About lora

[fenght96]

Any plan for lora?

[likemango]

As far as I know, it seems non sense for lora in 3D. Lora use for the 2D style. What is style about 3D?

[MangoLion]

lora is a concept, people apply lora to applicable model architectures. there's loras for llms and other stuff too

[likemango]

lora is a concept, people apply lora to applicable model architectures. there's loras for llms and other stuff too

Lora for customize output style. When it comes to 3D, how to define 3D style? As 3D style defined by 2D image.

[FishWoWater]

3D Models also have styles (e.g. lowpoly, stone style, pocket monster). The idea is to train a LoRA for the DiT model , as the practive of Section.5 of CLAY. Not sure whether TRELLIS will support this.

[likemango]

3D Models also have styles (e.g. lowpoly, stone style, pocket monster). The idea is to train a LoRA for the DiT model , as the practive of Section.5 of CLAY. Not sure whether TRELLIS will support this.

Thanks for your sharing, that inspire me. But CLAY demo has been depreacted and its project move to Rodin(commercial project)

[FishWoWater]

@YuDeng Hi, do you have any plan of releasing/training a lora for TRELLIS?

If not, I would like to have a try. I want to start from some lowpoly style, or optimizing for 3D cartoon human models.

The motivation is that many 3DAIGC works have bad performance on human/character models (faces or bodys).

I can filter out and clean the charcacter models in Objaverse-XL(perhaps along with other datasets), and use them to train a small lora for the two DiT models in TRELLIS.

Do you have any advice on this (is it feasible?) Thanks :)

[YuDeng]

Hi @FishWoWater, we do not have any specific plan for the lora part.

If you want to train a lora model for character, I think it's better to collect more data beyond the Objaverse-XL. The character models in Objaverse-XL should already be included in our training set.

[FishWoWater]

OK, thanks

Hi @FishWoWater, we do not have any specific plan for the lora part.

If you want to train a lora model for character, I think it's better to collect more data beyond the Objaverse-XL. The character models in Objaverse-XL should already be included in our training set.