Photoreal Blender scenes from a one-paragraph brief, built by AI agents: image→3D props on fal, scripted layout, lighting, camera and grade.
Behind-the-scenes videos — shared spec
One BTS video per scene (repair, ramen), about 60–75 s, showing how the scene was built: layout → generated props → placing them → materials and grime → lighting → post → final shot. The two videos must look like a pair, so both scenes follow this spec exactly.
Who makes what
- Scene agents (repair, ramen) render the four "dollhouse" clips below from their own scene files,
into
breakdown/<scene>/clips/. Nobody else knows your stage files and object collections. - Main session makes the concept, asset-generation, post-wipe and final-shot segments,
the captions, and assembles the video (
breakdown/assemble.py). Don't make captions or titles.
Output format for the dollhouse clips
- 1920×1080, 24 fps, PNG frames in
breakdown/<scene>/clips/<clip>/####.pngplus an mp4 of the same (h264_nvenc, any decent bitrate) atbreakdown/<scene>/clips/<clip>.mp4. No text, no grade, no letterbox. - Leave the bottom 140 px free of anything important (captions go there).
The dollhouse camera (same for all four clips)
One continuous orbit split into four clips, so they cut together as one move:
- A cutaway view of the whole interior from a high three-quarter angle (about 35–45° down), looking at the middle of the room. Hide the ceiling and the one or two walls nearest the camera at every angle, so the interior reads like a dollhouse. Hide anything that blocks the view for the whole orbit (exterior shell, awning, street) unless it is part of what that clip shows.
- Lens about 28–35 mm, f/8 or no DOF. Fit the whole interior with a little margin.
- Total orbit: 120° around the room's vertical axis, at constant angular speed, ease in on clip A only and ease out on clip D only. Clip A covers 0–30°, B 30–60°, C 60–90°, D 90–120°.
- Put the start angle where the room reads best (counter/bench in view all the way).
The four clips
| Clip | Frames | What it shows |
|---|---|---|
a_layout |
144 (6 s) | The empty shell as the layout: walls, floor, counter/bench blocks, shelves. Clay look: one flat light-grey material on everything (Workbench or Eevee with a soft studio light and AO). Optionally the shell pieces rise into place in the first 2 s. |
b_props |
240 (10 s) | Same clay look. The generated props appear one at a time in a sensible order (big furniture and shelves first, then big props, then small clutter), each scaling up from 0 with a small overshoot over ~6 frames. All props in place by frame ~200; hold after that. |
c_materials |
144 (6 s) | Switch from clay to the real textured materials with neutral flat lighting (bright, even, no night mood): a wipe or dissolve over the first ~2 s. Then fade in the grime and decals (stains, wear, posters, signage) over the next ~2 s. |
d_lighting |
240 (10 s) | Start from the textured scene in darkness (world near black, every light off). Turn the lights on one after another over ~7 s, with a short ramp each: key practicals first (lamp/pendants), then signage/neon, window/street, fills and bounce last. Then the volumetrics (steam, haze, rain) fade in. End on the full night look. |
Render engines: A–C in Eevee (fast). D in Cycles with low samples and the denoiser; it is the money shot, so keep it clean, but under ~10 GPU-minutes.
GPU rules (one 16 GB card shared by both scenes)
- Every Blender run, previews included, goes through
pipeline/gpu_run.sh. Never two Blenders on the GPU. - Don't hold the GPU between clips; release it so the other scene can interleave.
- Don't save over any existing .blend; build the BTS setup in a script on a copy (
breakdown/<scene>/bts.blend).
When done
Check a contact sheet of each clip yourself, then report: clip paths, frame counts, GPU minutes, and anything that looks off. Don't commit or publish.