Bezar Video Studio

Wax Wrap — AI Pipeline Test

First result of the pivot away from in-house 3D to pure AI image + video generation (Higgsfield / Fal / Kling), with the engine harvested from ChapterCut. Goal of this pass: prove we can produce a consistent product through all install stages and animate it — fast and cheap.

Real footage → OpenAI polish → AI image/video Studio product-viz look ~$8 total · ~1 evening Overnight autonomous build
What changed: the 3D approach was too error-prone for the effort. This pipeline starts from a real frame of the actual fitting, has OpenAI isolate & polish it into a clean hero plate, then evolves that one plate through every install stage so it stays the same fitting. No modeling, no sims — and it looks like the real product because it started as the real product.

The five-stage sequence (one fitting, evolving)

1bare cast tee + threaded steel pipes
Surface Prep
Cast-iron tee + threaded steel pipes
2primer
Wax Primer
Translucent amber coat
3putty
Profile Putty
Grey fill at the joints
4partial wrap
Wax Wrap (applying)
Tan tape over the joint
5finished
Finished
Full wax-wrap barrier

Motion — the sequence animated

Five 5-second AI clips (Kling v2.1 Pro, image→video), one per stage, cut together. Static locked-off camera + slow push-in to keep the product stable.

The very first proof clip — finished hero plate animated. Geometry held through the move (the consistency test that motivated the whole pivot).

How it's made (the recipe)

  1. Real frame → bare master. A real close-up of the actual fitting goes to OpenAI gpt-image-1, which isolates it on a clean studio background — giving the correct subject: a cast-iron tee with three galvanized pipes threaded into it. Gotcha: generators default to symmetry (kept trying to add a phantom 4th leg → a cross).
  2. Masked inpainting locks it. We build an edit mask from the bare plate — only the fitting is editable, everything else (background, the area below the body) is frozen. So the model physically cannot add a 4th leg or shift the framing.
  3. Bare → each stage. Every stage is generated independently from the one bare plate through that mask (amber primer, grey putty, partial wrap, full wrap). Result: all five share identical framing + correct geometry.
  4. Stage stills → motion. Kling v2.1 Pro image→video, locked camera + slow push-in.
  5. Assemble. FFmpeg cut, 1080p. (Next: Bryan VO + callouts.)

Honest assessment

Working

  • Correct subject — cast-iron tee with three threaded galvanized pipes (per the real part).
  • Identical framing + geometry across all five stages (masked edits lock it).
  • Looks like the real product (started as a real frame).
  • Fast & cheap — a full sequence in an evening, vs days of 3D. Geometry holds in motion.

Still to tighten

  • Primer color reads copper/amber rather than a subtle translucent honey — the one material still worth a pass.
  • Background is a warm cream; can standardize to neutral grey if preferred.
  • The push-in zoom varies a little per clip; can lock to one move.
  • Fine thread/seam detail isn't load-bearing yet (precision not required at this stage, per direction).

Where this goes next

The pivot is validated: AI gen gives us a believable, consistent product far faster than 3D. Next pass: (1) lock framing/background to one studio plate via masked edits; (2) train a Higgsfield Soul ID on the fitting so we can render new camera angles that stay identical; (3) build the real shot list from the storyboard with Bryan VO + animated callouts; (4) the storyboard ("cinematic") look as an opening/closing bookend. All keys + tooling are in place.

Bezar Video Studio · Polyguard Wax Wrap · AI pipeline test · engine harvested from ChapterCut · OpenAI + Flux Kontext + Kling.