Files
pansy/internal/imagenorm/imagenorm.go
T
steveandClaude Opus 4.8 0ab90a01cb
Build image / build-and-push (push) Successful in 8s
Gadfly review (reusable) / review (pull_request) Successful in 7m8s
Adversarial Review (Gadfly) / review (pull_request) Successful in 7m8s
EXIF orientation: bake it in when normalizing images (#103)
imagenorm decoded and re-encoded to JPEG but ignored the EXIF Orientation
tag. Phone cameras store the sensor pixels one way and set an EXIF flag to
rotate on display, so a "portrait" JPEG is really a landscape bitmap tagged
"rotate 90°" — and our re-encode strips EXIF, so without baking the rotation
in, a packet photographed in portrait reaches the vision model sideways
(bad OCR) and any future thumbnail is wrong.

- exifOrientation: a small pure-Go parser that walks the JPEG APP1/Exif
  segment for tag 0x0112, returning 1 (normal) for non-JPEG or unparseable
  input — never guess a rotation onto a correct image. No cgo, no new dep.
- applyOrientation: bakes in all 8 orientations (the 4 rotations + mirrors)
  after downscale (cheaper to rotate the small image; a 90° turn swaps the
  sides but not the longest edge, so the downscale bound still holds).
- Non-JPEG paths (HEIC/webp/png) are untouched — their decoders own
  orientation and they carry no JPEG EXIF.

Tests build oriented JPEGs (a corner marker + a spliced Exif APP1) and
assert the marker lands where each orientation says, dims swapping for the
quarter-turns; plus parser defaults for non-JPEG / no-EXIF / garbage.
Stays CGO_ENABLED=0.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01H3zbym8Doka2d7D48maSgZ
2026-07-22 01:23:08 -04:00

335 lines
12 KiB
Go
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
// Package imagenorm normalizes an uploaded image to a JPEG the rest of pansy
// (and majordomo's vision path) can rely on, decoding the formats a phone
// actually produces.
//
// # Why this exists
//
// majordomo's own media pipeline is stdlib-based, so it cannot decode HEIC or
// WebP — and HEIC is the iPhone camera default. The seed-packet feature's very
// first input is "a photo from my phone", so without this the feature fails on
// the exact device that motivates it. Normalizing at the upload boundary means
// everything downstream only ever sees JPEG.
//
// # The CGO constraint
//
// pansy is CGO_ENABLED=0 (a single static binary — the reason modernc/sqlite was
// chosen over the C one), so a libheif *binding* is out. github.com/gen2brain/heic
// runs libheif as WebAssembly via wazero: pure Go, no cgo, and it registers with
// image.Decode like any other format. golang.org/x/image/webp is pure Go too.
//
// # Import-driven registry
//
// Go's image decoders register via blank imports, and the failure mode is
// backwards from intuition: forget "image/png" and PNG uploads fail with
// "unknown format" while the exotic HEIC still works. So the blank imports below
// are load-bearing, and TestNormalizeAllFormats exercises all four formats to
// keep them so.
package imagenorm
import (
"bytes"
"encoding/binary"
"errors"
"fmt"
"image"
"image/jpeg"
"io"
"golang.org/x/image/draw"
// Decoders, registered with image.Decode by side effect. All four matter:
// jpeg/png are the common cases, heic is the iPhone default, webp is common
// on the web. Dropping any one silently breaks that format's uploads.
_ "image/jpeg"
_ "image/png"
_ "github.com/gen2brain/heic"
_ "golang.org/x/image/webp"
)
// Defaults chosen for the seed-packet path against ollama-cloud's limits (8
// images, 20 MiB, 2048px, jpeg+png). We re-encode to JPEG well under all of them.
const (
// DefaultMaxDim is the longest-edge ceiling. 2048 matches ollama-cloud's
// MaxDim; anything larger is downscaled. A packet photo has plenty of detail
// left at 2048.
DefaultMaxDim = 2048
// DefaultMaxBytes caps the *input* we will read. A phone photo is 38 MB; 25
// MiB leaves headroom for a large HEIC without inviting a decompression bomb
// as an unbounded read. The re-encoded output is far smaller.
DefaultMaxBytes = 25 << 20
// maxDecodePixels bounds the DECODED bitmap regardless of input byte size, so
// a small file claiming enormous dimensions (a decompression bomb) is refused
// before its ~4-bytes/px bitmap is allocated. 50 MP ≈ 200 MB peak — above any
// current phone camera (a 48 MP sensor is 48 MP) while capping the
// amplification a hostile header can force. maxDecodePixels and maxDimension
// are internal safety floors, not knobs — unlike MaxDim/MaxBytes there's no
// reason for a caller to raise them.
maxDecodePixels = 50_000_000
// maxDimension caps EACH side independently. It exists to make the pixel-count
// check overflow-safe: without it, a header claiming ~2^32 on a side could
// wrap int64(w)*int64(h) negative and slip past maxDecodePixels. No real image
// is 50k px on a side.
maxDimension = 50_000
// jpegQuality for the normalized output. 85 is visually clean and keeps the
// file small; the model reads text off it, not fine gradients.
jpegQuality = 85
)
// Options tunes Normalize. The zero value uses the Default* constants.
type Options struct {
MaxDim int // longest edge; 0 → DefaultMaxDim
MaxBytes int // input read cap; 0 → DefaultMaxBytes
}
func (o Options) maxDim() int {
if o.MaxDim > 0 {
return o.MaxDim
}
return DefaultMaxDim
}
func (o Options) maxBytes() int {
if o.MaxBytes > 0 {
return o.MaxBytes
}
return DefaultMaxBytes
}
// ErrTooLarge means the input exceeded the byte cap or decoded to an absurd
// pixel count. ErrUnsupported means the bytes weren't a decodable image format —
// or a decoder panicked on them (see the recover in Normalize).
var (
ErrTooLarge = errors.New("imagenorm: image too large")
ErrUnsupported = errors.New("imagenorm: unsupported or corrupt image")
)
// Normalize reads an image of any supported format (JPEG, PNG, HEIC, WebP),
// downscales it to fit opts.MaxDim on its longest edge, applies the JPEG EXIF
// orientation so the pixels come out upright, and returns it re-encoded as JPEG,
// plus the decoded format name (e.g. "heic") — handy for logging what a phone
// actually sent. On any error the returned bytes are nil and format is "".
//
// Errors, by cause:
// - input over opts.MaxBytes, or a decoded canvas over maxDecodePixels /
// maxDimension → ErrTooLarge, refused before the bitmap is allocated;
// - bytes that aren't a decodable image, or a decoder that panics on them →
// ErrUnsupported;
// - a genuine read or JPEG-encode I/O failure → a wrapped error (not a
// sentinel), since those are the caller's stream/environment, not the image.
//
// It bounds work against a hostile upload three ways: the byte cap, the
// pre-decode pixel/dimension check, and a recover around the third-party decoders
// (a malformed HEIC/WebP shouldn't take the process down).
//
// EXIF orientation: phone cameras store the sensor pixels in one orientation and
// set an EXIF tag to rotate on display, so a JPEG "portrait" photo is really a
// landscape bitmap tagged "rotate 90°" — and the re-encode below strips EXIF,
// which is exactly why the rotation must be BAKED IN here. applyOrientation does
// that for the JPEG path (the format phone uploads overwhelmingly arrive in);
// other formats carry no JPEG EXIF and their decoders own orientation, so they're
// left as decoded.
//
// One known gap, deferred to the upload handler (#81): there is no context —
// image.Decode is CPU-bound and not cancellable mid-decode, so a caller that
// needs a hard deadline should run Normalize under its own timeout. The size
// guards keep the work finite regardless.
func Normalize(r io.Reader, opts Options) (out []byte, format string, err error) {
// Cap the read at MaxBytes+1 so we can tell "exactly at the cap" from "over".
// maxBytes() is always a sane positive (default 25 MiB); guard the +1 anyway.
byteCap := opts.maxBytes()
limit := int64(byteCap) + 1
if limit < 1 {
limit = int64(DefaultMaxBytes) + 1
}
raw, err := io.ReadAll(io.LimitReader(r, limit))
if err != nil {
return nil, "", fmt.Errorf("imagenorm: read: %w", err)
}
if len(raw) > byteCap {
return nil, "", ErrTooLarge
}
// Check dimensions BEFORE a full decode, so a decompression bomb is refused
// before it allocates its bitmap. The per-side maxDimension check runs first
// so the pixel-count multiply below can't overflow.
cfg, _, err := image.DecodeConfig(bytes.NewReader(raw))
if err != nil {
return nil, "", ErrUnsupported
}
if cfg.Width <= 0 || cfg.Height <= 0 ||
cfg.Width > maxDimension || cfg.Height > maxDimension ||
int64(cfg.Width)*int64(cfg.Height) > maxDecodePixels {
return nil, "", ErrTooLarge
}
img, format, err := decodeSafely(raw)
if err != nil {
return nil, "", err
}
// Downscale first (cheaper to rotate the small image), then bake in the EXIF
// orientation so the JPEG we emit is upright. A 90° rotation swaps the sides
// but not the longest edge, so the downscale bound still holds after it.
img = downscale(img, opts.maxDim())
img = applyOrientation(img, exifOrientation(raw))
var buf bytes.Buffer
if err := jpeg.Encode(&buf, img, &jpeg.Options{Quality: jpegQuality}); err != nil {
return nil, "", fmt.Errorf("imagenorm: encode jpeg: %w", err)
}
return buf.Bytes(), format, nil
}
// decodeSafely decodes raw, converting both a decode error and a decoder PANIC
// into ErrUnsupported. The recover matters because the image comes from an
// untrusted upload and the HEIC/WebP decoders are third-party (libheif via WASM,
// x/image/webp): a malformed file that panics one of them must fail this one
// request, not crash the process.
func decodeSafely(raw []byte) (img image.Image, format string, err error) {
defer func() {
if r := recover(); r != nil {
img, format, err = nil, "", ErrUnsupported
}
}()
img, format, err = image.Decode(bytes.NewReader(raw))
if err != nil {
return nil, "", ErrUnsupported
}
return img, format, nil
}
// applyOrientation returns img with the EXIF orientation (1..8) baked in, so the
// pixels are upright and no display-time rotation is needed. Orientation 1 (and
// anything out of range) is a no-op. Values 5..8 are 90° rotations, which swap
// the output's width and height. Reads through image.Image.At and writes an RGBA,
// so it works whatever concrete type decode/downscale produced.
func applyOrientation(img image.Image, o int) image.Image {
if o <= 1 || o > 8 {
return img
}
b := img.Bounds()
w, h := b.Dx(), b.Dy()
swap := o >= 5 // a quarter-turn: output dimensions transpose
dst := image.NewRGBA(image.Rect(0, 0, w, h))
if swap {
dst = image.NewRGBA(image.Rect(0, 0, h, w))
}
for y := range h {
for x := range w {
c := img.At(b.Min.X+x, b.Min.Y+y)
var dx, dy int
switch o {
case 2: // mirror horizontal
dx, dy = w-1-x, y
case 3: // rotate 180
dx, dy = w-1-x, h-1-y
case 4: // mirror vertical
dx, dy = x, h-1-y
case 5: // transpose (mirror across the main diagonal)
dx, dy = y, x
case 6: // rotate 90° clockwise
dx, dy = h-1-y, x
case 7: // transverse (mirror across the anti-diagonal)
dx, dy = h-1-y, w-1-x
case 8: // rotate 90° counter-clockwise
dx, dy = y, w-1-x
}
dst.Set(dx, dy, c)
}
}
return dst
}
// exifOrientation extracts the EXIF Orientation tag (1..8) from raw image bytes,
// returning 1 (normal) when it's absent or unparseable — the safe default, since
// a wrong guess rotates a correct image. Only the JPEG APP1/Exif path is parsed:
// that's the format uploaded phone photos overwhelmingly arrive in, and the other
// decoders own their own orientation.
func exifOrientation(raw []byte) int {
// A JPEG is a run of FFxx marker segments after the SOI (FFD8). Walk them
// looking for APP1 (FFE1) carrying "Exif\0\0"; stop at the scan data (SOS).
if len(raw) < 4 || raw[0] != 0xFF || raw[1] != 0xD8 {
return 1
}
for i := 2; i+4 <= len(raw); {
if raw[i] != 0xFF {
return 1 // not aligned on a marker; give up rather than misread
}
marker := raw[i+1]
if marker == 0xD9 || marker == 0xDA {
return 1 // EOI / start-of-scan: no more headers to read
}
segLen := int(raw[i+2])<<8 | int(raw[i+3])
if segLen < 2 || i+2+segLen > len(raw) {
return 1
}
if marker == 0xE1 {
if o, ok := orientationFromApp1(raw[i+4 : i+2+segLen]); ok {
return o
}
}
i += 2 + segLen
}
return 1
}
// orientationFromApp1 reads the Orientation tag from a JPEG APP1 segment body
// (everything after the 2-byte length): "Exif\0\0" then a TIFF block holding
// IFD0. Returns (0, false) if the segment isn't Exif or the tag is missing.
func orientationFromApp1(seg []byte) (int, bool) {
const prefix = "Exif\x00\x00"
if len(seg) < len(prefix)+8 || string(seg[:len(prefix)]) != prefix {
return 0, false
}
tiff := seg[len(prefix):]
var bo binary.ByteOrder
switch string(tiff[0:2]) {
case "II":
bo = binary.LittleEndian
case "MM":
bo = binary.BigEndian
default:
return 0, false
}
ifd := int(bo.Uint32(tiff[4:8])) // offset to IFD0 from the TIFF start
if ifd < 8 || ifd+2 > len(tiff) {
return 0, false
}
n := int(bo.Uint16(tiff[ifd : ifd+2]))
for k := range n {
off := ifd + 2 + k*12 // each IFD entry is 12 bytes
if off+12 > len(tiff) {
return 0, false
}
if bo.Uint16(tiff[off:off+2]) != 0x0112 { // Orientation tag
continue
}
// SHORT value lives in the first 2 bytes of the entry's value field.
v := int(bo.Uint16(tiff[off+8 : off+10]))
if v >= 1 && v <= 8 {
return v, true
}
return 0, false
}
return 0, false
}
// downscale returns img shrunk so its longest edge is at most maxDim, preserving
// aspect ratio. An image already within bounds is returned unchanged (no
// re-sampling, no quality loss beyond the JPEG round-trip). Uses Catmull-Rom for
// a sharp result on text, which is what a packet photo is mostly made of.
func downscale(img image.Image, maxDim int) image.Image {
b := img.Bounds()
w, h := b.Dx(), b.Dy()
longest := max(w, h)
if longest <= maxDim || longest == 0 {
return img
}
scale := float64(maxDim) / float64(longest)
nw, nh := max(int(float64(w)*scale), 1), max(int(float64(h)*scale), 1)
dst := image.NewRGBA(image.Rect(0, 0, nw, nh))
draw.CatmullRom.Scale(dst, dst.Bounds(), img, b, draw.Over, nil)
return dst
}