Directing AI video with Higgsfield

The work behind the record

On 2 October 2026, we used ChatGPT with Higgsfield to explore several kinds of generated film: two vertical narrative concepts, a presenter-led biography with supporting inserts, and three automotive clips. The recovered Higgsfield history contains 15 completed video jobs. They all record the same generation model, seedance_2_5, with 1080p requested and generated audio enabled.

The useful progression is in how we specified the work. Early concepts put a complete story into a single request. Later requests separated identity references, performance references, shot boundaries, dialogue, sound, and continuity. The brief became more inspectable: we could point to a particular cut, wheel, camera position, or spoken line instead of asking for a generally better film.

This is a reconstruction from the saved generation requests and completion records. Those records establish what we requested and which jobs returned outputs. They do not establish a frame-by-frame quality score, a measured improvement rate, or a final assembled master. The full conversation and editing receipts are needed for those stronger claims.

The stack: director, service, generation model

ChatGPT was the conversational direction layer. It let us develop a brief in ordinary language and express it as a production request: story, shot list, reference inputs, duration, aspect ratio, and audio requirements. The ChatGPT role here comes from my account of using the service through ChatGPT; the recovered ledger is Higgsfield-side evidence. Its exact underlying model is not preserved in those generation records. Assigning it a model version from the date would be speculation.

Higgsfield was the connected generation service. Its tools expose model discovery, media inputs, generation submission, job state, and returned media. The general tool workflow is asynchronous: submit a request, retain its job ID, wait for a terminal state, then inspect the returned artifact. A submission acknowledgement and a completed job are different states. Model names belong to the actual generation records; a generic nested parameter such as model: "default" is not a reliable replacement for the job's top-level model ID.

Seedance 2.5 was the video generator. Higgsfield's current catalog identifies ByteDance as its provider and exposes text-to-video, multimodal reference generation, video editing, and extension. In this recovered set it generated videos with audio requested, using either text alone or images and an earlier video as references. These are hosted generation jobs, not a local GPU experiment. See Higgsfield's Seedance 2.5 guide.

The distinction matters technically. A fluent planning model can write a precise instruction without the media generator satisfying it. We evaluate the planning request and rendered artifact separately.

First stage: put a whole narrative into one short request

The first recovered job asked for a 20-second vertical portrait story. It used a consistent adult character, five time intervals, scene actions, synchronized sound, and a held end card. A second, 15-second vertical concept used a thread moving through a textile environment: immediate action, a physical obstacle, a reveal, and a proposed return to the opening composition.

Both were prompt-only. There was no supplied identity image or prior video. Each request therefore asked the model to invent the visual world and maintain it throughout its own clip.

The briefs already contained useful structure. They specified when the opening should become legible, where the payoff belonged, which sounds should accompany physical contact, and how the ending should behave. They also separated an aesthetic story from claims about physical transformation. Those are authoring decisions supported by the requests, not evidence that generated typography or a seamless loop passed inspection.

A 15-second limit is a temporal budget. Spending four seconds on setup leaves fewer seconds for an obstacle and resolution. We made that budget visible in the prompt rather than leaving the generator to allocate it implicitly.

Second stage: establish one performance before producing chapters

The presenter experiment changed the input contract. Its opening used three photographs of the same person, with each described as evidence of identity rather than a pose to hold. The ten-second request combined an introductory spoken line, a cup placed on a saucer, a physical camera push, wardrobe, and a specific harbor setting.

The progression then became a dependency graph:

text + three identity photographs
                  |
                  v
         opening performance video
                  |
          +-------+-------+
          |               |
  chapter requests    supporting inserts
  images + opening    independent text-only

Seven subsequent presenter chapters reused the three identity photographs and the opening video. Each explicitly asked to preserve the opening's voice and performance while changing location, wardrobe, camera angle, or dialogue. Two additional jobs supplied context: a five-second harbor establishing shot and a four-second cup insert.

Reusing the opening matters because voice descriptions are underspecified. “Warm, conversational” allows many voices. A concrete audiovisual reference supplies an example to condition on. That can reduce ambiguity, but the request alone does not prove speaker consistency or lip synchronization across outputs.

The camera instructions also became more anatomical. “Camera on her anatomical right, right cheek nearest” is more explicit than “three-quarter view,” whose orientation a model can interpret differently. Six of the seven chapter prompts named a precise dialogue phrase after which an immediate hard cut should occur; the remaining ten-second chapter requested one continuous shot. The intent was to tie editing to speech structure instead of allowing arbitrary cuts.

The prompts addressed a subtle reference failure: a photograph establishes appearance, not a held pose. Requests called for changing hand positions, posture, and gestures with the words. Several also constrained cup timing or drinking during speech. This made motion and performance part of the acceptance criteria.

The recovered result is ten completed presenter-related jobs: eight speaking clips and two inserts. Their requested durations sum to 114 seconds. That is source material; it is not evidence of a finished 114-second edit.

Third stage: carry continuity from one automotive clip to the next

The automotive experiment used four images with distinct roles: driver identity and front, side, and rear views of the same car. Multiple views gave the request concrete appearance evidence across changing camera angles. They did not create a calibrated 3D asset or guarantee exact geometry.

The first 15-second sequence specified four shots in a wet, imagined city: interior performance, wheel insert, front tracking, and side tracking. It was written to end with forward motion still alive for a proposed second segment.

The next 15-second request added that first generated video alongside the four original images. It described the continuation's entry state, prohibited repeating the opening, and carried across the car, driver, lighting, wardrobe, and soundscape. It used four further shots, including a controlled turning maneuver, an overhead view, and a moving rear composition.

That is reference-driven continuation. The saved parameters do not establish use of the dedicated video_extension mode. A request saying “the second fifteen seconds” should not be silently relabeled as a technical extension operation.

A later standalone 15-second Dubai sequence returned to the four images and tightened the physical constraints into six timed shots. Its wheel insert distinguished the rotating wheel from the brake caliper fixed relative to the suspension. Its closing shot described gradual deceleration before rest rather than an instantaneous transition from speed to a parked car.

One clear authoring shift visible in the record: general realism became a list of observable relationships. Tire rotation should agree with road travel; calipers should not rotate with the wheel; paint, wheels, and body should remain stable across cuts; the final stop should have a credible transition. These constraints are more useful to inspect than “cinematic and realistic.”

The recovered outputs are three completed 15-second automotive jobs. Two were designed as adjacent segments, and the third was a separate six-shot film request. A combined 30-second export is not established by these records.

What the settings actually tell us

The 194-second total adds requested generation parameters; it is not a measured runtime of the downloaded files.

High bitrate was requested for the eight speaking clips and three automotive clips. Standard bitrate was used for the initial concepts and two inserts. Bitrate describes output encoding; it is not a direct measure of identity fidelity or physical correctness.

The records show multi_shots: false, empty multi_prompt arrays, and detailed shot lists inside the main prompt. The timing and cuts were therefore authored as natural-language constraints rather than as populated structured per-shot entries. “Exactly four shots” remained an instruction for the generator to interpret.

Likewise, “24 fps,” “ARRI Alexa,” and “anamorphic lenses” appeared as cinematographic language in a prompt. That does not prove a measured frame rate or physical camera pipeline. A real deliverable receipt should report file metadata and visual inspection separately.

Why references and constraints improved the method

The progression targeted several kinds of ambiguity:

• Identity: separate original photographs from the scene and wardrobe instructions.
• Voice: reuse one opening performance rather than repeatedly describing a new voice.
• Continuity: provide the previous generated clip, its ending state, and a distinct next action.
• Editing: attach shot boundaries to time intervals or dialogue phrases.
• Physics: specify relationships between moving objects, their contact with the environment, and camera motion.
• Sound: distinguish foreground speech, environmental sound, mechanical effects, and music.

These improvements concern the production contract. A controlled model comparison, seed-matched baselines, and a measured numerical quality gain are not present in the recovered records. The recovered set is a same-model workflow exploration.

It also shows why expanding a prompt can reach diminishing returns. Identity, geometry, camera, speech, sound, exact cuts, and exact lettering compete for a short output. A written requirement needs an inspection point. If a feature remains unreliable, the next step is to isolate it into a separate shot or perform the edit explicitly.

The inspection contract

For this kind of work, I would close a clip with five checks: compare identity at the first, middle, and last frames; check reference-specific geometry across cuts; inspect motion and contact; listen to every spoken line and transition; and verify delivered duration, resolution, aspect ratio, and encoding.

For a sequence, add continuity at the joins. Does the next segment begin from the previous segment's motion and sound state? Does the voice remain the same? Has the character, wardrobe, or lighting drifted? Does the edit actually contain the prescribed cuts?

These are acceptance checks derived from the experiment, not a claim that all checks were performed on this recovered set. A status of “completed” with a media URL establishes the provider’s completion record and output reference. A final production receipt should identify the selected takes and assembled export, with inspection evidence for the properties that mattered.

What was completed

The recorded output set contains two vertical concepts, ten presenter-related clips, and three automotive clips, all marked completed. The technical progress was a more precise way of directing generation: establish identity and performance, reuse references deliberately, specify the next state, and turn vague realism into inspectable constraints.

The remaining evidence gap is the full human feedback and final editing history. Until those receipts are recovered, this report makes no claim that every instruction passed, that all clips were approved, or that a combined master was exported.

The reusable method is to give each reference a role, each shot a purpose, and each completion claim an observable artifact.