Cinematic AI: the juice.ext teardown

Primary-source harvest of the creator, his disclosed stack in his own words, and exactly what our account can and cannot reproduce.

01Who he actually is

Not a hobbyist. A studio with a price list.

Verified account, real name Juice, based in Amsterdam, bio reads "Futureproof Filmmaker & Creative Director | Cinematic world-building for premium brands". His site at extjuice.com is a working intake funnel built on GoHighLevel, and it sells three things: Premium Brand Films, World Build and Creative Direction.

The intake form asks for name, company, purpose, target length and budget, with tiers at under €2k, €2k to €5k, €5k to €10k, €10k to €50k, and €50k+, then promises a response in 48 hours. He also runs a YouTube channel at @juiceext.

02His stack, in his own words

Disclosed in captions, not inferred

03The method

Ten-page prompts, and regeneration until it locks

The single biggest technique revelation is prompt length. He is not writing a paragraph or a shot description. He writes ten pages per piece and picks the one model that will hold all of it. Everything else follows from that.

His stated loop is not generate-and-accept. It is recognise the pattern, locate the issue, regenerate. He treats visible drift as a defect to be hunted, which is the working version of the control hierarchy: subject before environment before camera before lighting.

The four laws, from the secondhand analysis, all still confirmed

Distribution is a serialised challenge

Across 74 posts from 9 March to 15 September, the dominant format is a numbered daily series: "day N of filmmaking in the AI era" appears 45 times, plus an 8-part "Juice in Brazil" run. He also built a three-post narrative arc where one story ran across three separate reels.

The engagement is brutally skewed. Median 956 likes, mean 3,299, top post 69,526. Roughly one post in ten carries the account. Captions are long and personal, median 655 characters, and 4.7 hashtags per post. He follows zero accounts.

04What we have already generated

The pipeline is proven, the output class is not

We have run this stack exactly once, on 29 August, during the Ventures work. That batch spent 28.05 credits, about $1.75, of an 80-credit ceiling.

• A soul/v2 probe at 0.050 credits confirmed credits were live.
• A dop-lite canary clip animated a static poster: 2.00 credits, 276 KB mp4 plus webm, verified as a slow push-in with the interface preserved.
• Thirteen failed calls taught us that dop-lite validates motions[].id as a UUID preset that no public endpoint lists. Those failures cost nothing.
• scripts/higgsfield_poster_to_video.py came out of it, dry-running by default.

So the plumbing works end to end. What we have never attempted is the thing this brief is about: a locked human character carried across multiple directed shots. Poster-to-video is a different and much easier problem.

05Cost by length

Nothing generates past ten seconds

Every validated video model on our account caps at 10 seconds per generation. kling-i2v and kling-turbo accept only [5, 10]; hailuo-02 only [6, 10]. Length is therefore a count of shots, not a render setting.

Quality path per shot is 8 explores on soul-v2, 2 hero stills at 1080p in 9:16, and 2 animations on kling-i2v at 10 seconds. Budget path drops to 720p stills and dop-lite.

Confirmed unit prices, measured today

num_images at 1, 2 and 4 all estimate to the same price, which would mean four variants per call at no extra cost. One real call should confirm that before anyone budgets on it. Separately, the DoP family's estimate returns a flat price for any duration including the literal string BOGUS, so it is not validating and its duration ceiling remains unproven.

06How to get extremely high quality

Copy the method, route around the missing model

1. Write ten pages, not ten lines. This is the highest-leverage change and it costs nothing. Full wardrobe, lens, lighting, blocking, weather, era and grade per shot, plus what must not change between shots.
2. Lock identity before motion. soul-character at $0.094 with a seed, reused across every shot. If the face drifts in the still, no video model will rescue it.
3. Still-first, always. Art-direct the hero frame at 1080p in the target ratio, then animate. Image-to-video inherits the framing instead of inventing one.
4. Use kling-i2v as the drift dial. It is the only family exposing negative_prompt and cfg_scale, which is the closest thing we have to his regenerate-until-it-locks loop.
5. Upscale after, not during. He upscales raw shots with a dedicated tool rather than asking the generator for resolution. Given our 1080p ceiling this is not optional, it is the only route to a 4K master.
6. Grade in Resolve. Same as him. The noir look is lens behaviour, halation and grain, and that belongs in a grade, not a prompt.
7. Unblock Seedance. Everything above is a workaround. He is explicit that Seedance 2.5 is what makes multi-shot consistency survive, and it is the one thing we cannot buy our way to.

07Honesty

Now verified that was not before

• He uses Higgsfield, Seedance 2.5, WAN, Lupa Upscaler, DaVinci Resolve and Udio. All self-disclosed in captions, not inferred.
• He writes ten-page prompts and names consistency, not prompting, as the hard part.
• He runs a commercial studio with published budget tiers up to €50k+.
• Seedance is blocked on our key with a 403, distinct from the 404s.

Still unverified

• The exact split between generation and manual post. He credits Resolve but never quantifies it.
• What he pays Higgsfield, and whether the Creator Program gives him model access we would not get.
• Whether Lupa or Udio are worth adding. Neither has been evaluated or priced.
• Whether our support request has moved. No reply logged since 29 August.
• The DoP duration ceiling, and whether num_images is genuinely free.