Creating a PDF from images involves constructing a valid PDF document from scratch — building the file header, object hierarchy, page tree, image resources, and cross-reference table that together form a compliant PDF file. Understanding this construction process reveals how images become pages and what determines the quality and size of the resulting document.

Each image becomes an XObject Image resource within the PDF. The process begins with reading the source image file and determining how to embed it. JPEG images have a significant advantage: their compressed data can be embedded directly using the DCTDecode filter, meaning the PDF wraps the raw JPEG data stream in a PDF object without decompressing and recompressing it. This is why converting JPEG to PDF is fast and perfectly lossless — the pixel data is identical bit-for-bit to the original JPEG file. The image XObject dictionary specifies DCTDecode as the filter, along with the image width, height, bits per component (typically 8), and color space (DeviceRGB for color JPEGs, DeviceGray for grayscale).

PNG images require different handling because PDF does not have a native PNG filter. The PNG pixel data must be extracted from the PNG file structure, which involves parsing the PNG chunks (IHDR for dimensions and color type, IDAT for compressed pixel data, and optionally tRNS or PLTE for transparency and palette information). The raw pixel data, once decompressed from the PNG's zlib stream, can be re-compressed using PDF's FlateDecode filter, which uses the same zlib algorithm. PNG alpha channels are handled by creating a separate SMask (soft mask) image object that stores the transparency data independently from the color data, since PDF's image model separates color and transparency.

Page construction for each image involves creating a page dictionary that specifies the MediaBox (page dimensions), a Resources dictionary containing the image XObject reference, and a content stream with the drawing operators that place the image on the page. The content stream typically contains a graphics state save (q), a transformation matrix (cm) that scales and positions the image, the image drawing operator (Do), and a graphics state restore (Q). The transformation matrix is a six-element array that maps the image's unit square (the image occupies a 1x1 area in its native coordinate system) to the desired position and size on the page.

Page layout algorithms determine how images are sized and positioned on the page. The most common approaches are fit-to-page (scaling the image to fill the page while maintaining aspect ratio), actual-size (placing the image at its native pixel dimensions based on its DPI), and fill-page (scaling to cover the entire page, potentially cropping edges). For fit-to-page, the algorithm calculates the scale factor as the minimum of (page width - margins) / image width and (page height - margins) / image height, then centers the scaled image within the available area.

The page tree is constructed as a root Pages node with a Kids array referencing all page objects, and a Count entry specifying the total number of pages. The document catalog references this page tree, and the cross-reference table at the end of the file indexes every object's byte offset. Finally, the file trailer points to the cross-reference table and the catalog, completing the PDF structure. The entire construction follows the PDF specification's requirements for a minimal valid document, ensuring the result opens correctly in any compliant PDF reader.