Itan-levin/sapset-adapters

SAPset per-asset adapters for Edit3D-Bench

100 LoRA adapters for the Hunyuan3D 2.1 shape diffusion transformer, one for each source object of Edit3D-Bench. Each adapter personalizes the generator to a single object, so that it regenerates that object from noise; SAPset edits the object by applying the adapter at a different strength per latent token.

Base model	Hunyuan3D 2.1 shape DiT
Rank / alpha	16 / 16
Target modules	feed-forward layers, including those inside the mixture-of-experts layers (fc1, fc2, net.0.proj, net.2)
Training data per object	266 renders (16 fixed benchmark-style views and 250 random cameras under randomized lighting), paired with latents encoded from points sampled on the object's surface
Objective	flow matching, base weights frozen
Checkpoint	the step with the lowest validation Chamfer distance, per object

manifest.json lists, for every folder, the Edit3D-Bench object it belongs to, the selected training step and its validation Chamfer distance. Each folder holds a standard PEFT adapter_config.json and adapter_model.safetensors.

Usage

With the SAPset code repository, installed as its README describes:

python scripts/download_adapters.py --out adapters --all          # or --objects GSO/SORTING_BUS ...
sapset-edit --bench-root Edit3D-Bench/data --case GSO/SORTING_BUS/prompt_2 \
            --adapters adapters --out outputs/sorting_bus_p2

The 12 whole-object edit cases are in Itan-levin/sapset-global-edits.

Citation

To be added.