apekshik/anyscene

What it does

AnyScene turns a short written brief into a finished photoreal Blender scene. An agent follows the repository's SKILL.md process: it generates concept and prop reference images, converts each prop to 3D with both Tripo H3.1 and Meshy 7.1 on fal and keeps the better model, scripts the room and dresses it from large furniture down to clutter, then checks props for collisions and floating. It adds grime, decals, night lighting, haze, steam and rain, renders cinematic and handheld walk-through shots in Cycles, grades them, and cuts a behind-the-scenes video of the build.

No prop is modelled by hand. The two included examples, a 1 AM Tokyo electronics repair counter and a Tokyo ramen counter at last orders, contain 86 and 74 generated props respectively. The repository ships the process and code; the large finished .blend files, generated models and renders are downloaded separately.

The workflow suits 3D artists and technical directors who want AI-assisted environment building while keeping scripted layout, lighting and camera control, and who are comfortable supervising an agent through the build.

Quick start

Prepare the following before starting:

• Blender 5.0+ available as BLENDER or on PATH.
• ffmpeg with h264_nvenc (or substitute libx264 in encode steps on non-NVIDIA GPUs).
• Node 20+.
• A fal API key exported as FAL_KEY.
• uv if you want the breakdown videos.

Start a new scene by pointing your agent at SKILL.md with a brief, for example: "Use SKILL.md to build a new scene: a Lisbon tram depot at dawn, one tram half-stripped for repair." Review the agent's output at each step's gate: concept, prop picks, clay layout, lit stills and shot previews. A new scene starts by copying the closer example scene and rewriting its room shell and dressing scripts, since those are specific to each place.

Usage constraints

The reference scenes were built on Windows with an NVIDIA GPU; ffmpeg encoding uses h264_nvenc by default and needs changing on other GPUs. Generation requires a paid fal account and API key. The two example scenes each took about a day of agent work and roughly $150 of generation, so cost and time scale with scene size and prop count. Scene generation depends on external Tripo and Meshy services through fal, so availability is tied to those APIs. The code is MIT-licensed, but generated assets are released separately under their own license.

HotRepo's take

On the Python new-project list for 2026-11-01, AnyScene recorded 117 stars and 15 forks, with 17 stars gained during the observed window. The star curve over the same morning rose from 100 to 117 across four hourly samples. No all-language daily rank is recorded for the project.

The repo's practical value is the process it documents: a gated, scripted pipeline that keeps the artist in control of layout, lighting and camera while outsourcing prop creation to two competing image-to-3D services. That makes it more of a supervised production workflow than a one-click generator, and the ~$150-per-scene figure is a useful cost anchor for anyone evaluating it.