← Lab

PILOT 2.5D

Made with generative AI music: MiniMax-Music3 video: MiniMax H3 ErosLoRA: Licon VBVR I2Vsigmas: HyperFlowmask: BiRefNet-HR3D: SAM 3D Bodyscene: Blender 5.2

How it was made: the ComfyUI clip, the Blender scene and the final effect

PILOT 2.5D is a 45-second test of a method from a creator who posts as @aicreataro: the AI clip comes first, then a world is built around the character in Blender, the 3D program. 2.5D means flat images set in 3D space for the camera to fly between.

A young man does parkour through subway cars at night. Around him the car breaks into tiles that pop toward the camera on the beat, alternating with worlds built from the clip’s frames. Big, colorful lyrics land on the floor and walls. At times he becomes a 3D body of cubes, which the camera circles before the body explodes.

How it was made

I brought the method in after WIRE ME UP, where events kept landing late against a Blender animatic (a simplified 3D animation) meant to guide the video model. the Claude Code agent wrote the scripts and prompts and built the Blender scene; I judged each version and decided what stayed.

  • Version one was 10 seconds of the WIRE ME UP chorus; version two, 15 seconds on a new track, looked like a kaleidoscope to me. Comparing it with the original, the agent saw why: there the world comes from the clip’s own frames, ours from unrelated images.
  • For the final cut the agent wrote the Music 3 lyrics and description; mixing Polish, Japanese, Korean, Russian and English words was my idea, and I picked the take.
  • The clip came from MiniMax H3, a video model, on my ComfyUI workflow: five pieces chained into one shot, each taking about 4 minutes to render. Long prompts full of don’ts gave a jog, so the agent cut them to one action sentence. I pushed for energy: use the pipes and rails, run like a Korean action hero.
  • A mask, an image separating figure from background, cuts the runner out. When I noted he was flat, unlike in the original, the agent added SAM 3D Body, a model that rebuilds a body mesh from video.
  • I dropped the karaoke captions and box jumps, suggesting a glowing stick figure between neon bars and an opening of words on black instead. When I asked about camera freedom, the agent proposed and built a 3D subway car wrapped in a panorama of clip frames.

What it shows

  • Build the world from the same footage. Frames of the same car, cut into tiles, a dot mosaic and falling cards, keep the perspective and match the runner.
  • Perspective has limits. The 3D body works from any angle. A one-frame world holds up to a 30–40° camera turn; full freedom takes a car built in 3D.

Who did what

Concept, choice of method, judging each version, the song's seed
Marcin: the idea, based on the method by @aicreataro, rejecting several variants, approving the final version
Scripts, Blender scene, prompts, lyrics and song description
Claude Code

Tools

  • ComfyUI (local)
  • MiniMax Music 3 (the song, with a collage of languages in the lyrics)
  • MiniMax H3 (Eros variant) with the Licon VBVR I2V LoRA, HyperFlow sigmas, shots chained from 5 pieces
  • BiRefNet-HR (runner mask)
  • SAM 3D Body with MoGe 2 (3D body mesh and skeleton)
  • SAM 3 (pipes and handrails)
  • Blender 5.2, Eevee (the 2.5D world)
  • ffmpeg and Python (edit)