Stage 1: Building the Assets

Setting up the workspace

Before generating anything, open Higgsfield and go into Collab instead of Images. A separate Collab project gives you every image and video model in one place, and keeps everything you make for this ad in a single shared space — handy if you're working with a team.

Every asset below gets made once, locked, and reused across every shot. If a character, robot, prop, or location wobbles between scenes, the final video won't hold together — so this stage is worth taking slowly.

The product — can sheet

The source is a single photo of a can. Dropped into GPT Image 2.0 (currently the strongest model for image editing) with a prompt asking for a front, back, and top shot, it turns into one product sheet. Now the video model has the packaging from every angle later on, so it won't hallucinate details it never saw.

The character — reference sheet

For the hero, the plan is to use the creator's own likeness, which means Soul Cinema — right now the best option for holding a face consistent across generations. Twenty photos of the creator go in as the character source; the model studies the face across angles and lighting and builds a reusable avatar from it.

The prompt only describes layout — full body on the left, a face close-up on the right — and deliberately skips outfit details, so the model proposes a range of looks to choose from. With the prompt enhancer on and a 16:9, 2K, 40-batch run, eight generations cost just one credit, so a large batch is cheap to explore. A casual outfit gets picked in the end — partly because it reads well, partly because it'll contrast against the massive robot suit built later.

The location — empty NYC street

This is called out as the single most important image in the set: video inherits its texture and lighting from whatever still image it's built on, so a sloppy or plasticky location will drag every shot down with it, no matter how good the scene prompt is later. Soul Cinema is the choice here too, since it renders images that already look like film stills.

The prompt goes for a modern, empty New York street on a sunny day with a clear sky, an anamorphic lens for the wide cinematic feel, and film grain to strip away the too-clean digital look. Ten variations are generated at one credit for eight images; the one that's picked reads as tall business buildings and wide streets — clean enough to build every later scene on.

Before moving on, the character and location get tested together: a quick Seedance prompt puts the hero walking down that exact street, with a dynamic camera and a dutch angle. The combination holds up cinematically, so both assets are locked as compatible before any more time goes into the rest of the set.

The hero robot — #7

Two robots are needed: one for the hero, one for the rival. Rather than writing the image prompt by hand, Claude is asked to draft a detailed, structured character-sheet prompt — front and back view, white and red armor, number 7 on the chest — which saves time and avoids wasted generations.

That prompt runs through Soul Cinema in batches. With the prompt enhancer on, the results vary a lot — some too cartoonish, some oversized — until one lands that reads as a believable human-scale body. That version gets locked in as the hero's suit.

The rival robot — #9

The opponent is built the same way, with Claude asked for the same style of prompt but a different number and a darker palette — black and red armor, number 9 — so both robots read as belonging to the same world. Again several batches come back before one hits: enough of a villain look to work as the antagonist.

The ball — chrome and red

For the football itself, Claude again writes the generation prompt — this time for a ball with a metallic, slightly futuristic red-and-silver design. Run once through Soul Cinema, the result keeps the classic panelled look but adds red detailing, and it's locked on the first pass.

The goal — electric energy shield

The last asset is the goal, and it's deliberately not a standard net-and-crossbar. Claude is asked for a futuristic version with a glowing electric effect, matching the sci-fi look of the robots. The result reads as futuristic enough on the first generation, closing out the asset set.

With the can, character, both robots, location, ball, and goal all locked, every recurring element of the commercial now exists as a consistent, reusable image — ready for the prompting framework in the next stage.