Seedance 2.5 ships in ComfyUI with 30-second runs and timeline shot control

• ComfyUI got the node-graph surface: ComfyUI's Partner Nodes post says Seedance 2.5 supports 30-second one-pass runs, up to 50 references, timeline shot control, editing, and native dialogue with lip sync.
• The launch spread across creator and API surfaces fast: fal exposed text, image, and reference routes in fal's endpoint links, while Flowith, Higgsfield, Pika, and Venice all announced access in the same wave.
• The big spec jump is native duration plus reference budget: ComfyUI's feature list repeats 30 seconds in one pass and 50 reference assets, matching ByteDance's official 30-image, 10-video, 10-audio framing.
• Prompting moved toward shot sheets: techhalla's prompt thread said 2.0-style prompting produced average results, and his examples used second-by-second timelines, camera notes, reference locks, and audio cues.
• Pricing is already part of the rollout: pika_labs' price post claimed Seedance 2.5 API access at up to 88% below other aggregators, while fal's docs price a 30-second 720p 16:9 run at about $13.87.

The Seed Team launch post says Seedance 2.5 doubled native single-pass generation from 15 to 30 seconds and added multi-round extensions. Fal's Reference to Video docs expose the implementation shape: [Image1], [Video1], [Audio1], billed input-video seconds, and a token formula. The Comfy page is mostly surreal prompt carpentry; the BytePlus rollout note buries the enterprise constraints, including C2PA credentials, visible watermarks, real-face controls, and no BytePlus availability in the United States.

ComfyUI's rollout is the most useful one for builders because it turns Seedance 2.5 into node-graph verbs. Christmas came early for video workflow nerds.

The launch post lists the operating surface:

• 30-second generation in one pass.
• Up to 50 reference assets per run: 30 images, 10 videos, 10 audio files.
• Second-level shot control, where the prompt is written as a timeline.
• Plain-language editing or extension of existing footage.
• Native dialogue and lip sync in more than 10 languages.

ComfyUI repeated the same controls in a shorter feature list, then pushed users to the live model through its access reply. The public Comfy landing page lists Image to Video and Text to Video partner nodes, pay-as-you-go credits, no watermark, and commercial use included.

Fal's demo posts gave the cleanest mental model for the new duration: tiny visual input, long camera plan, no keyframes.

The official Seed Team says Seedance 2.5 can generate 30-second audio-video clips in a single pass and then extend them over multiple rounds. Fal's Text to Video docs describe the model as reasoning over the whole shot so motion, lighting, and subject identity hold from first frame to last.

The key API parameter is boring and important: duration accepts auto or a fixed string from "4" through "30". The docs also warn that long generations should use the queue API rather than a synchronous call.

ByteDance's official count is more specific than the launch chatter: up to 30 images, 10 video clips, and 10 audio clips in one pass. That is the reference budget behind the simpler “50 references” shorthand.

Fal's Reference to Video docs map those assets into arrays named image_urls, video_urls, and audio_urls, then address them positionally in prompts as [Image1], [Video1], and [Audio1].

The evidence pool shows three reference-heavy use cases:

• Visual environment swaps, where pika_labs' editing example says one image reference can alter the landscape and main subject of a complex sequence.
• Graphic and typography integration, where pika_labs' graphics post says photographic, iconographic, typographic, and graphic references can be folded into a 30-second video.
• Identity restyling, where higgsfield_ai's restyle teaser says one reference can hold a subject through broad art-direction changes.

The prompt pattern that spread was a shot sheet.

The common structure in techhalla's prompt examples is easy to scan:

1. Style, camera, and atmosphere.
2. Image reference lock.
3. Timeline second by second.
4. Quality boosters.

Techhalla later told a commenter that timeline prompting now has twice the room and each section needs more detail in a timeline reply. His full wingsuit prompt used a single strict character reference, audio notes, and a 0-30s timeline covering launch, descent, near misses, landing, and a hot-dog beat in the wingsuit example.

Seedance 2.5's editing story is not just rerolling a whole clip. ByteDance says the model supports timestamp-level edits to characters, actions, and plot elements while keeping continuity before and after the edit.

The launch examples split into four practical routes:

• Animatic to render: higgsfield_ai showed hand-drawn timing and motion turning into finished animation.
• 3D blocking to final scene: higgsfield_ai showed rough Blender motion, physics, and layout becoming a textured cinematic shot.
• Video-to-video cleanup: higgsfield_ai's banana post framed 2.5 as frame-perfect video-to-video.
• Plate edits: higgsfield_ai's VFX insert showed adding or removing background elements without touching the plate.

One hybrid workflow used H3 for low-resolution scene adherence, then passed outputs and references into Seedance for lighting, camera angle, and cinematic polish, according to techhalla's H3 workflow reply.

The rollout looked less like a single launch button and more like a distribution scramble.

• fal: fal's endpoint links point to Text to Video, Image to Video, and Reference to Video routes.
• ComfyUI: ComfyUI's Partner Nodes post announced live access via Partner Nodes.
• Flowith: flowith's launch post promised 30-second videos and up to 50 references, while another flowith post claimed stronger character and visual consistency. Earlier Flowith teasers covered no-cut generation in a cinematic post, first-person POV in one POV teaser and another POV teaser, and animated shorts in its animated-shorts post.
• Higgsfield: higgsfield_ai's first promo offered 7 days of unlimited access, then higgsfield_ai's later post advertised 33 days of unlimited Seedance 2.5.
• Pika API Club: pika_labs' price post claimed access at up to 88% below other API aggregators, and pika_labs' partner note thanked BytePlusGlobal.
• Venice: AskVenice's teaser said Seedance 2.5 was coming, AskVenice's comparison post put 2.0 and 2.5 side by side anonymously, and AskVenice's access post opened creation.

Dreamina's official product page says the model is live globally on Dreamina, supports up to 30 seconds in standard mode, and extends to 180 seconds in beta long-video mode.

Fal prices Seedance 2.5 by output tokens for text-to-video and image-to-video: height * width * duration_seconds * 24 / 1024, at $0.0214 per 1,000 tokens. Its docs translate that into about $0.4730 per second for 720p 16:9 with audio, about $0.2205 per second for 480p, and about $13.87 for a 30-second 720p 16:9 generation.

Reference-to-video has the migration gotcha. Fal's docs bill input video duration along with output duration, apply a 0.6 multiplier when video references are used, and do not bill image or audio references.

Pika's API Club page says membership is $10 per month, new members get a $10 first-month credit, and usage stays pay-as-you-go. The same page says its API is built to be legible to agents and includes an onboarding prompt for Claude Code, Codex, Cursor, Lovable, Replit, and ChatGPT.

BytePlus is the official enterprise route, but its launch post adds limits that consumer demos rarely mention.

The BytePlus announcement says Dreamina Seedance 2.5 includes visible watermarks, C2PA Content Credentials, IP protection mechanisms, and technical controls restricting video generation from images or videos containing real faces. It also says vetted clients can use real-person verification and likeness authorization through ModelArk, or choose from more than 10,000 virtual human assets.

BytePlus says ModelArk access is available in supported markets globally, with one explicit caveat: BytePlus is not available in the United States. Fal filled part of that gap for enterprise buyers, saying in its Seedance enterprise post that select U.S. organizations can access Dreamina Seedance and Seedream models with U.S. hosting on fal's enterprise platform.