LFM2.5-VL-450M-Extract

<center> <div style="text-align: center;"> <img src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/2b08LKpev0DNEk6DlnWkY.png" alt="Liquid AI" style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;" /> </div> <div style="display: flex; justify-content: center; gap: 0.5em;"> <a href="https://playground.liquid.ai/chat?model=lfm2.5-vl-450m"><strong>Try LFM</strong></a> • <a href="https://docs.liquid.ai/lfm/getting-started/welcome"><strong>Docs</strong></a> • <a href="https://leap.liquid.ai/"><strong>LEAP</strong></a> • <a href="https://discord.com/invite/liquid-ai"><strong>Discord</strong></a> </div> </center> <br>

LFM2.5-VL-450M-Extract

LFM2.5-VL-450M-Extract extracts user-defined fields from images and returns them as JSON. It is Liquid AI's first vision model in the Liquid Nanos collection—compact, task-specific models built for production workflows—and extends the Extract family alongside LFM2-350M-Extract for text documents.

⚙️ How it works

You specify what to extract as a YAML field list in the system prompt, and the model returns a JSON object with those fields. Structured outputs integrate cleanly with rule-based systems and downstream pipelines. Use it out of the box or fine-tune for domain-specific extraction.

• System prompt:

wood_color: The overall coloration of the wood surface
wood_texture: The tactile quality of the wood surface 
wood_pattern: The partern types visible on the wood surface

• User prompt:

  <img src="https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract/resolve/main/sample_image.png" width="300">

• Output:

{
  "wood_color": "light to medium brown",
  "wood_texture": "smooth with visible grain",
  "wood_pattern": "parallel, irregular, wavy"
}

Our model supports the enum feature, which lets you provide a list of possible choices alongside the field description as follows, and the model will return one of the listed values as its answer.

• System prompt:

wood_color: The overall coloration of the wood surface, such as blue, red, or light tan
wood_texture: The tactile quality of the wood surface, select from smooth, rough, or grainy
wood_pattern: The partern types visible on the wood surface, e.g., straight, wavy, or curly

🌟 Use cases

• Detecting safety-critical events in images (e.g. fallen person, fire, leakage) to trigger automated safety systems.
• Collecting statistical information about objects across video frames for analytics pipelines.
• Auto-tag product images with structured attributes for Retail/E-commerce.

📄 Model details

📊 Performance

We evaluated LFM2.5-VL-450M-Extract on a 2,000-sample benchmark of (image, schema, JSON) triples, with reference labels generated by an ensemble of frontier multimodal models. Predictions are scored on the following three dimensions:

• JSON Validity — share of samples producing strict-parseable JSON
• Schema Consistency F1 Score — set-level F1 over predicted vs requested field names, macro-averaged across samples
• VLM Judge Score — match against the image directly, judged by a separate vision model (Qwen/Qwen3.5-35B-A3B)

<img src="https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract/resolve/main/lfm2_vl_450m_metrics.png" width="800">

LFM2-VL-450M-Extract outperforms similarly-sized (sub-1B) open-source VLMs on this benchmark and is competitive with models 4× its size.

Reproducing these numbers: The full evaluation pipeline, which includes extraction, VLM judging, and metric aggregation, is bundled in this repository under model_eval/. Setup, configuration, and run instructions are in the folder's README.

Scope: These numbers characterize the model on the input/output form it is designed for: a single input image, a YAML field list as the schema, and a flat JSON object as the output. Performance is not expected to transfer to largely different tasks, e.g. multi-image reasoning or free-form VQA.

<!-- > Generic instruction-tuned VLMs (SmolVLM, moondream) cannot perform > schema-based extraction zero-shot regardless of prompt strategy. > Under the most permissive prompt setups, they either: > - produce free-form captions ignoring the JSON instruction, or > - produce valid-shaped JSON but echo the schema descriptions or > few-shot example values as field values (zero faithfulness to > the image). > > LFM2-VL-Extract's task-specific training is what enables strict-JSON > output with faithful, image-grounded values in a single zero-shot > call — no few-shot examples, no grammar constraints, no inference > wrappers. -->

The full evaluation pipeline, which includes extraction, LLM/VLM judging, and metric aggregation, is included in this repository under model_eval/. Usage details are in the folder's README.

🏃 How to run

You can run LFM2.5-VL-450M-Extract with Hugging Face transformers v5.1 or newer:

pip install transformers pillow

from transformers import AutoProcessor, AutoModelForImageTextToText
from transformers.image_utils import load_image

model_id = "LiquidAI/LFM2.5-VL-450M-Extract"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16",
    trust_remote_code=True,
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

image = load_image("https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract/resolve/main/sample_image.png")

fields_yaml = """wood_color: The overall coloration of the wood surface
wood_texture: The tactile quality of the wood surface
wood_pattern: The pattern types visible on the wood surface"""

system_prompt = f"""Extract the following from the image:

{fields_yaml}

Respond with only a JSON object. Do not include any text outside the JSON."""

conversation = [
    {"role": "system", "content": system_prompt},
    {"role": "user",   "content": [{"type": "image", "image": image}]},
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    tokenize=True,
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
response = processor.batch_decode(
    outputs[:, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)[0]
print(response)
# {
#   "wood_color": "light to medium brown",
#   "wood_texture": "smooth with visible grain",
#   "wood_pattern": "parallel, irregular, wavy"
# }

[!WARNING] The model is intended for single-turn conversations. We recommend using greedy decoding (temperature=0).

📬 Contact

• Got questions or want to connect? Join our Discord community
• If you are interested in custom solutions with edge deployment, please contact our sales team.

Citation

@article{liquidai2025lfm2,
 title={LFM2 Technical Report},
 author={Liquid AI},
 journal={arXiv preprint arXiv:2511.23404},
 year={2025}
}