Eight minutes, one vendor, no crew list: what CANDY actually proves about Seedance 2.5

The film is called CANDY. It runs 8 minutes 59 seconds, and the player duration reads the same on the two posts that carried it, BytePlus Global's account and the director's own, so that much is settled. The claim wrapped around it, that a single creator made a real movie start to finish on one model, is not settled at all, and the people closest to the film have never said it.

CANDY is credited to DiDi_OK (@DiDi_OKK), who works on the festival circuit as Jiaze Li. It was made with Dreamina Seedance 2.5, the consumer surface for ByteDance's video model rather than the BytePlus ModelArk API, the paid developer route into the same model, which was still listed as coming soon at the model's 31 July launch. The distinction matters if you are trying to reproduce the result, because one is a subscription product with per-second pricing and the other is a developer route most people cannot buy yet.

The version of this story that traveled came from a quote post. Min Choi's post, the one our tip arrived on, quotes a BytePlus Global post that X renders with a Paid partnership label and a Made with AI label. DiDi_OK's own launch post is equally direct about it: "Happy to collaborate with @BytePlusGlobal on this project." So the film is a sponsored showcase, published ten days into the model's rollout cycle, and the account that promoted it hardest is the account selling the model. That does not make CANDY bad work. It does mean the framing around it is marketing, and any read of the film has to account for who chose it, and why.

The record it is supposed to have set already belonged to him

The pitch on this film is runtime, on the reasonable theory that sustaining character and continuity past a few minutes is where these pipelines fall apart. The theory is sound; the example undercuts it. DiDi_OK posted an 11:53 AI sci-fi piece in July and an 11:16 film in April, both longer than CANDY, both his. Neill Blomkamp's Nightborne ran 13 minutes in July on Seedance 2.0. In Korea, a 90-minute generated feature called Steelay reported roughly 13,000 generated outputs to get to length.

So 8:59 is not a ceiling being broken. It is a competent length for a sponsored short by a director who has gone longer twice this year, and the honest question is not how long the film runs but how the length was assembled.

The arithmetic the caption skipped

Nobody has published what CANDY cost, and the top reply under Min Choi's post asks exactly that. We can bound it from published rates, and the numbers below are our calculation rather than a disclosed budget, so treat them as a floor with stated assumptions.

• 539 seconds of finished footage at a 30 second generation ceiling means a minimum of 18 kept generations, and realistically more, because cuts run shorter than the ceiling.
• Dreamina lists Seedance 2.5 at $0.097 per generated second on its annual Advanced plan, which puts the kept footage at roughly $52. An observed credit rate of about 780 credits per 30 second pass, or $9 to $11, puts the same footage at $162 to $198.
• Apply a conservative 3x to 5x discard ratio and the generation spend lands somewhere between $156 and $260 on plan rates, or $486 to $990 on credit packs, before a single hour of human time, sound work or editing.

For scale, Steelay reported 500 to 600 million KRW on generative tools alone for its 90 minutes, and the French director Kevin Mendiboure logged 3,229 generations across 242 hours for one solo short. Generation is the cheap part of finishing a film, which is precisely why the credits question keeps mattering.

The director's own verdict is sharper than the campaign's

Jiaze Li is not an anonymous hobbyist who got lucky with a prompt, and writing him as one gets the story wrong twice over. He came out of advertising, won first prize and 1,000,000 RMB at Bilibili's first AI creation competition for the short 牌子, and took both the Reply AI Studios Grand Prix and the Community Award in Venice on 6 September for Website, from a finalist field of ten drawn out of more than 3,000 entries across 76 countries. His earlier films credit Seedance 2, Kling 3 Omni, TapNow, Veo 3 and Runway by name. The single vendor credit on CANDY is a break in his own pattern, which is roughly what a sponsorship is for.

He is also the most useful skeptic in the file. Talking to Chinese press about AI generated live action performance, he called it 预制菜, pre made supermarket food, and argued that AI video has not yet crossed its starting line. The moment nobody talks about the generation technology any more, he said, is the moment AI video truly passes the start. Set that next to a promoted caption about Hollywood not being ready and you have the whole tension of this moment in two sentences from the same person.

What this means for AI creators

Two practical things come out of this, and neither is in the vendor blog.

First, the edit is doing the work. A model with a 30 second ceiling and independent testers reporting drift in extension rounds two and three cannot sustain a nine minute performance, so a nine minute film is an assembled edit of short generations. If you are planning long form, budget your time for cut rhythm, sound design and continuity patching, not for one heroic generation run. Structural devices are legitimate here: Nightborne used documentary style interviews to disguise clip length limits, and that is a craft answer to a real constraint rather than a cheat.

Second, continuity is a workflow you build, not a feature you buy. Asked whether Seedance keeps character continuity across revisions, Min Choi's answer was that it does if you have a clear character sheet to use as reference for each revision. In practice, that means the reference discipline of an animation production. Build the character sheet before you generate anything, keep the wardrobe and prop details written down, and treat the 30 image reference slot as the place that discipline gets enforced. The people finishing long AI work are the ones who did the pre production, and that is a skill nobody can take away from you when the model version changes.

What the replies are asking

The conversation under these three posts is more honest than the campaign copy, and two of the questions are worth answering in public.

How many generations went in the bin, and what did the compute bill look like? Nobody involved has said, so our bounded estimate above is the closest anyone has come, and we would replace it with real numbers the moment the production shares them.

Hollywood has been not ready every eighteen months since 2022. That reply, sitting quietly at a couple of hundred views, is the tone trap. The interesting claim is never that an industry is doomed, it is that an award winning director can now finish a nine minute film on a consumer subscription while telling reporters the performances still taste pre made.

What we have not verified

Stated plainly, because the framing of this story depends on it.

• Crew size is unpublished, because there is no published credit list for CANDY. Solo authorship is asserted by a third party poster and quietly contradicted by the director's own language, which runs to "our bet", "our weird little film" and "we made an AI film". We are not printing a number either way, and we have asked.
• Whether any tool outside Seedance touched the final film. Edit, upscale, sound, music and titles are all unstated by both the creator and BytePlus, so "100% AI" is a generation claim at best.
• The commercial terms of the BytePlus collaboration, including fee, creative control and approval rights. X labels the post a paid partnership; the terms are private.
• Our own shot by shot read is still owed, because this piece is reported, not a watch through. The full film sits on Bilibili as well as X, and the teardown, cut length distribution, identity drift across scene boundaries, audio provenance and one hard physics shot, publishes as a follow up once we have watched it end to end on a playable copy.

Engagement figures move, so for the record: read on 14 September, Min Choi's post showed 73,595 views and 160 likes, DiDi_OK's own post 928,783 views and 1,829 likes, and BytePlus Global's 1,750,597 views. The film that is supposed to prove a solo creator can go it alone got most of its reach from the model vendor's own account.

CANDY, directed by DiDi_OK / Jiaze Li. Tools as credited by the creator and BytePlus: Dreamina Seedance 2.5. Sources: ByteDance Seed launch post, 31 Jul 2026; BytePlus Global and @DiDi_OKK posts, 10 Sep 2026; @minchoi post, 12 Sep 2026; Reply AI Film Festival finalist and winner announcements; Southern Weekly profile; independent Seedance 2.5 hands on testing; Dreamina published pricing. Cost figures are our calculation from published per second rates, not a disclosed budget. Crew size, non Seedance tools and deal terms remain unverified.