# The Room Remembers: production recipe

One room, six eras: 1550, 1750, 1925, 1975, 2026 and 2050. This recipe documents the selected production used by the interactive room and its six films. Recorded prompts are reproduced verbatim in `records/`. They are historical requests, not suggested replacements.

## 1. Build a common architectural foundation

In Blender, work in metres. Build an 11 x 14 metre room with a 6.6 metre ceiling and 0.55 metre walls. The gallery is 8.2 metres wide and 2 metres deep, with its deck at 3.4 metres. Keep the windows, stair, gallery, fireplace and doors in consistent locations across all six era collections. Separate architecture, furnishings, lighting and exterior so each can be reviewed independently.

Use `data/room-dimensions.json` for the recorded controls. These are foundation parameters, not an instruction to recreate every archived feature. In the final room, the door beside the fireplace is closed, the arched opening above the door is removed, and the 1750 curtains are absent. Preserve a clear walking route and keep wall-mounted decoration inside wall panels, away from columns, sconces and openings.

The architecture is authored Blender geometry. There is no single text prompt that generates this whole room. The ZIP contains instructions, parameters and placement records, not the original .blend file or the complete application source.

## 2. Generate one coherent collection direction per era

Open [GPT Image 2.5 Sunburst](https://app.scenario.com/models/model_openai-gpt-image-2-5-sunburst?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe). Start with that era's `YEAR-style.md` recipe. It describes a collection of matching furnishings, the material palette, carving language, textile colors and exclusions. Keep the selected collection image as the style anchor.

The six directions are walnut and oxblood Renaissance furniture; sage silk, ivory and aged gold Rococo; teal, walnut, brass and stepped Art Deco forms; teak, corduroy and warm 1970s colors; restrained contemporary linen and timber; and pale futuristic composite forms. The exact recorded text is in the individual recipes.

Generate each object separately using its recorded Sunburst prompt and the collection image in `referenceImages`. Request a complete isolated object with clear silhouette and a plain background. Do not generate an entire room and cut furniture out of it. Inspect the image for complete legs, supports, handles and believable construction before continuing.

## 3. Convert the selected images with Tripo P2

Open [Tripo P2](https://app.scenario.com/models/model_tripo-p2-image-to-3d?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe). Supply the individual object image as `image`. The recorded settings include textures and PBR enabled, `autoSize: false`, `quad: false`, and `orientation: "align_image"`. Read each recipe's actual `faceLimit` instead of assuming one value for every object.

This step uses image conditioning and parameters; it has no separate furniture text prompt. Preserve the image job, model job and output asset IDs. The package includes all 140 generated source objects. Of these, 139 distinct models are active across 324 placements. The unused source object is explicitly marked as not placed.

## 4. Assemble and inspect the room

Import the GLBs through the Scenario Blender plugin. Normalize each model with uniform metric scaling, a centered horizontal bounding box and the lowest point at the origin. Inspect its front direction, contact points and underside. Apply placement records from the object recipes; their positions are glTF Y-up, while the camera samples below use Blender Z-up. For Blender conversion, map glTF `[x, y, z]` to `[x, -z, y]`.

Review from the actual visitor viewpoints as well as from above. Check paintings and mirrors against wall panels, columns and sconces; test tabletop objects against their support surface; clear the fireplace; and leave camera routes open. Multiple placements can share one mesh. The per-era unique-model count is different from the placement count.

Export the architecture and furnishing collections to GLB, preserving mesh/texture relationships and the recorded transformations. In Three.js, use those same placements, era materials and camera landmarks. The room's object viewer presents each model, source reference and provenance, with material, solid and wireframe modes.

## 5. Generate surfaces and the grounded panorama

Use [Scenario Texture](https://app.scenario.com/models/model_scenario-texture?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe) for the eight material swatches and the revised 1975 wallpaper. These requests produce color textures. UV scale, roughness and physical material setup are separate authored decisions. The accepted wallpaper recipe records its exact crop, resampling and repeat scale.

Use [Scenario Skybox GPT](https://app.scenario.com/models/model_scenario-skybox-gpt?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe) for each panorama. The base prompt requests an eye-level view from flat short grass, with the lake and hills far away. The later eras reference the same location. Generate at 3840 x 1920, then use [Scenario Skybox Upscale](https://app.scenario.com/models/model_sc-upscale-flux-skybox?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe) at 2x for 7680 x 3840 delivery. This is an upscaled panorama.

Flatten the lower hemisphere beneath the room using `data/ground-projection.json`. The room meets this projected surface with contact shadows. There is no separate hill under the building.

## 6. Create six instrumental soundtracks

Open [ElevenLabs Music v2.5](https://app.scenario.com/models/model_elevenlabs-music-v2-5?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe). Each era recipe includes its exact prompt, seed, 60-second duration, instrumental setting and output format. The choices range from lute and harpsichord to jazz, electric piano and future chamber textures.

The selected tracks were processed with FFmpeg loudness normalization at -20 LUFS, -2 dBTP and LRA 9, with 0.08-second entry and 0.3-second exit fades, then exported as 48 kHz stereo MP3 at 192 kbps. The interactive room plays this music separately. The delivered films and montage are silent.

## 7. Render a shared moving camera reference

Use `data/camera-path.json`: 360 sampled frames at 24 fps, a 15-second path, a 22 mm lens and 1.65 metre eye height. Camera position and direction are Blender Z-up coordinates. The camera travels about 13.67 metres with a roughly 144-degree turn. It advances in the first six seconds, turns toward the opposite side around seconds six to nine, then moves back toward the gallery and closed-door side. Use the recorded frame samples to reproduce the path precisely.

Render a native motion preview for every era. Keep the same camera path, timing and lens while changing the era collection. Review the complete path for furniture intersections and obstructed frames. These previews were 720p motion references; they are not the final photographic films.

## 8. Create matching photographic references and films

Use Sunburst on an opening room capture, then on a reverse capture. The reverse request gives the Blender screenshot authority over the viewpoint and the opening photograph authority over materials and light. Keep every visible object, its position, the windows, gallery and stairs. The selected 1975 opening image uses take v003; the selected 2026 reverse uses v002. The per-era film recipe contains those exact requests.

Open [Seedance 2.5](https://app.scenario.com/models/model_bytedance-seedance-2-5?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe). Use the native preview as `referenceVideos[0]`, the opening photograph as `referenceImages[0]`, and the reverse photograph as `referenceImages[1]`. The prompt gives `@video1` exclusive control of the camera and gives `@image1` / `@image2` control of photographic appearance. It explicitly requires continued movement, the broad turn and the return, with no new objects or characters.

The recorded request uses `duration: -1`, `resolution: "1080p"`, `aspectRatio: "16:9"`, `generateAudio: false`, and MP4 output. These are the settings submitted for this production; consult the current model schema before a new run. Some provider outputs required a local duration conform to exactly 15 seconds / 360 frames. Inspect a dense sequence through the turn and play the complete delivered file. Generated details can differ from the native models even when the main layout is preserved.

## 9. Assemble the time-travel montage

Open [Scenario Video Studio](https://app.scenario.com/models/model_scenario-compose-video?referral=experiments-the-room-remembers&utm_source=scenario.com&utm_medium=experiment&utm_campaign=the-room-remembers&utm_content=making-of-recipe). Follow `data/montage-edit.json` and the exact layered request in `records/time-travel-montage.md`. Take 3.5 seconds from each of the six films, in chronological order, advancing the source time for each era. Overlap adjacent clips by 0.6 seconds and fade the upper layer in. Add the recorded year-label overlays. Six excerpts minus five overlaps produce 18 seconds at 1920 x 1080, 24 fps, with no audio.

## Reusing the recipes

Open the model links in your own Scenario workspace. Replace recorded input asset IDs with your own uploaded or newly generated assets, retaining their ordering and purpose. Job IDs document provenance; they are not generation parameters. For furniture, generate the style anchor before the object image, then the mesh. For film, make the native preview and both photographic references before the Seedance request.

The pack provides recorded prompts, model IDs, parameters, selected output IDs, dimensions, placements, camera samples and montage settings. It does not bundle GLB, image, audio, movie or Blender binaries. Published room asset paths are recorded for reference. Rerunning the same recipe does not guarantee byte-identical generative output. These historical rooms and the future setting are creative interpretations.