
Ali Özgür
ABOUT
A 30-second café scene directed like a shoot, camera first: blocked in Blender, generated with Seedance 2.5, then edited, scored and graded in-house.
SELECTED FRAMES — 4 OF 8
PROCESS
33 café drafts at 480p, 26 s each (about 20,600 frames, 2,604 credits). I kept one take, finished at 1080p, and cannibalized the other 32 and my first grade. I graded against the original and three looks on stills from every shot and light state, and picked B. Composition is decided in Blender, not in the prompt. Every shot is blocked first (script, crude layout, concept art, storyboard, then previz) and the previz goes into the prompt as @Video 1, followed second by second. Character sheets and location and prop sheets lock the elements, lit in the film's own look so the model doesn't drift. Human touch. Claude generates the camera move in Blender, then drops me onto it. I operate only the pan and tilt by hand, with a phone and a gamepad, to soften the move and add a person's weight and small corrections. The move's position stays as generated, and the model inherits the feel from @Video 1. Audio, four layers: Cloned voice. The director is my own voice. I recorded one sentence on my phone and attached it to every generation as an audio reference, so the model speaks as me. Generated voice. The other character's voice came from a written description (Seed Audio on Higgsfield). The room tone and the shots' own sound come with the picture from Seedance. I lift the dialogue and patch out stray blips. Sound FX and foley. All code-built, no library: the rain bed, the passing car's whoosh, the explosion's sub layer, the glitch hits, the cash register on INSERT TOKEN, and the clicks and scrubs under the making-of. Each is placed in its space and distance. Music. Suno, in two pieces: the café's score, locked, and a caper take cut on its bars to the making-of, stopping on a slap and landing its last hit on the lens flare. Mixed as dialogue, music and effects stems to -16 LUFS. Where it falls short: the model takes the Blender camera as a guide, not a lock. A prompt holds only so many words, so I spend them on the actors and action, and the set and frame edges.


