ImageConfiguration Class

DecodingStrictness.Lenient accepts the deviations that the reference decoders and the browsers accept, in identification, eager loads, sequential readers and collections alike. The resource limits are the same in both modes. Encoders and the other operations ignore the value.

Currently, ImageProcessingExtensions.Resize, the transformations, ImageProcessingExtensions.Convolve, the blurs (ImageProcessingExtensions.GaussianBlur, ImageProcessingExtensions.BoxBlur, ImageProcessingExtensions.MedianBlur), ImageProcessingExtensions.Sharpen and ImageProcessingExtensions.DetectEdges use it: the output rows of a large frame (at least 65,536 output pixels and 32 rows per worker; nearest-neighbor resizing excepted) are split into at most this many bands processed concurrently, the calling thread included. The result is identical whatever the value. Each worker needs its own row scratch, charged to the image's ImageResourceLimits.MaxLiveAllocationBytes budget. The ImageProcessingExtensions.Pixelate, ImageProcessingExtensions.OilPaint and ImageProcessingExtensions.Flatten effects split the rows the same way (the bands of the pixelation start on a row of blocks, and those of the oil painting are at least as tall as its window). The comparisons of ImageSimilarity (SSIM, PSNR, SSIMULACRA 2 and Butteraugli) use the value of the image of the expected frame: their floating-point sums are made by bands of 64 rows, added in order, so their results do not depend on it either. The HEIC and AVIF decoders use it too: the tiles of a grid of at least 65,536 samples are decoded by up to this many workers, each holding the decoder and the picture of its tile, with the pixels and the errors of a decoding by one worker. The HEIC and AVIF encoders code the tiles of such a grid (HeicEncoder.GridTileSize, AvifEncoder.GridTileSize) by up to this many workers, each with encoders of its own, and write the same bytes as one worker. The AVIF decoder also runs the post-processing filters of an AV1 frame of at least 65,536 pixels (the loop filter, CDEF, the super-resolution upscaling and the loop restoration) by bands of rows, with the pixels of a decoding by one worker, unless the frame is a tile decoded alongside other tiles. They take about 25% to 40% of the decoding time of a 3840x2160 image that uses CDEF, which then decodes about 1.3 to 1.45 times faster with 8 workers. Each worker needs its own rows, about 140 bytes per pixel of the width of the frame, charged to ImageResourceLimits.MaxLiveAllocationBytes. The JPEG XL encoder, the JPEG XL decoder and JpegXlTranscoder use it as well: the groups of a frame of at least 65,536 pixels (squares of 256x256 pixels, which the format codes independently) are encoded or decoded by up to this many workers, each with its own working memory, and the restoration filters and the conversion of a decoded frame to its color space are split into row bands; the bytes, the pixels and the errors are those of one worker. The codecs of the other formats split the independent parts of a frame of at least 65,536 pixels the same way, with the bytes, the pixels and the errors of one worker: the JPEG decoder and encoder transform batches of MCU rows (the entropy coding stays on the calling thread), the TIFF decoder decompresses the LZW, Deflate and CCITT strips and tiles and the TIFF encoder compresses the LZW and Deflate strips, the OpenEXR decoder and encoder process the chunks, the JPEG 2000 decoder and encoder code the code-blocks and split the wavelet transforms, the AVIF decoder and encoder code the tiles of an AV1 frame that has several (unless the frame is a grid tile coded alongside other tiles), and the HEIC decoder decodes the tiles and the rows of the wavefront of a picture that start at entry points. The other decoders (the HEVC in-loop filters included), the encoders and the other operations run on the calling thread.

Keep the default in services that already process several images concurrently: the bound applies per operation, so n concurrent operations may use up to n times this many threads.

Identification, eager loads, sequential readers and collections read them alike. The other formats ignore them.