Alessio Bevilacqua

Selected work

13th Floor: FW Drop

Spec campaign film · 2026

I pitched 13th Floor a late-'90s editorial for their bomber and denim, shot like it was on film: five cities, five people, one look. The owner signed off, and I built every character and city as an element so the product and the grade hold in every shot.

Seedance 2.5Soul CinemaSeedream 5.0 ProFinal Cut Pro37 generations

See how it was made ↓

The Last Fig

Action-comedy short · 2026

Four guys steal a fig, cut like a Nolan heist: establishing wides, a binocular POV, extreme close-ups, over-the-shoulders. I wrote it, blocked it with Blender previz and staging diagrams, generated it in Higgsfield and cut it in Final Cut.

Seedance 2.5Seedream 5.0 ProSoul 2.0Nano Banana Pro148 generations

See the process on Higgsfield →

The Tow Father

Branded spot, vertical · 2026

A Sopranos-style mob parody for a real towing company, built from the owner's creative direction. The lead stays the same man across about 25 shots and four locations. It ran on the client's socials and passed 40K views on their TikTok.

Cinema StudioSeedance 2.0Higgsfield CLI + skills80 generations

See the process on Higgsfield →

Also: Spider-Man: RAF Vol.1, fan film, sound design & foley (682K views)

Ego Death, music video: writer, sound design, artist; assisted on the edit

How I work

Every film runs through the same pipeline. Here it is, with the proof.

Claude Code drives my own local studio on the Higgsfield API, over an MCP server I wrote. Seedance 2.5 and Soul Cinema run through Higgsfield; the character sheets are Seedream 5.0 Pro.

Brief to locked plan

13th Floor started as a voice memo. It became a script where every decision is marked locked, agreed or open, and the look is picked before anything is generated.

“A high-end Vogue editorial commercial… a late '90s, 2000s kind of vibe. I want it shot on a film camera… old school grungy, but brought back in an expensive way.”From the voice memo, Sep 22

A look, turned into a recipe

I wrote the reference look (late-'90s fashion film on 16mm) as two locked lines of text: one closes every location prompt, one closes every video prompt. The grain and the crushed blacks held in every city with no lens words in the prompts.

late-1990s 16mm fashion-film texture: coarse moving film grain, crushed murky blacks, warm skin against cool cyan-green shadows, softly glowing practical lights, gentle gate weave, 24fps motion cadence, candid handheld energy

The video half of the recipe, as it closes every prompt.

Elements first

The product sheets are built from the client's real product photos, not AI. Every model and every city is saved as a tagged element before any video runs, so the zip, the patch and the grade come out right in every shot. The Toronto model is me, built from a selfie.

Reference sheet for the 13th Floor hooded workwear bomber — @prop_tf-bomber

Character sheet for the Tokyo model in the bomber and jeans — @char_TF_tokyofw

One generation, ten cuts

Every city is written in the same ten-cut order, so one 30-second generation gives a full pool of coverage and five cities cut together as one campaign. São Paulo and London were cut to seven for budget.

Ten frames from one New York generation: establishing, walk, sleeve detail, dice crew, dice, patch, slow shutter, knee pleats, front pose, back pose — New York, one generation: all ten cuts landed in order

Test, fix, write it down

Runs get a verdict. New York's back pose came out cropped at the shoulders, so I rewrote the camera line and Toronto came out head to heels. A fix only becomes a rule once a failed generation proves it.

Left: New York back pose cropped at the shoulders. Right: Toronto back pose, full length — Before: “three metres behind him … full length” / After: “five metres behind him … from the head to the heels, the ground showing”

Block it before paying for it

On The Last Fig I blocked the opening scenes in Blender with stand-in figures and gave Seedance one frame per shot. The staging held. Later scenes used staging diagrams instead.

Top row: three Blender previz frames with stand-in figures. Bottom row: the matching generated shots from The Last Fig — Top: Blender previz / Bottom: the generated shot

The system behind it

Studio

My own studio, agent-driven

A local studio on the Higgsfield API with an element library and a run log. I wrote its MCP server, so Claude Code can generate, save elements, pull runs and file them into projects on its own.

Claude Code skills take it from brief to locked script, to element sheets, to shot prompts in one fixed format.

Craft log

What Seedance actually does

16 settled rules. Everything else stays under test until a failed generation proves it. Three lessons from it:

It reads words literally.“Catch-lights blazing” gave glowing red eyes. Physical actions only.
It copies posed figures.Posed mannequins in a staging frame came back as men floating in those exact poses. Now: box figures.
Language lock goes at the top.With it only at the end of a long prompt, the English lines came back in another language.

Client loop

Brief, build, train, ship

On my own product, Recevia, I set up each client one on one: wrote their spec, trained the owner, took it live, then shipped what they ran into as the next release. 4,800+ calls answered for real businesses.

About

I make cinematic AI films and build the Claude Code pipelines behind them.

I came up in music as a songwriter and producer, and I've been making videos since I was 10. Eight years doing this professionally, the last two-plus hands-on with gen-AI.

I also co-run SelfieBooth, an event production company: 125+ events for clients including Scotiabank, BMO and CBRE, working with their event teams from brief to delivery.