laconiq-ai/technique-router

Re-hosted by the Laconiq project. Original model: chopratejas/technique-router by chopratejas, licensed under Apache-2.0. Hosted under laconiq-ai for distribution reliability and provenance control. All credit for the underlying model belongs to the original author; the Apache-2.0 license and attribution are retained.

Technique Router (MiniLM)

A fine-tuned MiniLM classifier that routes a natural-language query about an image to one of four image-optimization techniques. Originally built for the Headroom SDK by chopratejas. LaconIQ re-hosts it so deployments inside a controlled or air-gapped environment have a stable artifact source.

What it does

Given a question about an image, the model selects the technique best suited to answering it:

Technique	Best for
transcode	Text extraction and OCR tasks
crop	Region-specific queries
full_low	General understanding
preserve	Fine detail and counting

The routing decision is what matters: a question that only needs text extracted does not require the same image fidelity as a question that requires counting objects.

Model

• Base: microsoft/MiniLM-L12-H384-uncased
• Parameters: 33.4M
• Task: four-class text classification over image queries

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "laconiq-ai/technique-router"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

query = "What brand is the TV?"
inputs = tokenizer(query, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    pred_id = torch.argmax(probs, dim=-1).item()

technique = model.config.id2label[pred_id]
print(f"{query} -> {technique}")

With the LaconIQ Gateway

LaconIQ uses this model for image-technique selection in the Gateway's image handling. Gateway packages are provided with a pilot engagement rather than from a public index. See the LaconIQ documentation for how the Gateway uses this model.

For use with the original Headroom SDK, see the upstream model.

Intended use

Routing image-analysis queries to an appropriate optimization technique, so a vision model receives the detail a question actually needs.

Limitations

• English only.
• Trained on common image-understanding phrasings; it may not generalize to domain-specific terminology.
• The router selects a technique, it does not validate the outcome. A misrouted query can lose detail the question needed, so techniques that discard information should be reviewed against your own content before wide use.

Citation

@misc{headroom-technique-router,
  title={Technique Router for Image Token Optimization},
  author={Headroom AI},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/chopratejas/technique-router}
}

License

Apache-2.0, inherited from the original model. See the attribution note above.