Image compressors on Kodak

Side-by-side rate-distortion and encoding-throughput plots, sourced from the JSONs hardcoded below. RD metrics (bpp, PSNR, SSIM, DISTS) come from rate_distortion_<id>.json measured at native Kodak resolution; encode throughput (MPx/s) comes from encode_<id>.json measured on 512x512 center crops. The two are joined on the swept dimension so each operating point shares an encoder configuration but not a sample geometry.

Pick which codecs to plot by commenting / uncommenting lines in the CODECS list below.

Entropy-coding caveat for the throughput panel. The timed “encode” region of the cheng2020 and mbt2018 harnesses is a likelihood-only forward pass — no actual bitstream is produced (the rANS coder is not vendored), so their throughput is generous relative to the codecs whose timed region pays for real byte production (JPEG, AVIF, WaLLoC, LiVeAction, FRAPPE, MCUCoder). Those results are kept as-is. The first MS-ILLM / NoGAN-MS encode JSONs (results/msillm/encode_1786549370.json, results/noganms/encode_1786549820.json) were measured under the same generous convention; the harness has since been changed to include real entropy coding (a constriction range coder driven by the model’s own distributions) in a separate store stage, and later encode JSONs from compressors.msillm.evaluate_encode_complexity reflect that — check each JSON’s config.timed_region when comparing.

Show code cell source

Hide code cell source

import json
from pathlib import Path
import matplotlib.pyplot as plt
plt.rcParams["font.family"] = "serif"

Show code cell source

Hide code cell source

RD_PATHS = {
    "frappe":          "results/frappe/rate_distortion_1777315303.json",
    "frappe_v2_image": "results/frappe_v2_image/rate_distortion_1782987823.json",
    "jpeg":            "results/jpeg/rate_distortion_1777321251.json",
    "avif_default":    "results/avif/rate_distortion_1777321273.json",
    "avif_speed0":     "results/avif/rate_distortion_1777567112.json",
    "avif_speed10":    "results/avif/rate_distortion_speed10_1777561281.json",
    "jxl":             "results/jxl/rate_distortion_1777567114.json",
    "ldm_sr":          "results/ldm_sr_openimages/rate_distortion_1777418424.json",
    "vq_diff_ithq":    "results/vq_diffusion_ithq/rate_distortion_1777418464.json",
    "kandinsky_2_1":   "results/kandinsky_2_1/rate_distortion_1777418494.json",
    "stable_cascade":  "results/stable_cascade/rate_distortion_1777418663.json",
    "cheng2020":       "results/cheng2020/rate_distortion_1777567118.json",
    "msillm":          "results/msillm/rate_distortion_1786548517.json",
    "noganms":         "results/noganms/rate_distortion_1786548918.json",
    "mbt2018":         "results/mbt2018/rate_distortion_1777567120.json",
    "walloc":          "results/walloc/rate_distortion_1777562174.json",
    "liveaction":      "results/liveaction/rate_distortion_1777567116.json",
    "mcucoder_native": "results/mcucoder/rate_distortion_1777674947_native.json",
    "mcucoder_224":    "results/mcucoder/rate_distortion_1777679959_224.json",
    "e2a2_image":      "results/e2a2_image/rate_distortion_1788381387.json",
}

EC_PATHS = {
    "frappe":          "results/frappe/encode_1777342722.json",
    "frappe_v2_image": "results/frappe_v2_image/encode_1782987851.json",
    "jpeg":            "results/jpeg/encode_1777321267.json",
    "avif_default":    "results/avif/encode_1777321337.json",
    "avif_speed0":     "results/avif/encode_1777572007.json",
    "avif_speed10":    "results/avif/encode_speed10_1777561318.json",
    "jxl":             "results/jxl/encode_1777587942.json",
    "ldm_sr":          "results/ldm_sr_openimages/encode_PENDING_.json",
    "vq_diff_ithq":    "results/vq_diffusion_ithq/encode_PENDING_.json",
    "kandinsky_2_1":   "results/kandinsky_2_1/encode_1777587591.json",
    "stable_cascade":  "results/stable_cascade/encode_PENDING_.json",
    "cheng2020":       "results/cheng2020/encode_1777582109.json",
    "msillm":          "results/msillm/encode_1786549370.json",
    "noganms":         "results/noganms/encode_1786549820.json",
    "mbt2018":         "results/mbt2018/encode_1777585754.json",
    "walloc":          "results/walloc/encode_1777571761.json",
    "liveaction":      "results/liveaction/encode_1777571794.json",
    "mcucoder_native": "results/mcucoder/encode_1777680434.json",
    "mcucoder_224":    "results/mcucoder/encode_1777680434.json",
}

AXIS_LABEL = {
    "bpp":            "Rate [bits per pixel]",
    "throughput_MPx": "Encoding throughput [MPx/s]",
    "PSNR_dB":        "PSNR [dB]",
    "SSIM":           "SSIM",
    "DISTS_dB":       "DISTS [dB]",
}
TITLE_LABEL = {
    "bpp":            "Rate",
    "throughput_MPx": "Throughput",
    "PSNR_dB":        "PSNR",
    "SSIM":           "SSIM",
    "DISTS_dB":       "DISTS",
}
LOG_KEYS = ("bpp", "throughput_MPx")

Show code cell source

Hide code cell source

def load_codec(name):
    """RD points (+ throughput when an EC JSON exists for `name`; codecs with
    RD-only results get throughput_MPx = None and belong in CODECS_RD only)."""
    rd = json.loads(Path(RD_PATHS[name]).read_text())
    ec = json.loads(Path(EC_PATHS[name]).read_text()) if name in EC_PATHS else None
    swept_key = "channel_counts" if "channel_counts" in rd else "quality_values"
    points = []
    for k in rd[swept_key]:
        s = str(k)
        if ec is not None and s not in ec["results"]:
            continue
        rm = rd["results"][s]["mean"]
        tp = ec["results"][s]["throughput"]["median_MPx_per_s"] if ec is not None else None
        points.append({
            "sweep": k,
            "bpp": rm["bpp"],
            "PSNR_dB": rm["PSNR_dB"],
            "SSIM": rm["SSIM"],
            "DISTS_dB": rm["DISTS_dB"],
            "throughput_MPx": tp,
        })
    return points

def col(pts, key):
    return [p[key] for p in pts]

Show code cell source

Hide code cell source

def plot_panel(group, y_key, x_key, legend=True):
    """One small panel: y vs x for every codec in `group`.

    bpp and throughput axes are log-scaled wherever they appear.
    """
    plt.figure(figsize=(4.5, 4), dpi=180)
    ax = plt.gca()
    if x_key in LOG_KEYS:
        ax.set_xscale("log")
    if y_key in LOG_KEYS:
        ax.set_yscale("log")
    for c in group:
        ax.plot(col(c["data"], x_key), col(c["data"], y_key),
                marker=c.get("marker", "."),
                linestyle=c.get("linestyle", "-"),
                color=c.get("color"), label=c["name"])
    ax.set_xlabel(AXIS_LABEL[x_key])
    ax.set_ylabel(AXIS_LABEL[y_key])
    ax.set_title(f"{TITLE_LABEL[y_key]} vs {TITLE_LABEL[x_key]} (Kodak)")
    if legend:
        ax.legend(loc="best")
    log_either = (x_key in LOG_KEYS) or (y_key in LOG_KEYS)
    ax.grid(True, which="both" if log_either else "major", alpha=0.4)
    plt.tight_layout()
    plt.show()

Show code cell source

Hide code cell source

# Comment / uncomment to control which codecs appear on the plots
CODECS = [
    {"name": "JPEG",             "data": load_codec("jpeg"),           "color": "gray",   "linestyle": "-", "marker": "."},
    # {"name": "JPEG-XL",          "data": load_codec("jxl"),            "color": "tab:brown",  "linestyle": "-", "marker": "."},
    {"name": "AVIF",   "data": load_codec("avif_default"),   "color": "black", "linestyle": "-", "marker": "."},
    # {"name": "AVIF (speed=0)",   "data": load_codec("avif_speed0"),    "color": "tab:red",    "linestyle": "-", "marker": "."},
    # {"name": "AVIF (speed=10)",  "data": load_codec("avif_speed10"),   "color": "tab:pink",   "linestyle": "-", "marker": "."},
    {"name": "mbt2018",          "data": load_codec("mbt2018"),        "color": "red",       "linestyle": "-", "marker": "."},
    {"name": "cheng2020-anchor", "data": load_codec("cheng2020"),      "color": "firebrick",       "linestyle": "-", "marker": "."},
    # {"name": "LDM-SR",            "data": load_codec("ldm_sr"),         "color": "tab:purple", "linestyle": "-", "marker": "."},
    # {"name": "VQ-Diffusion ITHQ", "data": load_codec("vq_diff_ithq"),   "color": "tab:gray",   "linestyle": "-", "marker": "."},
    {"name": "Kandinsky 2.1",     "data": load_codec("kandinsky_2_1"),  "color": "orange",  "linestyle": "-", "marker": "."},
    # {"name": "Stable Cascade",    "data": load_codec("stable_cascade"), "color": "tab:cyan",   "linestyle": "-", "marker": "."},
    {"name": "WaLLoC",           "data": load_codec("walloc"),         "color": "blue",    "linestyle": "-", "marker": "."},
    {"name": "LiVeAction",       "data": load_codec("liveaction"),     "color": "green",      "linestyle": "-", "marker": "."},
    # {"name": "MCUCoder",         "data": load_codec("mcucoder_native"), "color": "green",    "linestyle": "-", "marker": "."},
    # {"name": "MCUCoder (224x224)", "data": load_codec("mcucoder_224"), "color": "tab:green",  "linestyle": "-", "marker": "."},
    {"name": "FRAPPE",           "data": load_codec("frappe"),         "color": "tab:purple",  "linestyle": "-", "marker": "."},
    {"name": "FRAPPE v2",        "data": load_codec("frappe_v2_image"), "color": "purple",  "linestyle": "-", "marker": "."},
]

# MS-ILLM / NoGAN-MS encode timings were measured on a different testbed
# than the controlled one used for every other codec, so they are excluded
# from the throughput panel and appear only on the PSNR vs bpp plot. E2A2
# image has no encode-complexity JSON yet (RD only), so it is likewise
# RD-only.
CODECS_RD = CODECS + [
    {"name": "MS-ILLM",          "data": load_codec("msillm"),         "color": "magenta",    "linestyle": "-", "marker": "."},
    {"name": "NoGAN-MS",         "data": load_codec("noganms"),        "color": "darkcyan",   "linestyle": "-", "marker": "."},
    {"name": "E2A2 image",       "data": load_codec("e2a2_image"),     "color": "tab:blue",   "linestyle": "-", "marker": "."},
]

for c in CODECS_RD:
    tps = [t for t in col(c['data'], 'throughput_MPx') if t is not None]
    tp_str = f"[{min(tps):.3f}, {max(tps):.3f}] MPx/s" if tps else "n/a (RD only)"
    print(f"{c['name']:>20}: {len(c['data'])} points,"
          f" bpp [{min(col(c['data'],'bpp')):.4f}, {max(col(c['data'],'bpp')):.4f}],"
          f" throughput {tp_str}")

                JPEG: 21 points, bpp [0.1731, 6.7867], throughput [212.354, 543.966] MPx/s
                AVIF: 21 points, bpp [0.0681, 5.8218], throughput [1.970, 6.037] MPx/s
             mbt2018: 8 points, bpp [0.1102, 1.5837], throughput [0.156, 0.174] MPx/s
    cheng2020-anchor: 6 points, bpp [0.1174, 0.8041], throughput [0.046, 0.094] MPx/s
       Kandinsky 2.1: 3 points, bpp [0.0625, 0.2500], throughput [0.207, 0.757] MPx/s
              WaLLoC: 10 points, bpp [0.0097, 0.6171], throughput [14.188, 74.832] MPx/s
          LiVeAction: 9 points, bpp [0.0341, 0.6921], throughput [1.542, 1.932] MPx/s
              FRAPPE: 21 points, bpp [0.0042, 0.9418], throughput [59.575, 914.938] MPx/s
           FRAPPE v2: 9 points, bpp [0.0109, 4.8747], throughput [21.299, 831.277] MPx/s
             MS-ILLM: 8 points, bpp [0.0036, 0.7054], throughput [0.571, 0.708] MPx/s
            NoGAN-MS: 6 points, bpp [0.0413, 0.7016], throughput [0.566, 0.717] MPx/s
          E2A2 image: 5 points, bpp [0.0107, 3.0311], throughput n/a (RD only)

Show code cell source

Hide code cell source

plot_panel(CODECS_RD, "PSNR_dB",  "bpp")

Show code cell source

Hide code cell source

plot_panel(CODECS, "throughput_MPx", "PSNR_dB", legend=False)

Show code cell source

Hide code cell source

plot_panel(CODECS, "SSIM",     "bpp",            legend=False)

Show code cell source

Hide code cell source

plot_panel(CODECS, "DISTS_dB", "bpp",            legend=False)