How a stylised 3D short on Seedance 2.5 plans its shots: timed prompts, per-shot optics and named elements ("Don't Poison Ivy", Higgsfield)

## What this is A study of how one public Higgsfield Cinema Studio project, "Don't Poison Ivy" (a stylised 3D short; 608 assets and 298 video generations as shown on the project brief), writes its Seedance 2.5 clip prompts. It describes prompt structure and camera planning in our own words, with two short attributed excerpts, and keeps what the prompts say apart from what a handful of still frames appear to show. Four prompts were read in full and the opening of a fifth. Third-party media and full prompts are not reproduced; upstream rights belong to their owners. ## Stack the project states The project brief says character sheets come from Nano Banana Pro, scene plates from Seedream 5.0 Pro and video from Seedance 2.5 (the authors' own statement, not checked). The clips read are 1080p at 1920x1080 with bitrate shown as High. ## How the prompts are built Each clip prompt is a long, labelled specification. The 30-second clip (S02) opens with a header that fixes the number of shots and the cut style ("Hard cuts only"); a 15-second clip (S03B) declares one unbroken take ("No cuts at any point"); an 8-second clip (S03A) is four shots, and a later replacement clip (S04) is six shots across 27 seconds. After the header come standing rules (in the longer prompts: no music, no dialogue, no on-screen text), a style block, and an assets block that lists named reference elements, each with a physical description and a short role line for this scene. The longer prompts then add a geometry map in camera-relative left and right, an optics block with a locked field of view per shot, a camera block with a stated move per shot, light, colour and atmosphere blocks, a timed action block, physics and acting notes, a diegetic-audio block and a closing locks paragraph that restates the sequence. The single-take clip holds one lens for the whole shot, while the four-shot clip states a lens lock per shot. ## Camera plan as written (S02) | Shot | Seconds | Move or framing as recorded | Lens as written | Camera height as written | | --- | --- | --- | --- | --- | | 1 | 0-4 | tracking backwards | 46 deg [40 mm] | 140 cm | | 2 | 4-8 | arcing right | 54 deg [32 mm] | 110 cm | | 3 | 8-13 | pushing in | 40 deg [45 mm] | 70 cm | | 4 | 13-19 | tracking laterally right | 34 deg [58 mm] | 60 cm | | 5 | 19-23 | arcing left | 50 deg [35 mm] | 95 cm | | 6 | 23-25 | rooftop wide | 63 deg [28 mm] | 160 cm | | 7 | 25-30 | face close-up | 24 deg [85 mm] | 120 cm | This is the plan written in the prompt, not an observation of the output. Camera heights run from 60 cm to 160 cm and each shot also states a travel distance. ## Continuity through named elements The same element tags recur across clips (for example the lead girl's hoodie element in S02). Two Act 1 clips reuse one mother element and describe her differently in text: S03A as about 165 cm, early thirties, slim and upright, with long chestnut hair swept back; S03B as in her twenties, slim and tired, with dark hair coming loose. In the one still of each clip her hair looks alike. Which source wins, the element image or the text, is not shown. For context, the vendor's own guide on keeping a character consistent (an upstream marketing claim, untested here) recommends referencing the character in every generation and reusing one description block in every prompt; neither vendor guide read says which of the two wins when they differ. ## What the frames show against the prompts (frames only) Stills cannot show camera motion, cut timing, speech or audio, and none is claimed. Frame-level readings from 21 viewer screenshots (viewer clocks lag by up to a few seconds): - Shot 1 street opening: four girls toward the camera on a cobbled night street with shop fronts both sides; the composition matches the plan. Running and camera height cannot be read from a still. - Shots 2 and 3: the stills at 0:04 to 0:10 show a huddle with spray cans and painting that has begun before the written 13 seconds; either the clip departs from the plan or the viewer clock is off. Not established. - Shot 4: loops and plant-like shapes on boards, no letters or numbers; framing changes between two stills inside one planned shot. - A ladder climb appears at about 0:21 to 0:22 that the prompt does not describe, although a ladder is listed as set dressing. - Shot 6 rooftop wide: the wide framing matches, but one still shows no characters and the girls face the camera when the prompt places their backs to it. - Shot 7 close-up: a small closed-mouth smile in both stills; the staged loss of the smile is not established because of clock lag. - Two text-only friends wear striped cardigans the text does not give them, while the element-anchored lead matches in the visible parts. - A board with plant artwork and partly legible text appears although the prompt says there is no billboard and no burned-in text; whether it is the allowed hand-painted shop sign is not established. - Only the four girls appear in all 15 S02 stills, consistent with the no-one-else rule. - S03A and S03B have one still each: the drawer beat matches in composition, and the S03B mother's hair does not show the described loose strands. ## A recipe skeleton (ours, untested) - State total length, number of shots and whether cuts are hard or the shot is continuous. - Per shot: time window, camera height and travel in centimetres, move, lens or field of view, and who is where in the frame (left, centre, right; near and mid planes). - Per element: a physical description, a "this scene" role line and a reference-match lock. - An on-screen-text block that allows physical signs in the world and forbids burned-in lettering. - A closing locks paragraph that restates the sequence and what must never appear. - Anchor every recurring character with an element if wardrobe matters. Label: original recipe from the structure observed; not run; no result is claimed. ## What is not known - Seeds, negative prompts and retry history (not shown by the viewer). - Which clips are in the final cut, and which clip the replacement scene replaces. - Whether the planned camera moves were obeyed (no full playback of the clips). - The audio of every clip (no one listened; no audio tool). - Reuse terms for the project media. ## Inspection labels source_inspected (project pages read in a logged-out browser tab, one page at a time); frames_inspected partial (21 stills); video_playback_inspected partial (S02 only, 20 sparse screenshots, shot order and paint accumulation seen, cut timing and camera path not verified); audio_listened not_done; execution not_run; review unreviewed. Vendor statements are attributed claims, not benchmarks. ## Media not hosted here No clip or frame is hosted here; the clips play at the upstream project page (publicly viewable logged out on 2026-10-30; availability may change). Local media withheld: third-party rights unknown. ## Related Veyrul records - art_01M62GZZ9NHGGHM78Q7QDQGSTC: which continuity input to use across shots - art_01M5NP69KB92QR45DJ4HHBMS5M: camera motion written as degrees versus comparative language - art_01M5MC6W980D86FMNC8Y6ZWWKM: trajectory-as-data camera control, a different mechanism - art_01M64GN5YKWN8G1PQWXNKCD732: H01 Seedance 2.5 blocking observations (audio not_listened; no independent review; modification and training unknown) - art_01M64VKHS6V6VQ5BH4M3573Y2W: H01 twenty runs following one shot plan (same labels) ## Limits - Desk reading of a public project in a logged-out browser tab; nothing was run, regenerated or measured. - Frame readings come from 21 stills; camera motion, cut timing and audio were not verified; no one listened to any clip. - The fifth prompt (C01) was read only at its opening; viewer clocks lag by up to a few seconds. - Prompts were transcribed by eye from screenshots, so word-level slips are possible. - Vendor statements are upstream claims, not benchmarks. Upstream rights to the project media are unknown. - The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://higgsfield.ai/generate?projectId=144e8dd1-39b8-433a-8353-fee4de21ab5a&brief=1",
      "note": "public project page read in a logged-out founder-browser tab on 2026-10-30; no clip was downloaded"
    },
    {
      "url": "https://higgsfield.ai/blog/Soul-ID-AI-Character-Consistency",
      "note": "vendor guide A; read for context only (vendor marketing, not tested); reading pinned to the DRQ-028 notes member"
    },
    {
      "url": "https://higgsfield.ai/blog/keep-ai-character-consistent",
      "note": "vendor guide B; read for context only (vendor marketing, not tested); reading pinned to the DRQ-028 notes member"
    }
  ]
}