Experiment 02 · Canvas 02 · KREA2 t2i + tldraw draw

Generate + Draw

Generate a photorealistic art gallery with KREA2, then hand-draw the same scene beside it using tldraw's vector shapes with converging perspective.

Prompts

Jon → Hermes (spoken input)
"Generate an image of a high-end art gallery interior with modern art sculptures, Jeff Koons balloon dog, interesting architecture. Then draw the same image next to it on the tldraw canvas." Later correction: "You gotta draw the perspective. Do the perspective lines so we see the floor, the far wall. Follow those lines of the room. Make it a three-dimensional drawing, like a sketch."
Hermes → ComfyUI (KREA2 turbo t2i)
A high-end contemporary art gallery interior, large open space with white walls, polished concrete floors, skylight natural lighting, large sculptures including a Jeff Koons balloon dog, abstract modern art installations, bronze figures on pedestals, architectural details, wide angle photography, 8k Model: KREA2 turbo fp8 · Steps: 8 · CFG: 1

Generated Gallery Image

Generated art gallery interior
KREA2 turbo t2i — photorealistic art gallery with Koons balloon dog, bronze figures, star sculpture, mirror sphere, and skylight lighting

tldraw Canvas — GUI Export

tldraw canvas02 GUI export showing gallery image and perspective drawing
Canvas 02 exported from tldraw GUI — the generated gallery image on the left, the hand-drawn perspective sketch on the right, with text labels. This is what the tldraw desktop app looks like during the workflow.

What Happened

Generation: The gallery image was generated in 38 seconds using KREA2 turbo text-to-image. The prompt called for a Jeff Koons balloon dog, modern art sculptures, and interesting architecture — the model delivered a convincing gallery space with a skylight, track lighting, a reflective balloon dog, a bronze figure on a pedestal, a colorful star sculpture, and a large mirror sphere.

Vision analysis: Before drawing, the agent used a vision model to analyze the generated image in precise spatial detail — identifying the room shape, vanishing point, each sculpture's position and size, the skylight angle, the light beam direction, and the architectural cutout in the back wall. This analysis guided the drawing.

First drawing attempt — flat: 34 geometric shapes (rectangles, ellipses, stars) placed as flat 2D blocks. Jon immediately said: "You gotta draw the perspective. Do the perspective lines so we see the floor, the far wall."

Second drawing attempt — perspective: 42 shapes with actual depth. The floor became a trapezoid narrowing toward the back wall. Left and right walls receded to a vanishing point. Pedestals were drawn as trapezoids (narrower at the back) to show perspective. Freehand draw shapes used compressLegacySegments for precise polyline paths — the skylight, light beam, track lights, and sculptures all placed in 3D space.

The result reads as a pen sketch — not photorealistic, but spatially coherent. The limitation: tldraw's draw shapes are uniform-stroke polylines with no fills or smooth curves, so the balloon dog (which has tube-like curves) is the weakest element, represented as overlapping circles. But the room perspective, the converging lines, and the sculpture placement clearly convey the 3D space.