This Astra feels truly amazing and could really shake up modeling! A while back we “modeled to render”, then moved to “render to model”… and now we’ve got image to model!
It’s still not working for AC29—only AC28, which I tested. It’s great to have it online, too.
One thing I’m wondering about is that image-to-model uses quite a few credits; I tried it yesterday and they disappeared fast. My main concerns are accuracy, consistency, and parametrics. The GLB file isn’t importable yet—will there be other formats?
Yes, we’re looking at how we can bring this to ArchiCAD, and can use some help with testing the early builds. We already have the 3d extraction in a beta build: Archicad 29 support - #8 by Ole
The version seems to have some issues, and we need to build to package it for macs. More to come on this.
Regarding tokens, yes, it’s an intensive process with any provider. However, these models get distilled, and are usually made available at fraction of the cost.
Regarding accuracy: these are the first models that are somewhat correct at 3d modeling with a certain LOD. The LLMs will only get better and more accurate over the coming months.
I agree that tools like Astra or other AI models could handle a lot of our repetitive tasks, but there’s definitely more to explore here. Building models for construction, points us toward BIM, and with Veras diving into this, the level of development but more importantly, the level of information need will be key. What we really want isn’t just renders or visuals—it’s structured information we can use for many outputs. Since BIM depends on the data inside model elements, could Veras eventually do it all?
At first, we might use “text to image” with prompts and references to brainstorm and create visuals. Once we have a few options, our creative team and clients can choose the direction. Then we could use Veras again for “image to model” through Astra or something similar. The big question is: who will pull out the information we need? Will it be Veras, a new BIM tool, or something else? If not, the models will need to work smoothly with Revit and ArchiCAD—otherwise, we’ll just get background data and have to rebuild it in our BIM tools.
Good points! When you look at hydrating generic data with items that are products, we already know that LLMs coupled with web-search tools are really good at that.
For example, it you have a building product that came in damaged in shipment, you can ask an LLM to find the closest store that you can purchase this product from.
Another example: Let’s say you have a generic piece of furniture, and need to procure that. An LLM can find a close product and product availability and give you the cutsheet.
It’s a matter of where you want to store that information. With Veras, that can be in a BIM app of choice (that we support), or directly in Veras. The cool part is that you can store it in a PDF or image, as LLMs can restructure and recompose it together in a unified format.
