Single-image inference

Not every question needs a pipeline. Any model wrapper can be called directly on one image, which is handy for prototyping prompts before running them over an inventory.

Source: examples/single_image_vlm_inference.py.

from rapidtools import download_dataset
from rapidtools.models import GenerationConfig, load

[image] = download_dataset('synthetic_landslide_image')

model = load('gemma4', model_id='google/gemma-4-E2B-it', temperature=0.4)
output = model.run_inference(
    image_inputs=image,
    prompt='Describe the condition of this post-disaster scene and what went wrong.',
)
print(output.text)

Ask for structured output with json_mode when you plan to turn the prompt into a pipeline attribute:

output = model.run_inference(
    image, 'Return JSON with keys hazard, severity (0-4) and evidence.',
    generation=GenerationConfig(json_mode=True, temperature=0.0),
)

examples/label_mapping.py shows another direct use of a model: asking it to translate plain-English object names into Mapillary’s label vocabulary through MapillaryLabelMapper.