AI video12 minAI video tools are shifting from one-shot clips to controllable motion and editable workflows. That is the difference between a toy and a production primitive.Read now →

Seedance 2.5 Just Went 4K. Here Is the Five-Hour, One-Person Workflow for a Full AI Commercial
ByteDance's Seedance 2.5 now renders native 4K, and a Christmas supermarket heist short shows what one person can ship in five hours. The trick isn't the video model. It's handing pre-production to Claude, wiring it to a generator over MCP, and locking the camera before you prompt a single frame.
Every Christmas Eve, an older couple named Brenda and Graeme get the last bronze turkey at Aldi. Six years running. This year, four friends decide it's their turn.
What follows is a tight little heist in four phases. Danny flirts shamelessly with Brenda ("Didn't know Santa delivered this early"). Mike traps Graeme in a revolving door, lap after lap. Sophie ambushes them at the free-samples table "by any means necessary." And while the couple is distracted, the ringleader slips in and bags the bird. "You're lucky," says the worker. "That's the last one."
It plays like a polished festive TV spot. It was made by one person, in about five hours, entirely with AI, and every shot was generated in native 4K.
The 4K part is new this week. The five-hour part is the more interesting story, because the speed doesn't come from the video model. It comes from what happens before anyone types a video prompt.
What actually changed this week
Seedance is ByteDance's flagship video generation model. Version 2.5 launched on July 31, 2026, building on the joint audio-video architecture of Seedance 2.0 and adding longer narratives, more reference inputs and video editing.
The news in November is resolution. Here's the timeline:
- October 27: BytePlus, ByteDance's enterprise cloud arm, opened the 4K edition to business customers only, according to reporting by The Inference.
- November 11: Four creative apps switched on Seedance 2.5 4K within roughly 90 minutes of each other: Magnific, Picsart, Pika and Krea.
- November 13: ByteDance's own creative platform, Dreamina (Jimeng in China), officially launched Seedance 2.5 4K.
The key specs, per Picsart's launch post and Chinese launch coverage:
- Native 4K output in 10-bit color, so there's no separate upscaling pass guessing at detail.
- Up to 30 seconds per shot as one continuous take, double Seedance 2.0.
- Up to 50 references (image, video and audio) to pin down characters, wardrobe, voices and locations, up from 15.
- Lip-synced dialogue in 20 languages.
- 480p Draft Mode that carries the prompt, references and seed into the final render, so the 4K version keeps the framing you approved. ByteDance estimates a session of about 20 previews plus one final render costs up to 77% less than 20 full-resolution takes, at nearly twice the speed.
That draft-then-final logic matters for everything below. 4K is expensive. The workflow is built to make sure you only pay for 4K once you already know the shot is right.
The real bottleneck was never generation
Anyone who has tried to make a multi-shot AI film knows the failure mode. Shot one looks amazing. Shot two has a different color grade. Shot three's lens feels like a phone. By shot six, the "film" is a slideshow of unrelated clips with the same characters pasted in.
The fix professional crews use is pre-production: mood boards, shot lists, character sheets, a locked camera package, music chosen before the edit. That work usually takes days.
The core idea of this workflow is simple: let Claude do pre-production, and make it do it like an art director.
Step 1: A pre-production board that Claude builds
The hub of the project is a Milanote board. Milanote is a visual mood-board tool where you can drop images, text, links and video onto a canvas.
On this board, Claude assembled:
- Character sheets for every role: four main friends plus the two elderly "rivals," designed ahead of generation.
- Three music references, each with suggested timestamps, so the editor knows exactly which section of which track could carry which moment.
- Visual references pulled from ShotDeck, covering composition, color grade and art direction.
- A scene-by-scene breakdown with timecodes. For example, the 0-to-2-second beat gets a composition reference, shot notes, the spoken dialogue and a "look and camera" note.
- All dialogue and voiceover, broken down per scene.
One honest caveat: Milanote does not ship an official MCP server or Claude connector. Connecting Claude to a Milanote board currently runs through third-party MCP bridges, so check what you're granting access to before you wire it up.
Why ShotDeck is the secret weapon
ShotDeck is a searchable library of high-resolution stills from films, TV, music videos and commercials. What makes it useful for AI work is the metadata. Each frame is tagged with camera and lens details, lighting, composition, location, aspect ratio and color, including the actual swatches and HEX values of the palette.
That turns a vague vibe ("make it look cinematic") into concrete, repeatable data:
- The exact HEX palette of a frame you love, which you can feed into prompts and later into your grade.
- The camera body, such as an ARRI Alexa Mini LF.
- The lens family, such as Panavision Panaspeed primes.
In this project, the camera package was set once (an Alexa Mini LF look with Panavision glass) and carried through every shot.
Lock the camera, or lose the film
This is the step most AI filmmakers skip, and it's the one that matters most. Continuity comes from locking the camera style for the whole project: the same body, lens character, palette and grade language in every prompt.
The workflow treats "steal like an artist" with some care. Lifting a whole shot is copying. Borrowing a lens choice, a palette or a composition principle is how cinematographers have always learned. The point is to pull ingredients, not finished dishes.
Step 2: Wire Claude to a generator with MCP
Here's where pre-production stops being a document and starts producing assets.
Claude was connected to Higgsfield through MCP (Model Context Protocol), the open standard that lets AI assistants call external tools. Once connected, Claude can generate images and video on your behalf and drop the results into the board.
In practice, that means Claude used ShotDeck like an art director to find visual direction, then used Higgsfield to generate original locations in the same grade and style as those references. The snowy supermarket exterior, the aisles, the samples table: all generated, all matching the look.
How to connect it
According to Higgsfield's own help center:
- In claude.ai or Claude Desktop, open Settings → Connectors.
- Choose Add custom connector, name it Higgsfield and paste the server URL
https://mcp.higgsfield.ai/mcp. - Click Connect, sign in to Higgsfield and authorize.
Higgsfield's MCP page also offers a one-click "add to Claude" flow, and the connector can be found by searching inside Claude's connector settings. The desktop app works well for long sessions.
Two fine-print points worth knowing before you hand an agent your credits:
- A paid Higgsfield subscription is required, and every generation through MCP deducts credits at standard rates. Higgsfield says unlimited and free generations apply only on its own website.
- There's no native credit cap for agents. Higgsfield suggests telling the agent in plain language, for example "do not spend more than X credits, and ask before generating." That's an instruction, not a hard limit.
Step 3: Build a board straight from your story
The board isn't just a one-off for the turkey film. It doubles as a reusable template with built-in instructions that tell Claude how the board is organized.
To show how fast it scales, a second concept was pitched to Claude by voice: a sunbed advert built around people huddling together. One long, rambling prompt described the story. Claude then took control of a fresh board and populated it with:
- Character sheets.
- Location references.
- Music choices.
- ShotDeck frames, downloaded and placed.
- Notes for voiceover and dialogue.
From there, the human's job is curation: tweak anything, swap in ShotDeck frames you prefer, tighten the story. That's the essence of the method. The AI produces a solid first-pass foundation; the human applies taste.
Step 4: Generate in Seedance 2.5, carefully
With pre-production done, it's time to make shots. Inside Higgsfield, you choose Video, select Seedance 2.5, and pick a resolution from 480p up to 4K.
The cost reality
Here's the number that shapes every decision: 4K Seedance is still very expensive per second.
So even though the model supports 30-second takes, nothing in this project ran longer than 10 seconds. The sensible pattern is to iterate cheaply at low resolution and save 4K for takes you're sure about.
A claim worth flagging: the project's maker argues Higgsfield offers the cheapest Seedance 2.5 4K on the market because it's a high-volume API distributor and negotiates better rates. That's plausible economies-of-scale logic, but treat it as a claim, not a fact. Public pricing is genuinely murky:
- The partner apps that launched 4K on November 11 did not publish a per-second 4K price, as The Inference noted.
- At 720p, published API rates already varied widely on launch day, from about $0.18 per second at one reseller to roughly $0.23 on BytePlus's own list and higher elsewhere.
- Dreamina's launch promo quoted Seedance 2.5 from 0.15 yuan per second and the Flash variant from 0.08 yuan per second, without saying which resolution those prices cover or when the promo ends.
Translation: compare per-second costs at the resolution you actually need, on the day you buy.
Upload your references first
Before prompting, every reference goes into the generator:
- The four main characters and the two elderly characters.
- Voice designs describing how each character should sound.
- A reference for the turkey itself (yes, the hero prop gets a sheet).
- Location references generated in Step 2.
Step 5: Let Claude write every shot prompt
Prompting follows a repeatable loop for every scene:
- Screenshot the scene card from the board: composition reference, shot notes, dialogue and the look-and-camera block.
- Upload the characters who appear in that scene, plus the location reference, into Claude.
- Ask for a Seedance 2.5 prompt, adding specific direction. For the opening shot, that meant: the friends standing outside a snowy supermarket, matching the uploaded grade and weather, with a subtle bit of cold breath fog while they talk "because we want that realism."
Claude returns a structured prompt that you copy into the generator. A custom Claude skill tuned for cinematic prompting can sharpen results, but it's optional. Claude's default prompts already work well when the references are strong.
The tagging trick that prevents chaos
One small instruction saves a lot of re-rolls. In the Claude chat, ask it to list all the @ references at the top of every prompt, in order.
In this project, that meant:
- @image1: the female lead
- @image2: Mike
- @image3: Danny
- @image4: Ru
- @image5: the supermarket
Why it matters: generators map tags to uploaded references by position. If Danny is supposed to speak but his image is in the wrong slot, the wrong face delivers the line. When Claude spells out the order every time, you upload references in exactly that sequence and the right person talks.
Step 6: Edit around the music
Once the generations are approved, everything moves into Premiere Pro. The edit follows a clear order:
- Music first. Download the track and lay it on the timeline before any picture.
- Build clips on the music. Here, a riser at the very top builds anticipation before the beat drops into "Meet Brenda and Graeme."
- Time every cut to the track. Timing is the whole game in a short like this.
- Color grade once the order is locked.
- Sound design and extras last.
The bonus tip: Scene Edit Detection
Seedance often produces multiple shots inside one generation. A single 10-second clip might contain three different camera angles with hard cuts between them.
Premiere Pro's Scene Edit Detection handles this. Right-click the clip, choose Scene Edit Detection, tick the options and hit Analyze. Adobe's machine learning finds the original cut points. Depending on the options you choose, it can:
- Cut the clip at each detected edit point.
- Create a bin of subclips.
- Add markers so you can review the cuts by hand.
Adobe notes it works best on hard cuts and may miss gradual transitions like dissolves. For AI multi-shot clips, which usually hard-cut, it's a big time saver. In a five-hour production, those minutes add up.
Why this matters
The headline is "AI video goes 4K," but the deeper shift is about who can make commercial-grade work.
- Pre-production is now automatable. Research, references, character design and shot planning used to take a team days. A well-structured board plus Claude compresses that into an hour or so of supervision.
- MCP turns chat into a studio. Once an assistant can call a generator directly, the gap between "plan" and "asset" shrinks to a sentence.
- 4K changes the delivery story. Native 4K with 10-bit color moves AI footage closer to deliverable spec for large screens and campaigns, not just social feeds.
- Draft modes change the economics. Paying 4K prices only for approved takes makes high resolution viable for small budgets.
For small agencies, brand teams and solo filmmakers, that's the real unlock: a festive spot with six characters, four locations and dialogue is no longer a crew-sized job.
Counterpoints and risks
None of this is free of trade-offs.
- Cost can spiral. 4K per-second pricing is high and often unpublished, and an agent with generator access can burn credits fast. Set explicit budgets in your instructions and watch usage.
- Brands and trademarks. The turkey short is set at a named real supermarket. That's fine for a personal piece, but a paid commercial using another company's brand, store or logos invites legal trouble. Use fictional or licensed settings for client work.
- Reference ethics. Borrowing palettes and lens choices is craft. Recreating someone's specific shot, character or likeness is not. Keep references as ingredients.
- Third-party connectors. Unofficial MCP bridges into tools like Milanote need access to your boards. Review permissions, and don't put client-confidential material on boards connected to tools you haven't vetted.
- Quality claims are still vendor claims. One launch partner called Seedance 2.5 4K its "highest quality to date," but there's no independent benchmark behind that yet.
- Disclosure. Platforms increasingly label AI-generated ads. Plan for it rather than hoping nobody notices.
Practical takeaways
If you want to try this workflow this week:
- Start with a board, not a prompt. Characters, locations, music with timestamps and a shot-by-shot breakdown before anything renders.
- Pick one camera package and never change it. Body, lens family, palette. Put it in every prompt.
- Use ShotDeck metadata, not adjectives. HEX values and lens names beat "moody and cinematic."
- Draft at 480p, finish at 4K. Only promote takes that already work.
- Keep shots to 10 seconds or less while 4K prices are high.
- Make Claude list the @ reference order at the top of every prompt.
- Cut to music first, then grade, then sound.
- Run Scene Edit Detection on every multi-shot generation.
FAQ
Is Seedance 2.5 4K a separate model? No. It's a resolution option inside Seedance 2.5, alongside 480p drafts and 1080p. You choose it at generation time.
Do I need Claude for this workflow? Not strictly. Any assistant that supports MCP connectors can drive a generator. Higgsfield also supports ChatGPT, Cursor and others. Claude is used here for planning, board building and prompt writing.
How long can a Seedance 2.5 shot be? Up to 30 seconds as one continuous take. In this project, shots stayed at 10 seconds or less to control 4K costs.
What does Seedance 2.5 4K cost? It depends on the platform. The four partner apps that launched it on November 11 didn't publish a 4K per-second price, and 720p API rates varied widely. Check the live rate where you generate.
Can I use the output commercially? Usually, within each platform's terms. Picsart, for example, says Seedance 2.5 videos made there can be used in ads and brand campaigns. Your own responsibility for trademarks, likenesses and disclosure still applies.
The bottom line
The best AI video in the world won't save a film with no plan. What makes this five-hour, one-person 4K commercial possible is the boring part done brilliantly: a pre-production board that Claude builds like an art director, a camera package locked from the first frame, and a generator wired in over MCP so references become assets in minutes.
Seedance 2.5 going 4K raises the ceiling. The workflow is what lets one person reach it. Build the board first, and the shots almost write themselves.
Sources
- Seedance 2.5 now generates 4K video in Picsart (Picsart)
- Four video apps switched on ByteDance's 4K model and none named a price (The Inference)
- Seedance 2.5 4K launches on Dreamina with 480p draft mode (AIHub)
- How to connect Higgsfield to Claude or ChatGPT (Higgsfield Help Center)
- ShotDeck: The Library of Cinematic Imagery
- Detect edit points using Scene Edit Detection in Premiere (Adobe)





