Tensor Preprocess Visualizer
Inspect how a computer-vision model input tensor is built from an image: resize, letterbox, channel order, normalization, dtype and 4D shape.
Upload an image and define the model input expected by your vision pipeline. Pixlane shows the exact resize path, channel order, normalization math, tensor layout and channel heatmaps so object detection, segmentation and classification inputs are easier to debug before inference.
How to Use Tensor Preprocess Visualizer
- Upload an image: Upload one of the formats listed by the tool (JPG, PNG, WebP) and wait for the local preview.
- Adjust settings and process: Adjust the available controls, then run the selected image processing workflow in your browser.
- Review and download: Review the result, then download it in one of the formats offered by this tool.
Frequently asked questions
What problem does this solve?
Vision models rarely consume the original image directly. They receive a resized, ordered and normalized tensor. This tool makes that hidden preprocessing step visible so shape and channel bugs can be spotted before inference.
Can I use it for object detection and segmentation models?
Yes. Letterbox 640x640 with RGB NCHW is useful for many YOLO-style object detection and segmentation pipelines, while NHWC and 224x224 presets are common for classification models.
Does the tensor computation run in WASM?
Yes. Pixlane sends the decoded frame into the Pixlane WASM runtime, where resize, channel extraction, normalization, dtype preview and heatmap rendering are computed.
Why do I see channel heatmaps instead of raw numbers?
A full 4D tensor is too large to inspect cell by cell. Heatmaps expose channel distribution, padding, crop effects and normalization ranges while the panel also shows the exact output shape.