Three weeks have passed since the TRELLIS.2 post, and the gap between a working generator and controllable infrastructure has shrunk almost to zero. The last two days completed three large pieces of the puzzle: a render studio for every signed-in user, a panel controlling another machine on the network and a small but important workload regulator.
Blender Renderer: a shared engine room
The Mac mini gained a new role: a .blend render server. I return to the ecosystem page with a scene file of up to 8 GiB, and an optimization script spreads the Cycles render across all CPU threads and starts the queue. The progress bar reports the real state. Afterwards you can download the original, an optimized project copy or the result as PNG or MP4. A fast EEVEE mode remains available as a fallback.
Everything goes onto persistent external storage and is visible to every signed-in user. This is deliberate: the render library is a shared resource rather than one user's private space.
.blendscene
optimizationCycles
10 threadsprogress
barPNG / MP4
shared library
A sandbox that really isolates
An independent audit highlighted something worth explaining openly: Blender's --disable-autoexec blocks embedded scripts, but does not stop the compositor writing files anywhere it likes. I moved outputs into a separate job directory, and renders now run in a macOS sandbox without network access or private-directory access, with a free-disk-space limit. External code review confirmed the paths, queue and upload were contained: 26 tests, no regressions.
A bridge to the RTX 5090
August's second pillar is the AI Manager panel. The virtual machine gets its own endpoint in the protected /ekosystem/ai-manager/ area. Browsers reach it through a secure bridge on the Mac mini that knows the tokens and allowed addresses. The configuration is deliberately asymmetric: clients see only what the bridge considers safe.
The panel shows the current mode and model, backend and GPU service states, nine live metric cards covering prefill, generation, GPU, VRAM, temperature, power and queue, plus nine historical charts spanning 15 minutes to 30 days. I drew the charts from scratch on Canvas 2D because the headless browser did not cooperate with the chart library and dependency-free implementation was in the specification.
During testing, the charts briefly showed blank canvases. A race was responsible: the live stream grew the time buffer, making the condition for unloaded history false, so history never entered the buffer. One flag and full buffer initialization fixed it, after which all series filled 80–93% of their length.
Concurrency you can set directly
Finally, per-profile request concurrency became configurable with a plain numeric field: enter 1, 2, 8 or 16 and apply. The backend confirms that a restart is needed, and the panel shows operation progress until it is ready. In the live test, the restart took about 66 seconds, the runtime recovered and the new value persisted.
Five quiet guardians
An audit of scheduled jobs revealed what keeps the site upright: five watchdogs, every minute, without involving models. Studio, crawler, audio studio, downloader and the new Blender Renderer. The logic is simple: silence means healthy. A script logs only when it has repaired something. Otherwise the scheduler records empty stdout and stays quiet.
What remains
August closed in three layers: content, with a shared render library and 3D generator; control, with a dashboard, telemetry and concurrency; and silence, with five quiet guardians and a clean repository pushed to GitHub. Everything passed tests, was checked in the browser and reached production without downtime.