IMPERIUM

KAHAYA ×

What Higgsfield's AI rendering does with the existing Kahaya render set. Click any image for full size, click again to zoom.

Every image is AI-generated in a single one-shot prompt from the original renders. This is an experiment across models, not final marketing material: each result can be tweaked and iterated toward a desired outcome.

Stills

New angles, light and settings from the same render set. Architecture checked against the master plan.

Aerial with the bay
The compound and the real bay. GPT Image 2
Blue hour
Blue hour. Generated from the daytime frame, not a separate render. Nano Banana Pro
Sunset over the bay
Sunset over the bay. GPT Image 2
Misty morning surroundings
Misty morning, real countryside. GPT Image 2
From the beach
From the beach, compound inland behind the grove. Nano Banana Pro

Before / after

Drag the handle. Left of the line is the original, right is the edit. All three pairs are the same frame at the same pixels.

Real surroundings edit
Original render
OriginalReal surroundings

Surroundings replaced with muted, hazy Kuta countryside at matching perspective. The compound is untouched: same camera, same buildings. Nano Banana Pro.

Villa removed
Villa circled
Circled villaRemoved

The circled villa removed: masked out and rebuilt by the model as landscaped garden, the three neighbours left as an L, everything else untouched. Drag right to reveal the circled original. Close-up below.

Close-up, villa circled
Close-up before. Target circled.
Close-up, villa removed
Close-up after. Villa gone, garden in its place, neighbours untouched.
Thatch roofs
Original roofs
ShingleAlang-alang thatch

Whole-site material change: every roof re-specified as thatch, identical ridge lines, solar array kept.

Models compared

Seven models, same source frame, same instruction, shown untouched for comparison. Where the tests landed after review: GPT Image 2 for scenic realism (the bay, sunset and misty stills above), Nano Banana Pro when the frame must stay pixel-exact and for targeted edits. Seedream renders at 5K but its output reads noticeably more AI, so it is kept here for comparison only.

Seedream 4.5
Seedream 4.5 · holds geometry, best photo feel, 5K
Nano Banana Pro
Nano Banana Pro · pixel-exact, best instruction following
Nano Banana 2
Nano Banana 2 · close, sometimes invents terrain
FLUX.2
FLUX.2 [max] · atmospheric, overdoes ground wear
Seedream 5 Pro
Seedream 5.0 Pro · faithful, stays render-like
Hazel
OpenAI Hazel · reframes and redraws, not reliable here
Kling O1
Kling O1 · barely changes, stays CGI

Plan graphics

The master plan redrawn as an exploded axonometric: parallel projection, road along the bottom as drawn, buildings colour-coded by unit type.

Exploded axonometric
Exploded axonometric. Nano Banana Pro, from the master plan + one aerial

Video

Camera moves generated from the stills above. Use the fullscreen control on each player.

Orbit · Kling 3.0 pro, 10s, from the photographic aerial

Rooftop pass to the sunset · Seedance 2.0, 10s

Beach approach · Cinema Studio 3.0, 10s

What Higgsfield can do with this render set

  • Re-photograph a render with real surroundings, same buildings, up to 4K output
  • Change light and time of day on the same frame: dusk, night, sunset
  • Edit a marked target: remove or change one building, wall or material, rest untouched
  • Whole-site material studies: roofs, finishes, landscaping in one pass
  • New angles anchored to references: beach level, high aerial, rooftop height
  • Plan graphics: exploded axonometrics and diagram styles from the master plan
  • 10-second camera moves from any approved still: orbit, pass, approach
  • Upscaling to 4K+ available for print or large-format use

Recommended pairing from this experiment: GPT Image 2 for scenic and photographic work, Nano Banana Pro for pixel-exact edits and targeted changes, Kling 3.0 pro and Seedance 2.0 for motion. Everything on this page is one prompt, one shot: with iteration each result gets sharper, cleaner and closer to a specific brief.

׋›