Hacker News Reader

Comment by xnx | Hacker News Reader

Anyone interested in this type of thing should check out the free and open source TRELLIS: https://huggingface.co/spaces/JeffreyXiang/TRELLIS

Upload an image and it outputs a 3D model.

Use a separate image generator to make a model of anything you can imagine.

View on HN

Trellis is very impressive (topology notwithstanding). I put some examples together of what Trellis can do, it's definitely better with angular type models (vehicles, etc.), but unsurprisingly can struggle a bit with more organic forms.

https://imgur.com/a/trellis-2d-3d-samples-eSWJhMj

I fed TRELLIS an image of the left side of my Vespa, and it correctly (mostly) added exhaust on the other side.

Super, super cool to see -- really hyped for what this means for 3D representations!

Ya trellis i think is the current best, i'd imagine what the OPS site is doing is a txt2img+rembg into trellis or another model, and then converted to a blender format... its a nice workflow, though he says he downgraded to a less intense model to save compute right now... which is pretty obvious, it should mention somewhere which model it is currently using

I took a masked image of a tank from a game, in a perspective format, presumably the best case for these models.

Trellis did far worse than BlenderGPT. Particularly Trellis tends to have little to no detail with nearly black texturing in the parts that are "hardest" to imagine.

Somewhat interesting as this seems to use Trellis under the hood, but again, this did a substantially better job.

Reply on news.ycombinator.com