3; p. 30, Section 5.1.3 continuation
Mask-guided pre-training spans I2V, continuation, endpoint transformation and interpolation; task-specific refinement follows. The image encoder branch is omitted in pre-training and introduced during SFT. I2V data are filtered using first-frame/remaining-frame SigLIP similarity above an unspecified threshold.
Go to primary source ↓e20PDF p. 24, Table 3 and its footnote; p. 30, Table 7 and its caption
Both human-evaluation tables are titled win rate gap but described as preference proportions. Table 7 contains a negative Matching value of −4.2% against CN-TopA, which cannot be an ordinary preference proportion. The report does not resolve this terminology.
Go to primary source ↓e21PDF pp. 30–32, Sections 5.2–5.2.1, Figure 21 and Equation (4); pp. 34–35, Sections 5.2.2–5.2.3
VACE unifies text, context frames and masks, with mask 1 identifying regions to edit. Reactive and preserved frame streams are separately encoded. Full fine-tuning or a context adapter is offered; adapter tuning preserves base weights. The subsection's evaluation presents qualitative examples.
Go to primary source ↓e22PDF pp. 35 and 37, Section 5.4.1 and Figure 26; pp. 37–38, Table 8 and Section 5.4.3
Personalization prepends reference faces in VAE latent space and uses channel-wise conditions and masks. On an unseen evaluation set, ArcFace similarities average faces sampled at 1 fps. Wan scores 0.5526 versus CN-TopA 0.5655, CN-TopB 0.5197 and CN-TopC 0.4998.
Go to primary source ↓e23PDF p. 38, Section 5.5; p. 40, Figure 28
Camera control transforms camera intrinsics/extrinsics into Plücker features, encoded and injected through scale/shift adapters. Figure 28 marks the camera encoder/adapter trainable and existing attention/FFN blocks frozen. Outputs are generated video conditioned on a supplied camera trajectory.
Go to primary source ↓e24PDF pp. 40–44, Sections 5.6.1–5.6.3, especially pp. 42–43 streaming algorithm
Streamer fine-tunes a pre-trained DiT with a fixed sliding denoising queue. Training uses 2w tokens, omitting the first w warmup tokens from the loss; inference similarly discards warmup tokens. Completed tokens return at zero noise for continuity. LCM/VideoLCM distillation reduces generation to four steps.
Go to primary source ↓e25PDF p. 42, Figure 30 caption; p. 43, Figure 31 caption; p. 44, Section 5.6.3
Figure 30 reports a 15-minute video at 8 FPS on eight A100 GPUs. Figure 31 reports 20 FPS on a single RTX 4090 with INT8 and TensorRT, while p. 44 states 8 FPS on a single 4090. The text acknowledges TensorRT-related artifacts, flicker and temporal incoherence; matching configurations are not established.
Go to primary source ↓e26PDF pp. 44–45, Section 5.7 and Figure 32; p. 47, implementation, evaluation and limitations
V2A uses waveform-based 1D-VAE latents, temporally aligned CLIP video features, frozen umT5 text features and a flow-matching DiT. It targets ambient sound/music, deliberately excludes speech/vocal data, and provides qualitative evaluation. Human vocal generation remains limited.
Go to primary source ↓e27PDF p. 47, Section 6 Limitation; p. 48, Conclusion continuation
The authors identify difficulty preserving fine details under large motion, high inference cost and insufficient specialized-domain expertise. They report approximately 30 minutes for unoptimized 14B inference on an unspecified single high-end GPU, without fixing clip duration/resolution in that statement.
Go to primary source ↓e28PDF p. 35, Section 5.3; p. 36, Figure 25; p. 39, Figure 27
The joint image/video model also generates still images. Figure 25 is an image-sample montage; Figure 27 shows reference identities and selected personalized video frames. These low-extracted-text pages contain visual examples rather than missing source material.
Go to primary source ↓8.5 Primary sources
Wan: Open and Advanced Large-Scale Video Generative Models ↗
PDF · 27,435 extracted words
Source fingerprint
ca859378a8296fe1a09fb5127680bcdc7e6499fb340e989a598c904720424e9c