The making of Myanmar Glam

One person, a home GPU, a few rented ones, open generative models and an agentic workflow. About a week.

Making-of film and notes · October 2026

The making-of, in HLG HDR. Narrated in a clone of the director's own voice, generated locally.

Best on an HDR screen: iPhone, iPad or a recent Mac in Safari, or Chrome on an HDR display. Other screens get a standard-range version automatically. The stills on this page are HDR too.

The point

Every frame of the couture film was generated. No camera, no set, no model on a runway, and no proprietary generation API: the images, the video, the HDR conversion and this narration all came from open models running on GPUs we control.

How far can one person push local generative AI, with an agentic workflow, a custom pipeline and tools built in-house?

Title card: how far can one person push local generative AI, over a dim HDR shot of the model — The question behind the project, and a frame from the film it produced.

The model in a gold beaded outfit at a candlelit vanity — The question behind the project, and a frame from the film it produced.

The system

Two tools hold it together. Pocket AI is where a shot gets built: references, prompt, settings. Tensor Master is the control room: it queues every job, ranks it, sends it to the home GPU or to rented cloud GPUs when the queue gets long, and keeps every take re-runnable from the gallery. An agentic workflow sits on top, writing the briefs, checking every take and sending the failures back to the queue.

Pocket AIbuild the shot · refs · prompt · settings

↓

Tensor Masterqueue · priorities · agentic QC · cost guard

↙ ↘

Home GPURTX 3090 · always on

Cloud GPUsrented RTX 5090s · by the hour

↘ ↙

Galleryevery take · edit & rerun

Same open models, wherever a job runs.

By the numbers

1,073jobs through Tensor Master

6days, start to final takes

426MiniMax H3 renders

69final takes, 8 looks

18rented cloud GPUs

$35cloud compute

GPU rendering time, 37 hours in all

home RTX 3090 · 30.1 hrented · 6.7 h

Diagram from the film: Pocket AI to Tensor Master to home and cloud GPUs to the gallery — Scale up when the queue grows, shut it all down when it is done.

Stats card from the film: 1,073 jobs, 6 days, 37 GPU hours, 18 rented GPUs, 35 dollars — Scale up when the queue grows, shut it all down when it is done.

The pipeline

Qwen Image 2.1Reference images: a rear view of every garment, location plates, a character sheet for the model.
MiniMax H3Video in reference-to-video mode: 768×1344, 24 fps, 4-second takes.
LTX-2.5 SDR → HDRAn ACES workflow that rebuilds the dynamic range at the takes' native size.
RIFE 4.9Frame interpolation to 48 fps.
Grade and deliveryPer-look 3D LUTs; Dolby Vision profile 8.4 plus HLG.
MiniMax Music3Original score for the earlier cuts and for this making-of.
CapCut AI EditThe final cut of the film, including its soundtrack.

The seven pipeline steps listed over an HDR shot

End card on the jetty at dusk: one person, open models, own GPUs

What broke

The back of the dress. The model invented zippers, or put the front buttons down her spine, until it was shown every garment from behind.
Shiny, flash-lit faces. The fix was to light her like a real set: soft light from the side, shadows left alone.
Moiré in HDR. An upscale slipped into the first conversion and turned fine lace into a rippling pattern. Every take was reconverted at its native size.
Takes that look fine at a glance. A second actress, a hidden cut, odd hands. The agentic QC pass catches these and re-queues them, so the human review only sees what survives.

Before and after: front buttons wrongly drawn on the back of a red dress, then the correct one-piece back — The back of the dress, fixed; moiré from an upscale against the native-size conversion.

Zoom on lace: upscaled version with moiré above, native-size version clean below — The back of the dress, fixed; moiré from an upscale against the native-size conversion.

Made with Pocket AI and Tensor Master, open-weight models on a home RTX 3090 and rented RTX 5090s, and an agentic workflow. Narration: Qwen3-TTS, cloned from the director's voice. Film soundtrack from CapCut; making-of score by MiniMax Music3.