Astra tops Blender 3D modeling benchmark on first day of testing
A user with Astra access ran their in-progress Blender 3D modeling benchmark, where it modeled a drill from a single reference image angle over multiple turns, and calls it the best result seen from any model so far.
.... holy crap... i am amazed... got gpt-6 astra access and finished up meetings and stuff i had and was finally able to dive into it, first thing i did was actually run them through the blender 3d modeling benchmark i'm making on the first task test and UHHHHHHH... yeah i was right earlier that 3d modeling is only a few generations from being completely solved by LLMs ... this is a multi-turn oneshot, modeled from only seeing ONE REFERENCE IMAGE ANGLE of the target drill... this is run on a no image gen test and i'm still refining the prompt guidance for the benchmark, but under this test as it is, this is the best result i've seen from any model (vs fable 5.1 and 5.6 sol)... i'll be continuing to develop this benchmark for blender 3d modeling and publishing the results at a later date, but wow at this result from unrefined eval prompting on day one of me working on the bench