(a)
Image as human-readable plain text — no binary
Not anticipated
No prior art encodes photos as human-readable, semantically structured plain text intended as a compression codec. SVG tracing (ImageTracer.js, Potrace, Vector Magic, StarVector) outputs readable text, but traced photos typically compress worse than JPEG and aren't photo codecs. Base85 containers (AEON SAPI, ZipText/ZTAR) are text but opaque alphabet soup — and still need a full image decoder.
Closest: StarVector (arXiv:2312.11556, 2024) — a CodeLLM that writes SVG code as text tokens, but generative, not a codec. Codewars ASCII-art codec encodes ASCII art, not photos.
(b)
Progressive layers: L0 / L1 / L2 / L3
Partially anticipated
Text + edge two-layer semantic compression exists. Edge + low-res + residual layering is classic (1988–2012). But the specific four-layer text stack — semantic text, RLE text edge map, RLE text palette fills, original-file pointer — was not found.
Closest: “Text + Sketch: Image Compression at Ultra Low Rates” / PICS (arXiv:2307.01944, Jul 2023) — text + HED edge map, decoded via ControlNet / Stable Diffusion. Canon patent US 8,218,908 (2012). AGDL (Zhang & Wu, 2021). JPEG XL DC frames; SlimCAE (CVPR 2021).
(c)
Decoder-free — no NN, no special software
Not anticipated*
Not anticipated for photo-quality reconstruction. Trivially-decodable text image codecs exist only for toy / ASCII-art domains. All semantic-compression academic work (textual transform coding, PICS, PerCo, Cross-Modal Compression) requires large generative models as decoders. SVG still needs a renderer; base85 still needs decoders plus the image codec.
*The Codewars ASCII quadtree codec needs only a trivial parser — decoder-free in spirit, but for ASCII art, not photos.
(d)
Coherence: keep long strokes, drop speckle
Half anticipated
The negative half — drop speckle — is standard and 20+ years old: Potrace and ImageTracer.js despeckle via a “turd size” threshold, deleting small connected components as noise. JPEG XL patches and noise synthesis make the same move: don't spend bits on noise. The positive half — explicitly keeping long connected strokes because they carry text / outline semantics — was not found as a named mechanism outside MIC.
Also: PICS found HED edge maps beat Canny on rate-distortion — HED suppresses texture noise, i.e. an edge-coherence selection.
(e)
L0G scene graph — queryable without decoding
Partially anticipated
Textual queryability is an active research theme and per-object attribute graphs exist — but a schema-fixed scene-graph layer (bboxes + color coverage % + transcriptions + face counts) inside an image codec was not found.
Closest: LOST-3DSG (arXiv:2601.02905) — per-object label, 3D bbox, color, material, description; ~157 bytes/object, ~3 KB/scene, queryable — but for 3D robotics, built with neural encoders. Weissman textual transform coding (2023; Phil. Trans. R. Soc. A 2025). TextBoost (arXiv:2603.04115) validates exact transcriptions as side info.
(f)
50–80:1 vs JPEG, text-legible wireframe
Not anticipated as claimed
No source found claims a classical edge + palette codec reaching 50–80:1 over JPEG on photos. Generative codecs exceed the ratio by orders of magnitude — PerCo at 0.003 bpp (≈8000:1 vs raw), PICS text+sketch at ≈0.013 bpp — but with hallucinated pixels and neural decoders, not wireframe-legible reconstructions. For reference, JPEG XL and AVIF achieve only ~2–2.5:1 over JPEG at equal quality.
The report flags this plainly: MIC's claimed ratio sits in generative-compression territory, achieved here by semantic rather than generative means — see Open Questions below.