---
license: gemma
base_model: google/gemma-3-4b-it
tags:
- vision-language-model
- TEM
- microscopy
- materials-science
- gemma
- scientific-VLM
language:
- en
pipeline_tag: image-text-to-text
---

# ATOMIC-Gemma

ATOMIC-Gemma is a domain-specific Vision-Language Model for Transmission Electron Microscopy (TEM), fine-tuned from Gemma3-4B-IT using Stage 2 instruction tuning on TEM conversation data.

> **Note:** ATOMIC-Gemma is developed after the ECCV 2026 submission deadline and is **not part of the published paper**. It is released here to demonstrate the generalizability of the ATOMIC training pipeline across different base model architectures.

For the published paper and full pipeline, please refer to our GitHub repository:
👉 [https://github.com/SemiMRTLab-NCKU/ATOMIC](https://github.com/SemiMRTLab-NCKU/ATOMIC)

---

## Model Details

| | |
|---|---|
| **Base Model** | Gemma3-4B-IT (`google/gemma-3-4b-it`) |
| **Training Stage** | Stage 2 (instruction tuning) only |
| **Training Data** | 60K Stage 2 conversations |
| **Domain** | Transmission Electron Microscopy (TEM) |
| **Modalities** | CTEM, HR-TEM, STEM, Diffraction |

---

## Inference

ATOMIC-Gemma can be loaded directly via `transformers`:

```python
from transformers import AutoProcessor, Gemma3ForConditionalGeneration
from PIL import Image
import torch

model_id = "LabSmart/ATOMIC-Gemma"

model = Gemma3ForConditionalGeneration.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16
).eval()
processor = AutoProcessor.from_pretrained(model_id)

image = Image.open("your_TEM_image.png").convert("RGB")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "What type of TEM image is this?"}
        ]
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device, dtype=torch.bfloat16)

input_len = inputs["input_ids"].shape[-1]

with torch.inference_mode():
    generation = model.generate(**inputs, max_new_tokens=256, do_sample=False)

generation = generation[0][input_len:]
response = processor.decode(generation, skip_special_tokens=True)
print(response)
```

---

## Training Data

Training data is available on HuggingFace:
👉 [https://huggingface.co/datasets/LabSmart/ATOMIC_dataset](https://huggingface.co/datasets/LabSmart/ATOMIC_dataset)

---

## Citation

```bibtex
@inproceedings{atomic2026eccv,
  title     = {ATOMIC: A Domain-Specific Vision-Language Model
               for Transmission Electron Microscopy},
  author    = {Tu, C. and Hsu, Shu-han and others},
  booktitle = {Proceedings of ECCV 2026},
  year      = {2026},
  note      = {BibTeX will be updated upon publication}
}
```

---

## License

This model is released under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). It is intended for academic research purposes only.