diff --git a/doc/note/quirks.md b/doc/note/quirks.md
index f919da3..750b209 100644
--- a/doc/note/quirks.md
+++ b/doc/note/quirks.md
@@ -32,3 +32,8 @@
 GIF-specific quirks have a `uint32_t` value of `((0xF8586 << 10) | n)`, for
 some small integer `n`. The generated `C` language file defines human-readable
 names for those constant values.
+
+
+## Listing
+
+- [GIF image decoder quirks](std/gif/decode_quirks.wuffs)
diff --git a/doc/std/image-decoders.md b/doc/std/image-decoders.md
new file mode 100644
index 0000000..fae0c23
--- /dev/null
+++ b/doc/std/image-decoders.md
@@ -0,0 +1,131 @@
+# Image Decoders
+
+Regardless of the file format (GIF, JPEG, PNG, etc), every Wuffs image decoder
+has a similar API. This document gives `wuffs_gif__decoder__foo_bar` examples,
+but the `gif` could be replaced by `jpeg`, `png`, etc.
+
+Some image formats (GIF) are animated, consisting of an image header and then N
+frames. The image header gives e.g. the overall image's width and height. Each
+frame consists of a frame header (e.g. frame rectangle bounds, display
+duration) and a payload (the pixels).
+
+In general, there is one `wuffs_base__image_config` and then N pairs of
+(`wuffs_base__frame_config`, frame). Non-animated (still) image formats are
+treated the same way: that their N is 1 and their single frame's bounds equals
+the overall image bounds.
+
+To decode everything (without knowing N in advance) sequentially:
+
+```
+wuffs_base__io_buffer src = etc;
+wuffs_gif__decoder* dec = etc;
+
+wuffs_base__image_config ic;
+const char* status = wuffs_gif__decoder__decode_image_config(dec, &ic, &src);
+// Error checking (inspecting the status variable) is not shown, for brevity.
+// See the example programs, listed below, for how to handle errors.
+
+// Allocate the pixel buffer, based on ic's width and height, etc.
+wuffs_base__pixel_buffer pb = etc;
+
+while (true) {
+  wuffs_base__frame_config fc;
+  status = wuffs_gif__decoder__decode_frame_config(dec, &fc, &src);
+  if (status == wuffs_base__warning__end_of_data) {
+    break;
+  }
+  // Ditto re error checking.
+
+  status = wuffs_gif__decoder__decode_frame(dec, &pb, &etc);
+  // Ditto re error checking.
+}
+```
+
+The second argument to each `decode_xxx` method (`&ic`, `&fc` or `&pb`), is the
+destination data structure to store the decoded information.
+
+For random (instead of sequential) access to an image's frames, call
+`wuffs_gif__decoder__restart_frame(dec, i, io_pos)` to prepare to decode the
+i'th frame. Essentially, it restores the state to be at the top of the while
+loop above. The `wuffs_base__io_buffer`'s reader position will also need to be
+set to the `io_pos` position in the source data stream. The position for the
+i'th frame is calculated by the i'th `decode_frame_config` call and saved in
+the `frame_config`. You can only call `restart_frame` after
+`decode_image_config` is called, explicitly or implicitly (see below), as
+decoding a single frame might require for-all-frames information like the
+overall image dimensions and the global palette.
+
+All of those `decode_xxx` calls are optional. For example, if
+`decode_image_config` is not called, then the first `decode_frame_config` call
+will implicitly parse and verify the image header, before parsing the first
+frame's header. Similarly, you can call only `decode_frame` N times, without
+calling `decode_image_config` or `decode_frame_config`, if you already know
+metadata like N and each frame's rectangle bounds by some other means (e.g.
+this is a first party, statically known image).
+
+Specifically, starting with an unknown (but re-windable) animated image, if you
+want to just find N (i.e. count the number of frames), you can loop calling
+only the `decode_frame_config` method and avoid calling the more expensive
+`decode_frame` method. For GIF, in terms of the underlying wire format, this
+will skip over (instead of decompress) the LZW-encoded pixel data, which is
+faster.
+
+Those `decode_xxx` methods are also suspendible
+[coroutines](/doc/note/coroutines.md). They will return early (with a status
+code that `is_suspendible` and therefore isn't `is_complete`) if there isn't
+enough source data to complete the operation: an incremental decode. Calling
+`decode_xxx` again with additional source data will resume the previous
+operation instead of starting a new one. Calling `decode_yyy` whilst
+`decode_xxx` is suspended will result in an error.
+
+Once an error is encountered, whether from invalid source data or from a
+programming error, such as calling `decode_yyy` while suspended in
+`decode_xxx`, all subsequent calls will be no-ops that return an error. To
+reset the decoder into something that does productive work,
+[re-initialize](/doc/note/initialization.md) and then, in order to be able to
+call `restart_frame`, call `decode_image_config`. The `io_buffer` and its
+associated stream will also need to be rewound.
+
+
+## Metadata
+
+Images can also contain metadata (e.g. color profiles, time stamps). By
+default, Wuffs' image decoders skip over metadata, but calling
+`set_report_metadata` will opt in to having `decode_image_config` return early
+when encountering metadata in the file. Calling `set_report_metadata` can be
+done multiple times, each with a different
+[FourCC](/doc/note/base38-and-fourcc.md) code such as `0x49434350` "ICCP" or
+`0x584D5020` "XMP ", to indicate what sorts of metadata the caller is
+interested in. Conversely, when the parser encounters metadata (and returns a
+"@metadata reported" [status](/doc/note/statuses.md)), call `metadata_fourcc`
+to see what sort of metadata it is.
+
+Embedded metadata needs to be processed by a separate parser. For example,
+processing XMP metadata usually involves some sort of XML parser, regardless of
+what particular image format that XMP metadata was embedded in. That metadata
+might also be in multiple (non-contiguous) chunks. The caller needs to loop,
+repeatedly calling `metadata_chunk_length`, advancing the `io_buffer` by that
+many bytes (after diverting those bytes to the separate parser) and calling
+`ack_metadata_chunk`. If the latter returns "@metadata reported", then repeat
+the loop. If it returns ok, then the metadata was completely consumed, and the
+caller can go back to the `decode_image_config` method.
+
+
+## Implementations
+
+- [std/gif](/std/gif)
+
+
+## Examples
+
+- [example/gifplayer](/example/gifplayer)
+
+
+## [Quirks](/doc/note/quirks.md)
+
+- [GIF decoder quirks](std/gif/decode_quirks.wuffs)
+
+
+---
+
+See also the general remarks on [Wuffs' standard library](/doc/std/README.md).