Model Details
Table with columns: Item, Description| Item | Description |
|---|
| Model | SEELE-AI/EVA01-2B-Instruct-LoRA |
| Release type | UND-side LoRA adapter |
| Base model | Qwen/Qwen3-VL-2B-Instruct |
| 3D input | .glb mesh input through the EVA01 mesh UND processor |
| Components | PEFT adapter, EVA01 mesh UND encoder, connector, tokenizer, processor, and config files |
| Training recipe | Alignment followed by instruction tuning |
| Code | SeeleAI/OpenEVA |
| Project page | EVA01 |
| Paper | arXiv:2605.16745 |
This repo does not include the full Qwen3-VL base weights. The EVA01 loader resolves the base model from the LoRA config by default. For offline or local deployment, pass a local base path through base_model_name_or_path.
Installation
EVA01 requires the OpenEVA code package in addition to the checkpoint files.
git clone https://github.com/SeeleAI/OpenEVA.git
cd OpenEVA/EVA01
bash install.sh
source .venv/bin/activate
If a different CUDA wheel index is required, set TORCH_INDEX_URL before running the install script.
TORCH_INDEX_URL=https://download.pytorch.org/whl/cu121 bash install.sh
Quick Inference
CLI:
python infer.py \
--checkpoint SEELE-AI/EVA01-2B-Instruct-LoRA \
--mesh assets/examples/construction_backhoe.glb \
--question "Describe this 3D object in detail."
Python API:
import torch
from eva01 import EVA01ForConditionalGeneration, EVA01Processor
model = EVA01ForConditionalGeneration.from_pretrained(
"SEELE-AI/EVA01-2B-Instruct-LoRA",
torch_dtype=torch.bfloat16,
device_map="auto",
)
processor = EVA01Processor.from_pretrained("SEELE-AI/EVA01-2B-Instruct-LoRA")
messages = [{
"role": "user",
"content": [
{"type": "mesh", "mesh": "assets/examples/construction_backhoe.glb"},
{"type": "text", "text": "Describe this 3D object in detail."},
],
}]
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt",
).to(model.device)
output_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)
text = processor.batch_decode(
output_ids[:, inputs.input_ids.shape[1]:],
skip_special_tokens=True,
clean_up_tokenization_spaces=False,
)[0]
print(text)
Local base override:
model = EVA01ForConditionalGeneration.from_pretrained(
"SEELE-AI/EVA01-2B-Instruct-LoRA",
base_model_name_or_path="/path/to/Qwen3-VL-2B-Instruct",
torch_dtype=torch.bfloat16,
device_map="auto",
)
The public mesh token is <|mesh_und_pad|>. The processor returns input_ids, attention_mask, and mesh_und_values.
Gradio Chat
python app.py --checkpoint SEELE-AI/EVA01-2B-Instruct-LoRA --host 127.0.0.1 --port 7860
The OpenEVA app supports uploaded .glb files and includes 10 built-in TexVerse examples. The OpenEVA GitHub repository also contains a PBR-rendered example gallery.
PointLLM-200 Evaluation
The OpenEVA eval script downloads PointLLM-200 benchmark files staged in the Full checkpoint repo and writes results under EVA01/outputs/pointllm200/.
python eval_pointllm200.py --variant lora
The deterministic path computes BLEU, ROUGE, METEOR, Sentence-BERT, and SimCSE with fixed seed 20260615 and greedy generation. GPT-ref and GPT-img judge paths are available when OPENAI_API_KEY is set.
The PointLLM-200 benchmark source is cited as RunsenXu/PointLLM.
Metrics
All metrics below are reported on PointLLM-200 with 200 samples, fixed seed 20260615, and greedy decoding. GPT-ref and GPT-img are judge metrics and may vary with judge model and API settings.
Table with columns: Model, B-1, B-4, R-L, METEOR, SBERT, SimCSE, GPT-ref, GPT-img| Model | B-1 | B-4 | R-L | METEOR | SBERT | SimCSE | GPT-ref | GPT-img |
|---|
| PointLLM-13B | 7.873 | 0.649 | 10.519 | 13.620 | 47.539 | 48.602 | 51.735 | 49.745 |
| ShapeLLM-13B | 10.542 | 1.050 |
Intended Use
This checkpoint is intended for research and development around 3D mesh understanding, 3D asset captioning, and mesh-grounded question answering. It expects mesh input through the EVA01 processor and should be used with the OpenEVA runtime/API.
Limitations
- The public release focuses on the UND-side path.
- This LoRA repo does not include the full Qwen3-VL base weights.
- Model quality depends on mesh geometry, scale, topology, materials, and texture availability.
- Outputs may contain incorrect or unsupported details when the mesh is ambiguous, incomplete, or visually underspecified.
- GPT-ref and GPT-img scores are judge references and are not bitwise reproducible across judge model or API changes.
Citation
@misc{eva01_2026,
title = {EVA01: Unified Native 3D Understanding and Generation via Mixture-of-Transformers},
author = {Zongyuan Yang and Mingjing Yi and Wanli Ma and Chenzhuo Fan and Bocheng Li and Baolin Liu and Yuke Lou and Yingde Song and Yongping Xiong and Zhengdong Guo and Shimu Wang},
year = {2026},
eprint = {2605.16745},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}