3-Step Workflow To Make Ultra-Realistic AI Ads

2026-06-23 · case-study tutorial · Adil · recording, 35:37 · watch on YouTube

Topic: Video generation & shot direction (also: Assets, references & consistency, Prompting & AI assistants) · mode: Tutorial · tags: fiction, prompts shared, one creator's experience

Adil builds a cinematic headphones commercial on Higgsfield in three steps: locked assets made in Soul Cinema and GPT Image 2.0, a Claude skill that writes the video prompts, and scenes generated in Seedance 2.0. He tests candidate faces and locations in video, swapping one asset at a time, and fixes face drift by erasing the extra face from the character sheet. Across five scenes he adds a second sheet when a character changes, a schematic map to pin positions, and his music track to time the dance. The finished ad, he says, is the best seconds of about a hundred tries: iteration is the skill.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — The finished headphones commercial and Adil's introduction

00:01:15 — The three-step workflow

★★★ Adil's ad workflow has three steps: the assets (product, characters, locations, props) in Soul Cinema and GPT Image 2.0; a Claude skill that turns the story into prompts; the scenes generated in Seedance 2.0, all on Higgsfield. — 00:01:15 · Soul Cinema, GPT Image 2, Claude, Seedance 2.0, Higgsfield AI · also in: The 3-Step Realistic AI Ad Workflow: The finished ad, and the 3-step system

00:01:43 — Stage 1: the product sheet in GPT Image 2.0, the hero's character sheet in Soul Cinema, the boss in Cast

★★ With one photo of the headphones, Adil has GPT Image 2.0 make a product sheet with front and three-quarter views, since the video model must see the product from every angle or it hallucinates mid-scene. — 00:01:43 · GPT Image 2 · also in: The 3-Step Realistic AI Ad Workflow: The product sheet (GPT Image 2.0)
★★ Adil calls GPT Image 2.0 the best model right now for product editing. — 00:01:43 · GPT Image 2 · also in: The 3-Step Realistic AI Ad Workflow: The product sheet (GPT Image 2.0)
★★ For the hero of the ad Adil uses Soul Cinema, which he calls the best model for photorealistic looks. — 00:02:59 · Soul Cinema
★★ Adil has Claude write the character-sheet prompt: a young man with an expressive face, two panels (a close-up and a full body front and back) on a grey background. — 00:02:59 · Claude, character sheet · also in: The 3-Step Realistic AI Ad Workflow: The hero character (Soul Cinema)
★★ Adil says a grey background matters on a character sheet: the team found the win rate much higher on grey, because no background clutter competes with the character. — 00:03:25 · character sheet · also in: The 3-Step Realistic AI Ad Workflow: The hero character (Soul Cinema)
★★ Eight Soul Cinema images cost one credit, so Adil runs many batches and picks the face that stands out. — 00:03:25 · Soul Cinema · also in: The 3-Step Realistic AI Ad Workflow: The hero character (Soul Cinema)
★★★ Adil advises keeping a couple of candidate faces instead of one and testing them in video, because a face that looks great as a still can fall apart once it moves. — 00:03:55 · also in: The 3-Step Realistic AI Ad Workflow: The hero character (Soul Cinema)
★ Adil uses the canvas as his workspace: every candidate side by side for comparison, the winner dragged to the top to lock it. — 00:03:55 · Canvas · also in: The 3-Step Realistic AI Ad Workflow: The hero character (Soul Cinema)
★★ For a side character such as the office boss, Adil uses Cast, which gives instant character sheets and several styling options from one prompt. — 00:04:42 · Cast, character sheet

00:05:10 — Building and testing the locations; erasing the duplicate face on the character sheets

★★ Adil says the locations make or break the video: if a location looks fake or plasticky, no prompting will save it. — 00:05:10 · AI look · also in: The 3-Step Realistic AI Ad Workflow: Locking the first location: the kitchen
★★ Adil generates each location in Cinematic Locations at a three-quarter angle: it gives the room more depth for Seedance 2.0 to hold when the camera moves and a much better win rate than a flat head-on shot. — 00:05:40 · Seedance 2.0 · also in: The 3-Step Realistic AI Ad Workflow: Locking the first location: the kitchen
★★★ To choose between candidate heroes and kitchens, Adil tests them with the same simple video prompt every time and swaps only one asset at a time, so he can see what makes the difference. — 00:06:16 · also in: The 3-Step Realistic AI Ad Workflow: Locking the first location: the kitchen
★★ A character sheet with several faces in one image makes the video model drift when animating, because it does not know which face to grab. — 00:07:45 · character sheet · also in: The 3-Step Realistic AI Ad Workflow: One face per character sheet
★★★ Adil's fix for face drift: in GPT Image 2.0, erase the face from the full-body panel of the character sheet, leaving one face to lock onto; he does the same for the boss. — 00:08:13 · character sheet, GPT Image 2 · also in: The 3-Step Realistic AI Ad Workflow: One face per character sheet

00:09:50 — Wardrobe: outfit ideas from Claude, the outfit edit and masking it over the original; the props

★★ Adil says testing every location in video only feels slow: it avoids wasting many credits later on full scenes generated with the wrong assets. — 00:09:50 · also in: The 3-Step Realistic AI Ad Workflow: Locking the stadium, street & office
★★ For the hero's wardrobe Adil gives Claude an image of the character, asks for casual outfit ideas each written as a prompt, and runs all ten in GPT Image 2.0 to compare. — 00:10:18 · Claude, GPT Image 2 · also in: The 3-Step Realistic AI Ad Workflow: The outfit edit, kept photoreal
★★★ Adil says every edit of an asset costs quality: running his crisp Soul Cinema character through GPT Image 2 edits softened everything, which is how the flat, plastic AI-slop look comes in. — 00:10:58 · GPT Image 2, Soul Cinema, AI look · also in: The 3-Step Realistic AI Ad Workflow: The outfit edit, kept photoreal
★★★ To keep an edited asset real, Adil layers the original Soul Cinema shot over the GPT Image 2 edit in any photo editor and masks out the old outfit, so only the outfit comes from the edit. — 00:11:19 · Soul Cinema, GPT Image 2, AI look · also in: The 3-Step Realistic AI Ad Workflow: The outfit edit, kept photoreal
★★ Props that recur across scenes (the sky dancer, the sneakers, the backpack) each get a clean reference sheet in GPT Image 2.0 and no motion test, since objects do not perform. — 00:11:49 · GPT Image 2 · also in: The 3-Step Realistic AI Ad Workflow: The props: sky dancer, sneakers, backpack

00:12:16 — Stage 2: the shot-list Claude skill, the style prefix and named prompts

★★ Adil says most people waste credits by typing an overloaded Seedance 2.0 prompt from scratch and tweaking it until twenty generations have cost a fortune. — 00:12:19 · Seedance 2.0 · also in: The 3-Step Realistic AI Ad Workflow: The shot list, and the Claude skill behind it
★★★ Adil never writes video prompts by hand: a skill he built in Claude writes them. — 00:12:19 · Claude · also in: The 3-Step Realistic AI Ad Workflow: The shot list, and the Claude skill behind it
★★ Adil defines a Claude skill as a file loaded into Claude that teaches it one job done your way; his writes shot lists with prompts tailored for Seedance 2.0. — 00:12:46 · Claude, Seedance 2.0 · also in: The 3-Step Realistic AI Ad Workflow: The shot list, and the Claude skill behind it

“A skill is a file you load into Claude that teaches it to do one job exactly your way.”
★★ Adil's shot-list skill packs what the team learned making its last film (how Seedance 2.0 wants a shot described, what holds or breaks character consistency) and is shared in the description. — 00:12:46 · Seedance 2.0 · also in: The 3-Step Realistic AI Ad Workflow: The shot list, and the Claude skill behind it
★★★ Adil gives the shot-list skill the script and every locked asset uploaded as images, not described in words, which he says is why most people's prompts come out generic. — 00:13:15 · Claude · also in: The 3-Step Realistic AI Ad Workflow: The shot list, and the Claude skill behind it
★★★ The skill's shot list is one connected document: a style prefix (lighting, camera, colour) glued to every prompt changes everywhere at once, and each prompt has a name so Claude can edit one while the rest stay locked. — 00:14:17 · Claude · also in: The 3-Step Realistic AI Ad Workflow: The shot list, and the Claude skill behind it

00:14:46 — Scene 1, the kitchen: Elements, the lighting fix, the coffee montage and the dancing exit

★★ Adil saves every locked asset as an Element in his project under the name he gave it in Claude, so pasting a shot-list prompt attaches the required images automatically. — 00:15:13 · Elements, Claude · also in: The 3-Step Realistic AI Ad Workflow: Scene 1 — the kitchen
★★ When the kitchen shot crammed the coffee making into its end, Adil gave the coffee its own prompt, because the one shot was doing too much. — 00:15:39
★★ The moka pot and mug came out different in every take, which a tight close-up shows first, so Adil locks each with one clean reference. — 00:17:22 · also in: The 3-Step Realistic AI Ad Workflow: Scene 1 — the kitchen
★★★ Adil says "he dances" or a generic description of motion means nothing to Seedance 2.0: the dance must be spelled out move by move. — 00:19:16 · Seedance 2.0 · also in: The 3-Step Realistic AI Ad Workflow: Scene 1 — the kitchen

00:20:46 — Scene 2, the stadium: dry and wet sheets, a lighting override, the match cut and the body-rig shot

★★ Scene one of the headphones ad took three prompts and nine generations, with keeper phases cut from nine 15-second clips. — 00:20:46 · also in: The 3-Step Realistic AI Ad Workflow: Scene 1 — the kitchen
★★ Because he stitches slices from different takes, Adil locks the hero's athletic outfit as its own reference; otherwise every generation gives different shorts or shoes. — 00:21:16 · character sheet · also in: The 3-Step Realistic AI Ad Workflow: Scene 2 — the stadium
★★★ Adil's rule: when a character must change within a scene (dry, then soaked after the sprint), build a second reference sheet rather than fight one, because images are cheap and videos are not. — 00:22:41 · character sheet · also in: The 3-Step Realistic AI Ad Workflow: Scene 2 — the stadium
★★ Adil treats the style prefix as a default and overrides it for a single scene that needs its own look, such as the harsh midday light of the stadium. — 00:23:08 · also in: The 3-Step Realistic AI Ad Workflow: Scene 2 — the stadium
★★ To join scenes, Adil has Claude match each scene's opening ear-cup tap exactly to the closing tap of the previous one, same hand and same motion, so they match-cut cleanly. — 00:23:58 · Claude
★★ For a pure product shot Adil prompts a body-rig snorricam locked on the right ear cup: the headphones dead centre while everything behind rips past in motion blur. — 00:25:06 · also in: The 3-Step Realistic AI Ad Workflow: Scene 2 — the stadium
★★ Adil says a body-rig shot normally means a rig, a camera operator, a day on location and a few thousand dollars before recording; here he generated it. — 00:25:06 · also in: The 3-Step Realistic AI Ad Workflow: Scene 2 — the stadium
★ The stadium scene of the headphones ad took three prompts, with six videos pulled into the cut. — 00:25:06

00:25:48 — Scene 3, the street: the layout map, cut-by-cut prompts and choreography to the music track

★★★ When the hero's direction and the sky dancer's size and spot changed every take, Adil gave the model a map: text cannot pin a location down; he calls it the number one hack of the video. — 00:26:28 · also in: The 3-Step Realistic AI Ad Workflow: Scene 3 — the street, and the layout map trick
★★★ Instead of brute-forcing twenty generations, Adil builds a schematic map in GPT Image 2.0 (the fire hydrant marked, the sky dancer locked to its right at twice a person's height), so the model places things right take after take. — 00:26:51 · GPT Image 2 · also in: The 3-Step Realistic AI Ad Workflow: Scene 3 — the street, and the layout map trick
★★ When the hero kept drifting around the street between takes, Adil pinned him to a tree on the left of the frame as an anchor. — 00:27:53 · also in: The 3-Step Realistic AI Ad Workflow: Scene 3 — the street, and the layout map trick
★★ When some batches dropped the hero's backpack, Adil wrote into the prompt that it stays on both shoulders in every cut, which fixed it. — 00:29:06 · also in: The 3-Step Realistic AI Ad Workflow: Scene 3 — the street, and the layout map trick
★★★ Adil adds his music track to the Seedance 2.0 prompt as an input and asks for the hero to dance in time with its beat, so every move lands on the beat. — 00:29:56 · Seedance 2.0 · also in: The 3-Step Realistic AI Ad Workflow: Scene 3 — the street, and the layout map trick
★★ The street scene, six videos in the cut, came from iterating one prompt three times rather than fighting three separate prompts. — 00:30:25

00:31:16 — Scene 4, the office, and scene 5, the boss dancing by the sky dancer

00:33:31 — The full commercial played top to bottom

00:34:28 — Recap: the workflow and iteration as the skill

★★★ Adil says the finished ad is the best few seconds of about a hundred tries cut together, and that iteration is the skill. — 00:34:28 · also in: The 3-Step Realistic AI Ad Workflow: The full commercial, and the recap

“Iteration is the skill.”
★ Adil says the ad workflow is the same for any product, not just headphones. — 00:34:28 · also in: The 3-Step Realistic AI Ad Workflow: The full commercial, and the recap

In the description and comments

★★ Higgsfield's description says Seedance 2.0 4K has just launched and that the headphones commercial was made with it in 4K on a laptop. — video description · Seedance 2.0
★★ Higgsfield claims Seedance 2.0 now renders in 4K, so the footage comes out broadcast-grade with no upscaling and no soft AI-slop look. — video description · Seedance 2.0, AI look

Threads in this video

milestone · Seedance on Higgsfield — Higgsfield's description says Seedance 2.0 4K has just launched and that the headphones commercial was made with it in 4K on a laptop. (video description)
restate · Which image model should be used for keyframes, edits, text and outfits? — For the hero of the ad Adil uses Soul Cinema, which he calls the best model for photorealistic looks. (00:02:59)
refine · Which image model should be used for keyframes, edits, text and outfits? — Adil calls GPT Image 2.0 the best model right now for product editing. (00:01:43)
restate · How do you keep a location, product or prop the same across shots? — With one photo of the headphones, Adil has GPT Image 2.0 make a product sheet with front and three-quarter views, since the video model must see the product from every angle or it hallucinates mid-scene. (00:01:43)
refine · How do you keep a location, product or prop the same across shots? — Instead of brute-forcing twenty generations, Adil builds a schematic map in GPT Image 2.0 (the fire hydrant marked, the sky dancer locked to its right at twice a person's height), so the model places things right take after take. (00:26:51)
refine · How do you keep a character the same across shots? — Adil's fix for face drift: in GPT Image 2.0, erase the face from the full-body panel of the character sheet, leaving one face to lock onto; he does the same for the boss. (00:08:13)
restate · How do you keep a character the same across shots? — Adil's rule: when a character must change within a scene (dry, then soaked after the sprint), build a second reference sheet rather than fight one, because images are cheap and videos are not. (00:22:41)
restate · How do you avoid the AI (plastic) look? — Adil says the locations make or break the video: if a location looks fake or plasticky, no prompting will save it. (00:05:10)
refine · How do you avoid the AI (plastic) look? — To keep an edited asset real, Adil layers the original Soul Cinema shot over the GPT Image 2 edit in any photo editor and masks out the old outfit, so only the outfit comes from the edit. (00:11:19)
refine · How do you spend fewer credits? — Adil says testing every location in video only feels slow: it avoids wasting many credits later on full scenes generated with the wrong assets. (00:09:50)
restate · Which LLM should be used to write video prompts? — Adil never writes video prompts by hand: a skill he built in Claude writes them. (00:12:19)
refine · How do you direct a multi-shot sequence? — The skill's shot list is one connected document: a style prefix (lighting, camera, colour) glued to every prompt changes everywhere at once, and each prompt has a name so Claude can edit one while the rest stay locked. (00:14:17)
restate · How do you direct a multi-shot sequence? — When the kitchen shot crammed the coffee making into its end, Adil gave the coffee its own prompt, because the one shot was doing too much. (00:15:39)
restate · How do you get precise motion (dance, action, choreography)? — Adil says "he dances" or a generic description of motion means nothing to Seedance 2.0: the dance must be spelled out move by move. (00:19:16)
refine · How do you get precise motion (dance, action, choreography)? — Adil adds his music track to the Seedance 2.0 prompt as an input and asks for the hero to dance in time with its beat, so every move lands on the beat. (00:29:56)
restate · When a generation fails, should you generate more takes, fix the prompt or start over? — Adil says the finished ad is the best few seconds of about a hundred tries cut together, and that iteration is the skill. (00:34:28)
refine · When a generation fails, should you generate more takes, fix the prompt or start over? — The street scene, six videos in the cut, came from iterating one prompt three times rather than fighting three separate prompts. (00:30:25)
refine · How do you make a transition between two shots? — To join scenes, Adil has Claude match each scene's opening ear-cup tap exactly to the closing tap of the previous one, same hand and same motion, so they match-cut cleanly. (00:23:58)
restate · How do you put a new outfit on a person in a photo? — To keep an edited asset real, Adil layers the original Soul Cinema shot over the GPT Image 2 edit in any photo editor and masks out the old outfit, so only the outfit comes from the edit. (00:11:19)

Where the stories that moved here went next

Seedance on Higgsfield — running; last: milestone, 2026-10-22

Mentioned

Central: character sheet · Claude · GPT Image 2 · Headphones ad (Higgsfield, 2026) · Map technique · Seedance 2.0 · Seedance shotlist Claude Skill (headphones ad) · Soul Cinema · style prefix

Also: AI look · Canvas · Cast · Cinematic Locations · Elements · product sheet

Also centred on Headphones ad (Higgsfield, 2026): 3-Step Workflow To Make Ultra-Realistic AI Ads · The 3-Step Realistic AI Ad Workflow: The finished ad, and the 3-step system · The 3-Step Realistic AI Ad Workflow: Locking the first location: the kitchen · The 3-Step Realistic AI Ad Workflow: Scene 3 — the street, and the layout map trick · The 3-Step Realistic AI Ad Workflow: The full commercial, and the recap · The 3-Step Realistic AI Ad Workflow: The product sheet (GPT Image 2.0)

Also centred on Map technique: How To Make Cinematic Football Ads With AI (Full Workflow) · I Made a Cinematic Ad Using Claude Fable 5 + Higgsfield AI (Full Workflow) · The 3-Step Realistic AI Ad Workflow: Scene 3 — the street, and the layout map trick · Make a Cinematic Ad End-to-End: Scenes 5-6: The Duel & Round One Loss · AI Got the Seating Wrong So I Drew a Map

Also centred on Seedance shotlist Claude Skill (headphones ad): 3-Step Workflow To Make Ultra-Realistic AI Ads · The 3-Step Realistic AI Ad Workflow: The finished ad, and the 3-step system · The 3-Step Realistic AI Ad Workflow: The shot list, and the Claude skill behind it

Also centred on style prefix: How to Make AI Films That Don't Look AI (Full Tutorial)

← Previous: AI Space Fleet: 4K vs 1080p Resolution Test! 🌌🛸 #shorts #scifi · Next → Black Hole Warp: 4K vs 1080p 🌀🚀 #shorts #space