AI movie models: Dreamina Seedance 2.5 vs Alibaba Wan 3.0

AI movie models, episode 6. Previously: Kling, Gemini Omni vs Utopai X, xAI Grok Imagine, Runway Gen-4.5 vs Luma Ray3.2 and MiniMax H3 vs Tencent HunyuanVideo 1.5.

This series explains one or two AI video models per post, using the same twelve questions every time. For how a clip is made from a prompt, see how AI video generation works. Facts below come from the vendor’s own pages unless marked as a user report or our opinion. Prices and leaderboard ranks were checked on 8 November 2026. We did not generate clips, and this post does not embed vendor videos.

Which versions. Dreamina Seedance 2.5 is ByteDance’s joint audio-video model, announced on 31 July 2026 for products such as Dreamina. On BytePlus ModelArk the current id is dreamina-seedance-2-5-260628. It follows Seedance 2.0. Alibaba’s Wan 3.0 is the API model wan3.0-video on Model Studio, with a faster sibling, wan3.0-video-prime. An Alibaba Cloud Community post dated 27 August 2026 says general availability began on 24 August. The API reference, last updated 15 October 2026 and read 8 November, still says “Currently in preview.” Both lines are on Alibaba’s own pages. Neither vendor publishes downloadable weights for these two models on the pages we read.

What each does differently. ByteDance pitches Seedance 2.5 on longer clips, up to 50 references in one request, and editing inside a clip. Alibaba pitches Wan 3.0 as one model for text, frames, references, documents, editing and extension, with a published price per second of output. Those are the vendors’ claims. The practical split on the pages we read is references and image-to-video rank on one side, and a simpler per-second price plus the text-to-video rank on the other.

Dreamina Seedance 2.5 against the checklist

1. Resolution. The ModelArk tutorial, updated 2 November 2026, lists 480p at 8-bit, 720p at 8-bit, and 1080p at 10-bit. The price table marks 4K “Not supported” for dreamina-seedance-2-5-260628. A 4K path for selected enterprise customers, with a 480p draft before a 4K render, is the limited trial we covered in Seedance 2.5 4K drafts. The public price table still does not list 4K.

2. Aspect ratios. 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, and adaptive. First-frame and first-and-last-frame jobs must use adaptive, and the tutorial says the output follows the first frame.

3. Reference types. Text, a first frame, a last frame, reference images, reference video, and reference audio. Audio can be a reference on its own. The same inputs also cover video editing and video extension when the prompt says so. The Seed product page adds vendor claims for white-model control, green-screen editing, camera movement and performance blocking. A style reference as its own input type is not documented. The model is described as joint audio-video generation. A separate lip-sync specification was not on the pages we read. The tutorial’s example prompts do include spoken lines.

4. References at once. Up to 50 assets: 30 images, 10 videos and 10 audio clips. The tutorial says the maximum total duration of reference audio or video assets rose from 15 seconds on the previous series to 30 seconds. For an edit, the reference video must be 4 to 30 seconds long. First-and-last-frame mode uses those two frames, and the tutorial treats that mode as separate from the omni-reference pile.

5. Price (checked 8 November). BytePlus bills successful videos only, in tokens. Online, without video input, Seedance 2.5 is $10.70 per million tokens at 480p and 720p, and $11.70 at 1080p. With video input the rates are $6.40 and $7.00. The lower rate is per token. Tokens still count input video time, so a clip with video input can cost more, which the examples below show. 4K and offline inference are marked “Not supported.”

For a 16:9 clip with no video input, BytePlus’s own examples are $0.514 for 5 seconds at 480p ($0.103 a second), $1.156 at 720p ($0.231 a second), and $2.843 at 1080p ($0.569 a second). Twice those 5-second examples, our arithmetic, is about $1.03, $2.31 and $5.69 for 10 seconds. BytePlus says the token formula is an estimate and the bill uses usage.completion_tokens.

With video input, the same page’s 5-second 16:9 examples run from $0.553 to $2.152 at 480p, $1.244 to $4.838 at 720p, and $3.062 to $11.907 at 1080p, depending on input length. A minimum token charge applies when the input includes video. We did not reprint that minimum table.

Draft mode on this public API makes a 480p video at the 480p token rate. That is separate from the enterprise 4K draft flow. A free tier for Seedance 2.5 was not on the price page. Dreamina’s consumer plan prices were not read on 8 November. Frame rate for 2.5 was not on the spec table.

6. Performance. A generation time in seconds is not published. The tutorial only says generation takes some time and shows a poll every 10 or 30 seconds. On Artificial Analysis’s text-to-video board with audio (AA-Video-T2V v2.1, read 8 November), Seedance 2.5 is 5th of 32, Elo 1155 plus or minus 13, rank range 3 to 7, 7,436 samples, listed as released July 2026. Wan 3.0 is 1st, Elo 1172 plus or minus 13, range 1 to 3. The 95 percent intervals overlap (1142 to 1168, and 1159 to 1185). On the image-to-video board with audio, Seedance 2.5 is 1st, Elo 1158 plus or minus 15, range 1 to 2, 6,297 samples. How to read a gap like that is the subject of what a leaderboard leaves off the podium. The board’s FAQ for the silent image-to-video list places Seedance 2.5 second, Elo 1131. Its silent text-to-video top five does not include Seedance 2.5, so this episode does not invent that rank. Artificial Analysis lists Seedance at $34.12 a minute at the settings it benchmarks. That sits next to BytePlus’s 1080p example of $0.569 a second. We use BytePlus for the price.

7. Character consistency. The method in the docs is reference images, up to 30, with prompts that point at “Image 1.” ByteDance’s launch post says 30-second clips make “stronger character consistency” easier. That is ByteDance’s claim. A separate identity-lock control is not documented.

8. Scene consistency. One generation runs 4 to 30 seconds, or the model may pick a length. The Seed page says you can extend twice. The tutorial says extension works over multiple rounds, forward or backward, and keeps the chosen video’s aspect ratio. An edit keeps the selected video’s aspect ratio and duration. The long tutorial example is one continuous shot with timed beats. A named multi-shot storyboard mode is not in the spec table.

9. Compared with the previous model. Against Seedance 2.0, the tutorial’s table shows duration rising from 4 to 15 seconds to 4 to 30, references rising from 15 (9 images, 3 videos, 3 audio) to 50, audio allowed on its own, draft mode added, and 1080p moving to 10-bit. Seedance 2.0’s public price table includes 4K. Seedance 2.5’s does not. ByteDance’s July post says the new model builds on 2.0 with longer clips, more references and timestamp editing.

10. What users think (user reports, not facts). Reddit returned a block page on 8 November, so these lines are from X posts we opened.

• Praise, pointed at Wan and useful here as a comparison: @Martinolli wrote on 15 October that Wan 3.0’s “visual quality and realism can be excellent, sometimes even more photorealistic than Seedance 2.5.”
• Complaint, same post: Wan is “still weaker for complex camera choreography.” The next clause was cut off in the text we received, so we do not quote it.
• @ImaStudio_ai wrote on 3 November that “the clarity gap is smaller than the price gap” and that quality across MiniMax H3, Seedance 2.5 and Wan 3.0 is “getting surprisingly close.” Dollar figures in that post are that account’s. They do not match the vendor pages above, so they are not repeated here as prices.

11. General quality. We have not generated Seedance 2.5 clips for this series. ByteDance’s launch post and the Seed page are the official galleries. The launch post claims smoother motion and more realistic visuals. We did not score hands, faces, physics or on-screen text ourselves.

12. Our verdict (opinion). On documented features, Seedance 2.5 is a candidate for many references and 10-bit 1080p; it leads Artificial Analysis’s image-to-video board with audio as read on 8 November, and BytePlus’s no-video-input example is about $2.84 for 5 seconds at 1080p. This is not confirmation of commercial-use or output rights. Skip the public API if you need 4K on a price list, open weights, or a low per-second rate. The United States is absent from BytePlus’s availability list, last updated 21 April 2026. That list predates Seedance 2.5’s launch and defers to availability at purchase; it establishes neither a current US ban nor current global availability.

Alibaba Wan 3.0 against the checklist

1. Resolution. 480p, 720p and 1080p. The usage guide says 1080p is the default. 4K is not in the resolution list.

2. Aspect ratios. The usage guide’s output section lists 16:9, 4:3, 1:1, 3:4, 9:16, and adaptive, which is the default and follows the input. The text-to-video parameter line on the same guide also lists 21:9. First-frame jobs are told to use adaptive.

3. Reference types. Text, first frame, last frame, reference images, reference video, reference audio, one file, or one web link. File and link cannot be used together, and frame mode cannot be combined with reference, file or link inputs. The usage guide says audio is on by default and can be turned off. It says the model natively outputs dialogue, background music and sound effects, and its text-to-video section calls that synchronized. An example prompt asks for lip sync. That is a prompt in the guide, and the 27 August post calls identity, voice and lip sync “first-class claims” of the release. The same post says audio texture and on-screen text are still maturing.

4. References at once. The API tables cap reference images at 10, reference videos at 5 clips whose durations add up to at most 15 seconds, and reference audio at 5 clips with the same 15-second total. One file or one link can sit with those references. The usage guide also says “up to 20” multimodal materials per request. When a video is passed in, the guide says input plus output duration stays at or under 30 seconds.

5. Price (checked 8 November). The wan3.0-video model page, last updated 11 October, bills per second of successfully generated video. It says the figures are original prices, excluding limited-time promotions, and points at the console for offers. A free-tier number was not on that page.

Singapore, marked international: $0.05 a second at 480p, $0.10 at 720p, $0.20 at 1080p. A 5-second clip is $0.25, $0.50 or $1.00, and a 10-second clip is $0.50, $1.00 or $2.00 (our arithmetic). A 30-second 1080p clip at the Singapore rate is $6.00.

Beijing, Tokyo, Frankfurt, US (Virginia) and Hong Kong: $0.041256, $0.082513 and $0.165025 a second. At the Virginia 1080p rate, 5 seconds is about $0.83 and 10 seconds about $1.65 (our arithmetic). The August community posts use the Singapore figures. Artificial Analysis lists $12.00 a minute, which matches $0.20 times 60, the Singapore 1080p rate, not the Virginia rate.

The faster wan3.0-video-prime page, same date, lists higher original prices: Singapore $0.068, $0.14 and $0.28 a second, and $0.063600, $0.127199 and $0.254399 in the other regions above. The API reference says prime is aligned with the standard model and faster end to end. It does not give a second count. The model page lists 300 requests a minute in each of those regions. Output URLs last 24 hours.

6. Performance. The API reference says a task typically takes 1 to 5 minutes, and the polling note says several minutes. Output is 30 frames a second, MP4. On the text-to-video board cited above, Wan 3.0 leads at Elo 1172, range 1 to 3, 7,307 samples, listed as released August 2026. On image-to-video it is 10th of 33, Elo 1109 plus or minus 15, range 7 to 12, 6,389 samples. Its image-to-video interval (1094 to 1124) does not overlap Seedance 2.5’s (1143 to 1173). The silent text-to-video FAQ places Wan 3.0 fifth, Elo 1132. Wan is not in that page’s silent image-to-video top five, so this episode does not invent that rank.

7. Character consistency. The documented method is reference images, up to 10, with prompts that point at “Image 1.” The 13 August post claims fine-grained consistency for faces, hair, clothes and accessories in reference-to-video. That is Alibaba’s claim. A separate identity-lock feature is not named.

8. Scene consistency. Clips run 2 to 30 seconds, or smart duration (-1), which the guide says defaults to 5 seconds. The text-to-video section says the model supports single-shot and multi-shot storyboards, 4 to 6 seconds a shot, with timestamps. Extension can run forward, backward, or both, and the guide says adaptive ratio keeps the source aspect. The August post also claims stable space and style in reference-to-video.

9. Compared with the previous model. The 27 August post says Wan 2.7 ran 2 to 15 seconds, and that editing from Wan 2.7 carries forward. The 13 August post says document input is new for the family, and that faces and reference consistency are the quality jump. The API FAQ says Wan 2.7’s first-frame mode could take driving audio, and Wan 3.0’s first-frame mode cannot. Audio-driven work on Wan 3.0 uses reference image plus reference audio instead.

10. What users think (user reports, not facts). The same two posts as above.

• Praise: @Martinolli said Wan’s realism can be excellent, sometimes more photorealistic than Seedance 2.5.
• Complaints: the same post says Wan is weaker for complex camera choreography. @ImaStudio_ai treats the visible quality gap as smaller than the price gap, which is that account’s opinion, not a measurement we repeated.

11. General quality. We have not generated Wan 3.0 clips. Alibaba’s 13 August post is the official write-up, and it says audio texture and on-screen text rendering are still improving. Motion, physics, hands and faces are not scores we made.

12. Our verdict (opinion). On documented features and pricing, Wan 3.0 is a candidate for document input and per-second billing; a Virginia price is listed and it leads Artificial Analysis’s text-to-video board with audio as read on 8 November. Its conflicting preview and general-availability statements leave applicable contract status unresolved; we have not verified production permission, commercial-use rights or output ownership. The pages read offered no downloadable weights. Prime is the faster SKU at a higher list price, not a different feature set on the page we read.

Rights and licensing

Seedance 2.5. On 8 November the bodies of BytePlus’s ModelArk terms and of the video-generation specific terms did not load past the navigation, so this episode does not state who owns the output. Commercial-use permission is also unverified. ByteDance’s July launch post says Dreamina adds invisible watermarks, C2PA Content Credentials and visible AI labels, and that it blocks hate, nudity, non-consensual sexual content, unauthorized likeness and harmful misinformation. That is a product-safeguard claim, not a copyright assignment. The international availability page, last updated 21 April 2026, lists many countries and does not list the United States. Its disclaimer says availability is decided at purchase. The April list predates this model’s launch and is not evidence of a current US ban or sanctions restriction. The July launch post said the first subscriber rollout, for accounts of people over 16, was Europe, Asia, the Middle East and South America. Weights are not published on these pages.

Wan 3.0. The pages read offered no downloadable Wan 3.0 weights, and the model page lists fine-tuning as unsupported. The API reference still says the model is in preview. Alibaba’s general preview-product terms, updated 20 August 2026, limit a preview product to internal testing, research and evaluation, and they keep ownership of the product itself with Alibaba Cloud. Wan 3.0 is named in that page’s model navigation, not inside the licence clauses, and a 27 August community post says general availability. We do not collapse those into one status. We have not established which terms govern a purchase or its output today; the generic preview clauses are not confirmed to govern Wan 3.0 output. Commercial-use permission and output ownership remain unverified. A clause that assigns copyright in generated video was not on the model page or in those preview terms. The usage guide says a watermark is added only when the caller sets watermark to true. Output links expire after 24 hours.

Summary

#	Checklist	Dreamina Seedance 2.5	Alibaba Wan 3.0
1	Resolution	480p, 720p, 1080p (10-bit). 4K not supported on the public price table	480p, 720p, 1080p. 4K not listed
2	Aspect ratios	21:9, 16:9, 4:3, 1:1, 3:4, 9:16, adaptive	16:9, 4:3, 1:1, 3:4, 9:16, adaptive. Text-to-video line also lists 21:9
3	Reference types	frames, images, video, audio, edit, extend	frames, images, video, audio, one file or link, edit, extend
4	References at once	30 images, 10 videos, 10 audio, 50 total	10 images, 5 videos, 5 audio. Guide also says up to 20 materials
5	Price (8 Nov)	about $2.84 per 5 s at 1080p, no video input (BytePlus example)	$1.00 per 5 s at 1080p in Singapore. About $0.83 in Virginia
6	Performance (AA boards with audio, read 8 Nov; vendor time)	T2V 5th, Elo 1155, range 3 to 7. I2V 1st, Elo 1158. Time not published	T2V 1st, Elo 1172, range 1 to 3. I2V 10th, Elo 1109. Typically 1 to 5 min (Alibaba)
7	Character consistency	reference images, up to 30 (ByteDance claim on consistency)	reference images, up to 10 (Alibaba claim on consistency)
8	Scene consistency	4 to 30 s. Extend (twice on the Seed page, multiple rounds in the tutorial)	2 to 30 s, multi-shot storyboard, extend
9	vs previous	longer than 2.0, more references, no public 4K	longer than Wan 2.7, documents added, first-frame audio driver dropped
10	Users	labelled reports only: realism and camera control compared with Wan	same posts: realism praised, camera choreography criticised
11	Quality	not tested by us	not tested by us. Alibaba says text and audio texture are still maturing
12	Verdict (opinion)	references and 8 Nov AA image-to-video rank with audio, at a high 1080p price	published per-second price and 8 Nov AA text-to-video lead with audio, preview label still on the API page

For builders comparing documented API features and pricing, Seedance 2.5 offers token billing and a long reference list, with 4K absent from its public price table. Wan 3.0 lists per-second prices by region, while its preview label and general-availability post conflict. On Artificial Analysis’s text-to-video board with audio read 8 November, Wan ranks first and Seedance fifth, but their 95 percent intervals overlap. These are feature and benchmark comparisons, not verified contracts or clearance for production or commercial output use.