Product image preparation
A catalogue needs consistent image sizes before publication or indexing. Define the target dimensions, colour handling and acceptable image quality before choosing hardware.
Research implementation · GPU testing pending
GPU image processing for repeatable batches.
Prepare consistent images for a product catalogue, an authorised public dataset or an image-search experiment. Our first image runner focuses on one measurable operation: resizing RGB images, with reference checks and a PNG output for every item.
Create an account and describe your workload →What this batch could help with
Dataset preprocessing
An image dataset needs the same resize policy for every item. Record the operation and original ordering so a later AI experiment can be reproduced.
Large batch comparison
Compare CPU and GPU resizing on identical inputs. Include decoding, transfers and encoding in the total duration rather than timing the GPU operation alone.
How the research workload runs
- Define the image contract
Specify source dimensions, output size, RGB conversion and interpolation. The research runner uses bilinear interpolation with align_corners=false and rounded RGB PNG output.
- Run a bounded batch
Generate original synthetic gradient images in a fixed order. Process a limited number at a time so memory depends on batch size rather than the total corpus.
- Compare and retain evidence
Compare each floating-point output with a local CPU reference, save ordered PNG files and hashes, then compare with a separately generated CPU batch.
Inputs and limits
The current runner generates synthetic inputs only: 1–10,000 images, a batch size of 1–32, and source and target sides of 64–1,024 pixels. Real image uploads, image tagging and AI upscaling are separate features and are not included.
Software requirements
A tested CUDA PyTorch environment and Pillow. The runner uses no pretrained image model and downloads nothing. Its default is CUDA; it stops when CUDA is unavailable instead of silently processing on the CPU.
Results and checks
The kit returns PNG images plus a JSON report with output hashes, dimensions, interpolation policy, numerical error, GPU compute time, complete pipeline time and PyTorch allocator peaks. CPU mode is labelled as a reference run and cannot claim a GPU pass.
For the floating-point resize operation, the local CPU comparison permits a maximum absolute difference of 0.000002. The separate PNG comparison allows at most one 8-bit colour level. These are research tolerances for this operation, not a universal image quality score.
When a GPU is useful
Small resize batches can be faster and cheaper on a CPU because GPU transfers and PNG encoding add overhead. Measure both routes for the same output before deciding that GPU acceleration is useful. AI tagging, denoising and upscaling require different models and separate validation.
Useful future quote units are accepted images at a specified source/target resolution. Include transfer, encoding, independent validation, retries and storage when comparing costs.
No published performance or price is established for this new workload. The embeddings benchmark does not predict its speed or accuracy.
Try the controlled research kit
The implementation runs locally in an existing reviewed environment. It uses no IdleGPU token, creates no customer jobs, installs nothing and awards no badge. GPU compatibility and useful output quality still need a real-device test. Customer dispatch is not enabled for this workload.
Before a marketplace launch, this workload needs independent correctness checks, interruption/retry testing, a reviewed data path, workload-specific verification and an agreed acceptance rule.
Common questions
Is GPU image processing available in the marketplace?
The test runner is available for controlled research. Customer image dispatch and image workload badges are not enabled. You can create an account and describe a non-sensitive requirement.
Does the image runner generate or upscale images with AI?
No. The first operation is deterministic bilinear resizing. AI generation, tagging and learned upscaling need their own pinned model, licence review and quality evaluation.
Why does the benchmark use synthetic images?
Original synthetic gradients give every item a reproducible input without using private or unlicensed image collections. Real-data performance and visual quality still require a separate authorised evaluation.
Explore related GPU workloads
Text embeddings → GPU audio transcription → GPU rendering →