DREW

THE WIKI · 04 / 09 · 3 MIN READ · MARKDOWN

Stills in GPT Image 2.5

How the stills are made in GPT Image 2.5: face refs where they build or fix him, the world, and the fixed frames the hardest moving shots begin and land on.

The stills in the D.R.E.W film were made with GPT Image 2.5, on Higgsfield. The stills that build Drew from scratch or fix his hair carry his locked face and hair refs; stills that edit an earlier frame start from that frame, and the garage stills used only the garage and car images. They do three jobs: they fix his face, they fix the world, and they give the hardest moving shots a first and a last frame. The generation log records 11 still steps, each with its exact prompt and inputs; they are on the D.R.E.W film page under the shot they fed. Claudia's equivalent page, Plates in Midjourney, is worth reading for how stills get picked.

Order of work

References first. Before any video, these were made as stills:

  1. His face, from two angles, plus a tied-back hair ref.
  2. Claudia's face, and a ref of her in the orange tee (she is a guest; her canon is on claudia.gallery).
  3. The garage.
  4. The orange car.
  5. The rooftop sign.

Then the jewelry: the pendant and the watch each had a still of their own. Only then do start frames and video begin.

His face

The face is locked by references, not by long descriptions:

  • Face front and face side. Two clean portraits, straight to camera and a side view. They are attached to the stills that build him and to the video prompts.
  • A face sheet. Seven heads on one pale blue sheet, in black and white with only the blue loc in colour: front, three-quarters, eyes closed, mouth open, two profiles, laughing. The colour on the loc alone makes the one constant that matters impossible to miss. The face sheet was only ever a reference, never a production image.
  • One loc, at the front left of his hairline. Pick it, and use it everywhere.

Write the identity text into every prompt.

His world

Each place gets one still that every shot in it refers back to:

  • the garage: concrete, strip lights, wet floor, four supercars in a row, the giant iced moon
  • the hotel entrance and its marble lobby
  • the rooftop sign: his name in marquee bulbs on a steel frame over the skyline

No text, no logos, no tickers in any of it, except the lettering the film shows on purpose: his name on the roof.

Fix the still, then make the video

Every still step in the log is a fix, and each fed one shot:

  • Cars: a door-closed still and a door-open still of the orange P1 (shot 10), and a door fix on a hotel frame so only the driver door is raised (shot 15). See Cars.
  • Placing people: Drew on the sidewalk side of the car (shot 16), Claudia placed in the passenger seat and then a face fix so she matches her canon (shot 21).
  • Where he looks: an edited still where he watches the road, so he stops turning to the camera while driving (shot 17).
  • Swaps: the silver W1 swapped into the street frame (shot 18), and a hair fix on the ledge frame (shot 20).
  • A place: the hotel lobby, made from his face and hair refs (shot 14).

For a shot that has to land exactly (a door opening, a hand on a key), make both ends as stills and give the video model the pair. claudia.gallery's A first and a last frame shows the same method from the other side. Make the pair from the same scene still, changing only the one thing that moves.

The film look

Every still prompt ends in the film look: real 35mm film photo, natural grain, not glossy or CGI. A clean, glossy still makes a clean, glossy clip.

Picking

Make several, pick one, and check it before it feeds a video:

  • one blue loc at the front left, the cuff, a dense full beard, hazel-green eyes, an adult face
  • the jewelry the shot needs, matching its reference
  • no text anywhere
  • five fingers per hand