Nana's Big Ass Buns

Nana's Big Ass Buns is a 44 second vertical comedy spot I made as a test piece, no client, no brief, just a question I wanted answered. I'd seen a reel that had me laughing for a good minute, and I wanted to see whether we could take a handwritten script and push it through a fully AI pipeline without losing the joke. The premise is a diss track for bakers. An 84 year old retired baker in a maroon tracksuit and oversized shades looks straight down the lens and calmly tells every bakery in the ends she's coming for their customers. Then a montage: the training silhouette, the squat, the gold ring, the pensioner entourage walking out of a rainy London square, the velvet wheelchair with the catheter bag, and a title card. Doors open Monday. Tell your doctor.

The whole thing lives or dies on one contrast: a very old woman with a very gangster attitude. If Nana drifts even slightly between shots, if the hair changes or the chains disappear or the face softens, the audience stops believing her and the joke collapses. So the real test wasn't "can AI make this look nice," it was "can AI hold a single comedic character across a dozen shots, three locations and a full monologue, working from a script written before any image existed." Everything else, the look, the pacing, the type, was in service of that.

This was the first project where I properly built a character reference sheet before generating a single shot. Front, three quarter, side and back, an expression progression, micro expressions, head angles, hand gestures, wardrobe details down to the sneakers and the rope chains, all on one board. That sheet went into every generation as reference. The name on it, Dot "OG Dottie" Harmon, and the accent notes never made it into the film, we just called her Dottie on the canvas. The sheet is a visual anchor, not a bible. Even with it, a few generations still drifted, so the discipline was in checking the details every time, hair volume, frame size, chain count, before moving on.

I gave the wheelchair the same treatment. It has its own product sheet, the Sovereign Mobility Chair SV-12, with front, side, three quarter and back views, an exploded diagram, material swatches and callouts for the brass hub, the caster cap and the drainage bag. From one hero image I ran a nine camera grid to get coverage, then pulled the detail shots for the montage off that.

The pipeline was LibTV. Lib Image did most of the stills, with Seedream 5.0 Pro for cleanup passes and Topaz for upscaling. For motion, Kling O3 and Happy Horse 1.1 handled the bulk of the montage, and I used Seedance 2.0 for the three hero dialogue shots in the bakery where Nana has to talk. Because the script and the framing were locked before I generated anything, I didn't need to spray and pray. Each shot got its reference frame right first, then went to video. That's why the keep rate landed around 70 percent, with roughly 60 image generations and 36 video generations for the whole film.

Her voice was designed in ElevenLabs v3 from a written brief: an 84 year old British woman with a working class London accent, slightly raspy and thin, natural age wobble, calm, cool, quietly cocky, not frail, not sweet. The script went in with emotion tags on every line, calm for the opening, smug for "the biggest bakery in Britain," a whisper for "tell your doctor," and timed breaks between beats so the edit had room to breathe.

The look is boxer promo grammar. Warm tungsten in the bakery for the monologue, cold grey daylight for the exteriors, black and white title cards, and red kinetic type over her lines. I cut and graded in Premiere Pro, and re-used the track through the montage and end card so the whole thing holds together as one piece rather than a string of clips.

It works. Nana stays Nana in every shot, the monologue lands, and the montage earns the title card. It's sitting internally for now as a proof of concept, we'll put it out on socials when the time is right. What carries forward is the order of operations: script first, character sheet second, lock the frame before you generate, and treat the sheet as a reference not a rulebook. That's the workflow I've used on every character-led piece since.

Format

9:16, 44s vertical comedy spot

Generations

~60 image, 36 video, ~70% keep

Timeline

4 working days

Platform

LibTV (Lib Image, Seedream 5.0 Pro, Kling O3, Happy Horse 1.1, Seedance 2.0, Topaz)

Voice

ElevenLabs v3

Edit and grade

Premiere Pro

Team

solo, Kenneth Ong (@kennematic)

Status

unreleased, internal