The post
People are selling floor plan walkthroughs for thousands, so I wanted to see how far a 2D plan plus two models actually gets you. Here is the full recipe, with every step's output, and what I'd do differently. The video shows the plan with a position dot on the left and the walkthrough on the right, then the same route from one single image at the end.
**Step 1. Draw the plan**
GPT Image 2.6 made a clean top-down plan: entry, living room, kitchen, master bedroom, labels and door arcs. About 0.03.
**Step 2. Turn it into a 3D cutaway**
I fed the plan back into GPT Image 2.6 as a reference and asked for an isometric cutaway with the same layout and one set of materials (oak herringbone floor, warm white walls, one light direction). This image is the scene bible. About 0.04.
**Step 3. Render each room from it**
Using the cutaway as the reference again, I asked for eye-level photos of six viewpoints along the route. Same sofa, same marble island, same bedding in every frame. About 0.044 each at quality high.
**Step 4. Walk through with Seedance 2.5**
Each room render became the first frame of a 5 second image-to-video shot with a steadicam walking prompt. Six shots, about 1.5 each at 720p, 5 to 6 minutes when run in parallel.
**What went wrong first**
My first try used the next room as the last frame. Seedance cross-dissolved between the two, so doors turned see-through and stools doubled. Starting every shot from its own render and cutting between them fixed that. And with only one image, the same route (15 seconds, about 4.5) gets plausible but the camera has to guess the layout, so the rooms stop matching the plan.
|Step|Model|Cost|Time|
|:-|:-|:-|:-|
|Plan and cutaway|GPT Image 2.6|0.08|about 2 min|
|6 room renders|GPT Image 2.6|0.27|about 1 min each, parallel|
|6 walking shots|Seedance 2.5|9.08|5 to 6 min, parallel|
|Single image comparison|Seedance 2.5|4.52|5 min|
Prompts are in the comments. What would you add to make it feel more like a real show apartment tour?
View the original thread on Reddit →