LlamaFT (Stage 1): A fine-tuned Llama-3.2-11B-Vision model that predicts discrete CAD command structures from images
TrAssembler (Stage 2): A transformer-based network that predicts continuous CAD parameters using Gaussian Mixture Flow

Available Checkpoints

LlamaFT Adapters (LoRA)

Category	Path
Chair	`llamaft/chair/`
Table	`llamaft/table/`
Storage Furniture	`llamaft/storagefurniture/`

TrAssembler Checkpoints

Category	Path
Chair	`trassembler/chair/`
Table	`trassembler/table/`
Storage Furniture	`trassembler/storagefurniture/`

Part Segmentation (PointNet2) Checkpoints

Category	Path
Chair	`partseg/chair/`
Table	`partseg/table/`
Storage Furniture	`partseg/storagefurniture/`

Usage

Stage 1: LlamaFT (Discrete Structure Prediction)

python

from transformers import AutoProcessor, AutoModelForVision2Seq, BitsAndBytesConfig
from huggingface_hub import hf_hub_download
from PIL import Image
import torch

# Download adapter
adapter_path = hf_hub_download(
    repo_id="qq456cvb/img2cad",
    filename="llamaft/chair/adapter_model.safetensors"
)
adapter_dir = str(Path(adapter_path).parent)

# Load base model
model_id = "meta-llama/Llama-3.2-11B-Vision-Instruct"
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True, 
    bnb_4bit_use_double_quant=True, 
    bnb_4bit_quant_type="nf4", 
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForVision2Seq.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    quantization_config=bnb_config
)
processor = AutoProcessor.from_pretrained(model_id)

# Load adapter
model.load_adapter(adapter_dir)
model.eval()

# Inference
image = Image.open("your_image.png").convert("RGB")
# ... see full inference code in the repository

Stage 2: TrAssembler (Continuous Parameter Prediction)

python

from omegaconf import OmegaConf
from huggingface_hub import hf_hub_download
import torch
import numpy as np

# Download checkpoint and config
ckpt_path = hf_hub_download(
    repo_id="qq456cvb/img2cad",
    filename="trassembler/chair/last.ckpt"
)
config_path = hf_hub_download(
    repo_id="qq456cvb/img2cad",
    filename="trassembler/chair/config.yaml"
)

# Load model (requires Img2CAD repository)
from TrAssembler.model import GMFlowModel

config = OmegaConf.load(config_path)
model = GMFlowModel.load_from_checkpoint(
    ckpt_path,
    args=config,
    embed_dim=config.network.embed_dim,
    num_heads=config.network.num_heads,
    dropout=config.network.dropout,
    bias=True,
    scaling_factor=1.,
    args_range=np.array([-1., 1.])
).cuda().eval()

# Generate CAD parameters
# pred_args = model.sample(batch)

Paper

Img2CAD: Reverse Engineering 3D CAD Models from Images through VLM-Assisted Conditional Factorization

Architecture

LlamaFT

Base model: Llama-3.2-11B-Vision-Instruct
Fine-tuning: QLoRA (rank=16, alpha=16)
Target modules: q_proj, v_proj

TrAssembler

Image encoder: DINOv2-B (frozen)
Sketch transformer: 6-layer decoder
Part transformer: 6-layer decoder
Output: Gaussian Mixture Flow for continuous parameters

Citation

bibtex

@inproceedings{you2025img2cad,
  title={Img2cad: Reverse engineering 3d cad models from images through vlm-assisted conditional factorization},
  author={You, Yang and Uy, Mikaela Angelina and Han, Jiaqi and Thomas, Rahul and Zhang, Haotong and Du, Yi and Chen, Hansheng and Engelmann, Francis and You, Suya and Guibas, Leonidas},
  booktitle={Proceedings of the SIGGRAPH Asia 2025 Conference Papers},
  pages={1--12},
  year={2025}
}

License

This project is released under the MIT License.