neuro-media-pipeline

Use presigned URLs generated with @aws-sdk/s3-request-presigner so clients upload directly to R2, bypassing serverless function limits. The client uploads via XHR for progress tracking, then calls a confirm endpoint to save metadata to the database.

Use sharp(input).rotate().resize({fit: 'inside', withoutEnlargement: true}) then convert with .webp({quality: 80}) or .avif(). Generate responsive variants at 150/400/800/1600px widths and extract width/height metadata to prevent layout shift.

Use server proxy for files under 5MB where you need server-side optimization before storage. Use presigned URLs for larger files up to 500MB to avoid proxying through serverless functions, which have memory and timeout limits.

Sharp resize loses GIF animation frames, so animated GIFs should skip optimization and only have metadata extracted. Use gifsicle instead if GIF resizing is required.

Client-reported MIME types and extensions can be spoofed to upload malicious files. The file-type library detects the real format from magic bytes in the buffer, ensuring only allowed image, video, and document types are accepted.

BlurHash is a ~20 character string encoding a blurred image placeholder, generated by encoding a 32px Sharp resize. Display it while the full image loads to eliminate visible pop-in and improve perceived performance.