I tested Astra 6 with a furniture reference and it did a really good job helping recreate it in SketchUp. I only gave it a link to the furniture set, and it handled the rest of the research itself.
What I find interesting is that you could now go shopping, take pictures of the exact furniture you want, and then build a version of it for your model instead of searching for something similar online.
Visually it looks quite neat, but 11 MB and 41k polygons for a single wooden stool is where practical modeling runs into trouble.
When you put four or five such assets into an actual interior scene, you easily add 50 to 60 MB of dense, unoptimized mesh. That immediately drags down viewport navigation, slows down auto-saves, and creates heavy sync issues on client files. A simple piece like this usually just needs clean low-poly geometry and instanced components for repetitive parts like the legs.
Generating the visual form from a reference is definitely an interesting step forward, but until these tools can handle proper mesh retopology and keep weights light, using them in real daily production setups will remain tricky.
It really depends on the workflow. These would mainly be used in D5 Render, on my end. The furniture can be created in SketchUp, but there’s no need to load all that detailed geometry into the architectural model itself. Keep the SketchUp model light and use the detailed furniture later in D5. Complete D5 projects can easily go over 1 GB anyway, so larger assets are pretty normal in that kind of workflow.
It is good to see people experimenting with these new workflows, and the visual recreation itself is quite neat.
External renderers can definitely take the load, but the practical catch is when 11 MB assets sit on 3D Warehouse as standard SKPs-they inevitably get pulled straight into live projects. A large chunk of the crash threads on this forum trace back to unoptimized Warehouse drops like this.
Until these AI pipelines handle automated low-poly retopology alongside the high-res mesh, it just passes the bloat downstream to everyday users.
One thing to do would be to set goal parameters if you don’t want a bloated sketchup model; tell it things like “stick to quads” or “no ngons” and give it an average target of edges/faces and texture size/format per component to restrain wonton proliferation of excessive edges and faces or bloated texture files.
I actually went in the opposite direction with the stool. I specifically asked for ultra realism, very refined details, visible imperfections in the wood, and a wabi-sabi look. I had to stop it short because it could have kept adding detail and gotten really bloated.
