Waver 1.0 vs Seedance: ByteDance's Other Video Model

Waver 1.0 is a real ByteDance video model, but it is not a product — it is a research release, and it is not part of the Seedance line. It has an arXiv technical report (2508.15761, submitted August 21, 2025) and an official repository under ByteDance’s FoundationVision organisation, and it is a 12-billion-parameter model that generates 720p video and upscales to 1080p. What it does not have, as of our 2026-08-24 check, is a listing on any ByteDance product surface: zero mentions on the Seed model page, the Seed research index, the Seed blog index, or the BytePlus ModelArk model list — all four of which list Seedance.

That gap is the whole story of this comparison. If you found Waver while searching for “Waver AI” and expected a signup page, the short answer is that you have found a paper, not a product. This page lays out exactly what is documented, exactly what is claimed, and exactly what nobody outside ByteDance has verified.

What Waver 1.0 Actually Is — and Who Built It

Waver 1.0 is described in its own README as “a next-generation, universal foundation model family for unified image and video generation, built on rectified flow Transformers and engineered for industry-grade performance” (read 2026-08-24). “Unified” is the key word: one model handles text-to-video, image-to-video and text-to-image rather than three separate checkpoints.

The ByteDance connection is documented, not inferred. The repository lives at github.com/FoundationVision/Waver, and the FoundationVision GitHub organisation’s own bio reads “Bytedance’s opensource FoundationVision models” (read 2026-08-24). The same organisation hosts VAR — the NeurIPS 2024 best-paper visual autoregressive model — plus LlamaGen, Infinity and ByteTrack. The README’s “try it” link points at opensource.bytedance.com/discord/invite, a bytedance.com host.

The technical report is credited to Yifu Zhang, Hao Yang, Yuqi Zhang, Yifei Hu, Fengda Zhu, Chuang Lin, Xiaofeng Mei, Yi Jiang, Zehuan Yuan and Bingyue Peng. The arXiv listing page does not print institutional affiliations, so “ByteDance” here rests on the repository and organisation evidence above rather than on an affiliation line in the paper.

One more piece of dating: the repository was created on 2025-08-05 and its last code push was 2025-08-27 (GitHub API, read 2026-08-24). Waver 1.0 is roughly a year old at the time of writing, and nothing in the official repository has moved since a month after launch.

Waver 1.0 vs Seedance: Side by Side

The two models differ less on specs than on status — one is shipped, the other is published.

Two rows deserve a note. The parameter row is one of the few places where Waver is more transparent than Seedance: the README prints a model table giving 12B parameters, 16 dual-stream blocks, 40 single-stream blocks, 24 attention heads and a head dimension of 128. ByteDance has never published an equivalent table for Seedance. And the resolution row is not a fair fight — a 2025 research model at 720p native is being compared against a 2026 flagship that added native 4K at the June 23, 2026 upgrade. Our Seedance version timeline tracks which Seedance versions are live on which surface; nothing in it has ever needed a Waver row.

Where the “Better Than Veo 3” Claims Come From

Search for Waver 1.0 and you will hit a wall of headlines saying it beats Kling, Wan and Veo 3. It is worth being precise about what ByteDance actually wrote, because the official wording is narrower than the circulated version.

The README makes two performance claims in text. The first: Waver “Ranks in the Top 3 on both the T2V leaderboard & I2V leaderboard at Artificial Analysis (data up to 2025-08-05 12:00 (GMT+8)), consistently outperforms existing open-source models and matches or exceeds state-of-the-art commercial solutions.” The arXiv abstract makes the same claim with an earlier cut-off, “data as of 2025-07-30 10:00 GMT+8”. Both are snapshots of a leaderboard that reshuffles constantly — a “Top 3” from August 2025 says nothing about August 2026.

The second: after describing comparisons on Waver-Bench 1.0 and the Hermes Motion Testset, the README states “Through manual evaluation, Waver1.0 are superior to those from both closed-source and open-source models” (quoted as written, including the grammatical slip). Note what is missing — no competing model is named anywhere in that sentence, or anywhere else in the README’s prose. The specific match-ups you see in third-party write-ups come from a chart image in the repository and from those articles’ own reading of it, not from a named comparison in the official text.

Both benchmarks are ByteDance’s own constructions, and the README describes them openly: Waver-Bench 1.0 is “304 samples covering a wide range of scenarios”, and the Hermes Motion Testset “comprises 96 prompts, encompassing 32 distinct types of sports activities”. A vendor-built benchmark evaluated by vendor-recruited annotators is a legitimate way to show what a model is good at. It is not third-party verification, and as of 2026-08-24 we have not found an independent lab reproducing these results.

The honest summary: every performance claim about Waver 1.0 in circulation traces back to ByteDance’s own August 2025 materials. Treat “better than Veo 3” as a vendor claim from a year ago, not a current finding.

Inside Waver 1.0: 12B Parameters and a Separate Refiner

The README is unusually detailed about how the model is built, which is the main thing Waver offers that Seedance does not.

Components. Waver uses Wan-VAE — the VAE shipped with Alibaba’s Wan 2.1 — to compress video into latents, and takes text features from two encoders, flan-t5-xxl and Qwen2.5-32B-Instruct. Building a ByteDance model on an Alibaba VAE and an Alibaba LLM is a genuinely unusual detail, and it is stated plainly in the official README.

Architecture. The DiT is “built upon rectified flow Transformers”, with video and text fused through what the README calls “a Dual Stream + Single Stream approach” and the paper names a “Hybrid Stream DiT architecture”. Text-to-video and image-to-video are trained jointly in one model by changing the input channels — 16 for the video token, 16 for the first-frame image token and 4 for a task mask — with the image latent mixed in 20% of the time during joint training.

The refiner. This is the part most summaries get wrong. Waver does not generate 1080p directly. A separate model, Waver-Refiner, upsamples a 480p or 720p video to 1080p, adds noise, and denoises it back. The README gives the payoff in speed: inference time drops “~40% for 720p→1080p” and “~60% for 480p→1080p” compared with generating 1080p directly. So when a listing advertises “Waver 1.0, 1080p”, the honest reading is 720p generation plus an upscale pass.

Training recipe. ByteDance says it “invested substantial computational resources in training on 192p videos” before stepping up to 480p and then 720p, on the finding that low-resolution training is what teaches motion. Inference specifics are published too: sigma shift 3.0 in training and 7.0 at inference for 720p, and an adapted APG guidance scheme with a normalisation threshold of 27 and guidance scale 8.

None of this has a Seedance equivalent. ByteDance has published no architecture diagram, parameter count or training recipe for Seedance — a gap we cover in is Seedance open source. If you came looking for ByteDance video-model internals, Waver’s README is the closest thing that exists.

Can You Actually Use Waver 1.0 Today?

Not from ByteDance, and not by downloading it.

The official repository ships nothing runnable. As of 2026-08-24, FoundationVision/Waver contains exactly two entries at root: README.md and assets/. There is no inference script, no weights, no configuration and no licence file. The repository has 951 stars and has not received a code push since 2025-08-27. Whatever “open source” means in the FoundationVision org bio, for Waver it currently means an open technical report.

ByteDance’s own access route is a Discord. The only try-it link in the README points to opensource.bytedance.com/discord/invite. There is no Waver web app, no Waver tier inside Dreamina, and no Waver model ID on ModelArk.

The project page did not load for us. The README links to www.waver.video as the project page. On 2026-08-24 we could not reach it from our test environment — our DNS resolved the hostname to a reserved-range address, which points at our own network path rather than at the site’s status. We are not going to report a site as down on the strength of a local resolver quirk.

Third-party hosts run it. WaveSpeedAI lists a model page titled “ByteDance Waver 1.0” describing it as “an all-in-one model for text-to-video, image-to-video, and text-to-image generation”, priced at “$0.30 per run” with the note that “the final charge scales with the parameters you set”, and reporting a median end-to-end generation time of about 123 seconds per request (third-party page, read 2026-08-24). That is a third-party channel price on a third-party endpoint — ByteDance publishes no Waver price of any kind, and we have not independently verified that the endpoint serves the model described in the paper.

Watch the domain. The official repository links to exactly three places: arxiv.org/pdf/2508.15761, www.waver.video, and the ByteDance Discord invite. Several other “Waver” domains rank for the search term and none of them is referenced by the official repository. The same pattern-matching we describe in official site vs the clones applies here, and it applies harder: with no official product page to anchor to, a lookalike has almost nothing to be caught against.

How to Fact-Check a Waver 1.0 Page in Two Minutes

Because there is no official product page to compare against, the fastest defence against a made-up Waver spec sheet is a short list of checks you can run yourself.

Check 1: does the page cite the arXiv report or only itself? The one document everything traces back to is arXiv 2508.15761. A page that gives numbers with no route back to that report, the GitHub README or a named third-party host is repeating something rather than sourcing it.

Check 2: does it say 720p or 1080p, and does it say which is generated? Any page presenting 1080p as native generation has skipped the refiner stage the README describes in detail. That is the single most common error we saw while researching this article.

Check 3: does it quote a leaderboard result without the date stamp? ByteDance’s own claim carries “data up to 2025-08-05”. A “Top 3” without a date is a year-old snapshot presented as current standing.

Check 4: does it name a price as ByteDance’s price? ByteDance has published no Waver price. Every figure you see belongs to a third-party host and applies only to that host’s endpoint.

Check 5: does it describe a subscription, credit pack or plan tier? None exists. That is the clearest signal that a page has confused Waver with a productised model — usually Seedance, whose credit mechanics we break down in the Seedance pricing review.

Why ByteDance Has Two Video Model Lines

This is the question behind the search, so here is the structural answer.

ByteDance runs several research groups that publish generative models, and a product organisation that ships them. Waver came out of FoundationVision, the group behind VAR, LlamaGen and Infinity — an org whose output is papers and repositories. Seedance comes out of the Seed line, whose models get product pages on seed.bytedance.com, a consumer surface in Dreamina, and billable IDs in the BytePlus ModelArk catalogue. On our 2026-08-24 reads, the Seed research index mentions Seedance 41 times and Seedream 26 times, and Waver zero times.

So the two are not competing versions and there is no upgrade path between them. They are two different outputs of the same company: Waver is what ByteDance published, Seedance is what ByteDance sells. A useful sanity check for anything you read about Waver: if a page describes a Waver subscription, a Waver credit system or a Waver tier list, it is describing something ByteDance has never announced.

The image-side parallel is worth knowing. Seedream is the Seed line’s image model, and it sits in the same catalogue as Seedance — that is a sibling relationship in a way Waver is not, despite Waver also doing text-to-image.

Which One You Should Care About

If you want to make video today, Waver 1.0 is not an option worth planning around. There is no official access, no published price, no weights to self-host and no update in a year. Seedance is the line with published rate cards, model IDs and a consumer app — see our Seedance API guide for the current model IDs and the Seedance pricing review for what each path costs.

If you are researching how ByteDance builds video models, Waver 1.0 is the more useful document by a wide margin. The parameter table, the hybrid-stream design, the cascade refiner and the resolution-laddered training recipe are all public. None of that exists for Seedance.

If you are evaluating a third-party platform that offers “Waver 1.0”, the questions to ask are whether it states the output resolution and whether that resolution is generated or refined, since the official model generates at 720p and upscales. Compare the per-run price against the per-second Seedance figures rather than assuming a research model is cheaper.

What We Could Not Verify (2026-08-24)

Stating the gaps is part of the answer here, because Waver’s information environment is thin enough that confident-sounding pages are usually filling them in.

• Whether Waver 1.0 is still being developed. No official statement either way; the repository has been static since 2025-08-27 and there is no announced successor.
• What the head-to-head chart actually says. The comparison figures are images in the repository; we quote only the README’s prose, which names no competitor.
• Any independent benchmark. We found no third-party lab reproduction of ByteDance’s Waver results.
• The project page’s current contents. Our fetch of www.waver.video failed on a local network path, not on the site’s side, so we make no claim about it.
• Whether third-party endpoints serve the paper’s model. Hosts advertise “ByteDance Waver 1.0”; we cannot confirm the weights behind those endpoints.
• Waver’s licensing. No licence file exists in the repository, so the terms of any use are undefined in official materials.

FAQ

Is Waver 1.0 real, or is it a fake ByteDance product?

It is real. Waver 1.0 has a technical report on arXiv (2508.15761, “Waver: Wave Your Way to Lifelike Video Generation”, submitted August 21, 2025) and an official repository at github.com/FoundationVision/Waver, under a GitHub organisation whose own bio reads “Bytedance’s opensource FoundationVision models”. What it does not have is a place on ByteDance’s product surfaces: on our 2026-08-24 check, “Waver” appears zero times on the Seed model page, the Seed research index, the Seed blog index and the BytePlus ModelArk model list, all of which list Seedance.

Is Waver 1.0 the same thing as Seedance?

No. They are different models with different lineages. Seedance is ByteDance’s productised video line — listed on the Seed model page, sold through Dreamina, billed through ModelArk. Waver 1.0 is a research release from ByteDance’s FoundationVision group with a paper and a README, no published weights and no ByteDance product page as of 2026-08-24.

Can I download Waver 1.0 weights?

Not from the official repository. As of 2026-08-24 it contains only README.md and an assets folder — no inference code, no weights, no licence file — and its last code push was 2025-08-27. Third-party hosts run an endpoint labelled “ByteDance Waver 1.0”, which is hosted inference, not a weight download.

What resolution and clip length does Waver 1.0 support?

The two official sources differ slightly, so here are both. The README says the model “Supports up to 1080p high-resolution image and video generation, with flexible resolution, aspect ratio, and video length (2-10 seconds)”. The arXiv abstract says Waver “can directly generate videos with durations ranging from 5 to 10 seconds at a native resolution of 720p, which are subsequently upscaled to 1080p”. Both agree 1080p arrives via the refiner rather than direct generation.

Does Waver 1.0 really beat Veo 3 and Kling?

That claim is ByteDance’s own, and the official wording is narrower than the circulated version. The README says only that Waver ranks “Top 3” on the Artificial Analysis text-to-video and image-to-video leaderboards with “data up to 2025-08-05”, and that in ByteDance’s manual evaluation it was “superior to those from both closed-source and open-source models” — naming no competitor. Named head-to-head match-ups come from third-party articles and a repository chart image, not from an independent benchmark.

Which ByteDance model should I use for video in 2026?

Seedance, because it is the only one you can actually buy. Which Seedance version is a separate question — start from our version timeline.

Related Content

• Seedance Version Timeline — which Seedance versions are live on which official surface.
• Is Seedance Open Source? — what ByteDance has and has not published about the Seedance line.
• Seedance API Guide — current model IDs and rate cards.
• Where to Access Seedance 2.0 — every official and third-party channel.
• Official Site vs the Clones — how to spot a lookalike domain.

SeedanceTips is an independent publication. We are not affiliated with ByteDance, and this article contains no affiliate links. Every official claim above is quoted from a page you can open yourself, with the date we read it.