fairtalking-second-work / scripts /plot_cta_framework.py
huahua123313's picture
Add files using upload-large-folder tool
a17b394 verified
Raw
History Blame Contribute Delete
12.3 kB
"""Render the CTA framework figure (paper-style).
Style is inspired by images/{1,2,3}.jpg:
* rounded rectangles with pastel fills (pink / lavender / mint / cream)
* black arrows, dashed for detach / auxiliary paths
* snowflake glyph for frozen(ish) / stop-grad boundaries
* math annotations for l_av, l_va, asym
Output: images/cta_framework.png (and .pdf)
"""
from __future__ import annotations
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import FancyBboxPatch, FancyArrowPatch, Rectangle
from matplotlib.lines import Line2D
# ---------- palette (matched to images/*.jpg) --------------------------------
COL_PINK = "#F7C6C7" # main blocks
COL_LAVEND = "#D9C7EA" # cross-modal predictor blocks
COL_MINT = "#CFE9D7" # loss / metric blocks
COL_CREAM = "#F6E7C1" # classifier / feature vector
COL_BLUE = "#CBDDF2" # backbone
COL_BG = "#FFFFFF"
COL_TEXT = "#111111"
COL_MUTED = "#666666"
COL_ACCENT = "#B22222" # asym highlight (dark red)
FIG_W, FIG_H = 17.0, 11.5
def rbox(ax, x, y, w, h, facecolor, text, fontsize=11, edgecolor="black", lw=1.4,
boxstyle="round,pad=0.02,rounding_size=0.10", weight="normal", zorder=3):
"""Rounded-rectangle labelled block."""
patch = FancyBboxPatch(
(x, y), w, h,
boxstyle=boxstyle,
facecolor=facecolor, edgecolor=edgecolor, linewidth=lw, zorder=zorder,
)
ax.add_patch(patch)
ax.text(x + w / 2, y + h / 2, text,
ha="center", va="center", fontsize=fontsize,
color=COL_TEXT, weight=weight, zorder=zorder + 1)
return patch
def arrow(ax, xy_from, xy_to, color="black", lw=1.6, style="-", mutation_scale=14,
zorder=4, curved=False, rad=0.0):
connector = f"arc3,rad={rad}" if curved else "arc3,rad=0"
a = FancyArrowPatch(
xy_from, xy_to,
arrowstyle="-|>", mutation_scale=mutation_scale,
color=color, lw=lw, linestyle=style,
connectionstyle=connector, zorder=zorder,
)
ax.add_patch(a)
return a
def snowflake(ax, x, y, size=0.13, color="#2F6DB5"):
"""Simple 6-arm snowflake glyph -- marks stop-grad / detach."""
import numpy as np
for k in range(6):
ang = np.deg2rad(60 * k)
dx, dy = size * np.cos(ang), size * np.sin(ang)
ax.plot([x - dx, x + dx], [y - dy, y + dy], color=color, lw=1.2, zorder=6)
ax.plot([x], [y], "o", ms=3.2, color=color, zorder=7)
def circle_op(ax, x, y, r=0.11, glyph="+", color="black"):
from matplotlib.patches import Circle
ax.add_patch(Circle((x, y), r, facecolor="white", edgecolor=color, lw=1.4, zorder=6))
ax.text(x, y, glyph, ha="center", va="center", fontsize=10, weight="bold", zorder=7)
# =============================================================================
fig, ax = plt.subplots(figsize=(FIG_W, FIG_H))
ax.set_xlim(0, 17)
ax.set_ylim(0, 11.5)
ax.set_aspect("equal")
ax.set_axis_off()
fig.patch.set_facecolor(COL_BG)
# ------ Title ---------------------------------------------------------------
ax.text(8.5, 11.10,
"CTA — Cross-modal Asymmetry for Talking-Face Deepfake Detection",
ha="center", va="center", fontsize=17, weight="bold")
ax.text(8.5, 10.65,
r"Core signal: $\mathrm{asym} \;=\; \ell_{V\!A}-\ell_{AV}$"
r" (direction-signed reconstruction gap)",
ha="center", va="center", fontsize=12.5, color=COL_ACCENT, style="italic")
# ------ Inputs (left column) -------------------------------------------------
rbox(ax, 0.35, 8.15, 1.55, 1.05, COL_BG,
"Video clip\n" r"$X_v \in \mathbb{R}^{T\times3\times H\times W}$",
fontsize=10, edgecolor="#333")
rbox(ax, 0.35, 6.15, 1.55, 1.05, COL_BG,
"Audio clip\n" r"$X_a$",
fontsize=10, edgecolor="#333")
# ------ Backbones ------------------------------------------------------------
rbox(ax, 2.35, 8.10, 2.05, 1.15, COL_BLUE, "Video Backbone\n(TimeSformer)", fontsize=10.5)
rbox(ax, 2.35, 6.10, 2.05, 1.15, COL_BLUE, "Audio Backbone\n(HF encoder)", fontsize=10.5)
# arrows: X -> backbone
arrow(ax, (1.90, 8.67), (2.35, 8.67))
arrow(ax, (1.90, 6.67), (2.35, 6.67))
# backbone output labels: tokens / pooled
ax.text(4.55, 9.10, r"$v_\text{tok}$", fontsize=11, color=COL_TEXT)
ax.text(4.55, 8.65, r"$v_\text{pool}$", fontsize=11, color=COL_TEXT)
ax.text(4.55, 7.10, r"$a_\text{tok}$", fontsize=11, color=COL_TEXT)
ax.text(4.55, 6.65, r"$a_\text{pool}$", fontsize=11, color=COL_TEXT)
# ------ Cross-modal predictors (center) -------------------------------------
rbox(ax, 5.60, 8.10, 2.55, 1.15, COL_LAVEND,
r"$\mathbf{A\!\to\!V}$ predictor" "\n(cross-attention)",
fontsize=10.5)
rbox(ax, 5.60, 6.10, 2.55, 1.15, COL_LAVEND,
r"$\mathbf{V\!\to\!A}$ predictor" "\n(cross-attention)",
fontsize=10.5)
arrow(ax, (4.40, 8.65), (5.60, 8.65))
arrow(ax, (4.40, 6.90), (5.60, 8.35), curved=True, rad=-0.28)
arrow(ax, (4.40, 6.65), (5.60, 6.65))
arrow(ax, (4.40, 8.40), (5.60, 7.00), curved=True, rad=0.28)
ax.text(5.05, 8.78, "query", fontsize=8, color=COL_MUTED)
ax.text(5.05, 8.12, "src (a)", fontsize=8, color=COL_MUTED)
ax.text(5.05, 6.78, "query", fontsize=8, color=COL_MUTED)
ax.text(5.05, 7.10, "src (v)", fontsize=8, color=COL_MUTED)
# ------ Per-sample losses ---------------------------------------------------
rbox(ax, 8.55, 8.35, 2.10, 0.80, COL_MINT,
r"$\ell_{AV}=\mathrm{MSE}(\hat v_\text{tok},v_\text{tok})$",
fontsize=10)
rbox(ax, 8.55, 6.35, 2.10, 0.80, COL_MINT,
r"$\ell_{VA}=\mathrm{MSE}(\hat a_\text{tok},a_\text{tok})$",
fontsize=10)
arrow(ax, (8.15, 8.67), (8.55, 8.75))
arrow(ax, (8.15, 6.67), (8.55, 6.75))
ax.text(9.60, 8.15, "(per-sample, keeps batch dim)", fontsize=8, color=COL_MUTED,
ha="center", style="italic")
ax.text(9.60, 6.15, "(per-sample, keeps batch dim)", fontsize=8, color=COL_MUTED,
ha="center", style="italic")
# ------ Asymmetry node ------------------------------------------------------
rbox(ax, 11.35, 7.15, 2.75, 1.05, COL_PINK,
r"$\mathrm{asym} = \ell_{VA}-\ell_{AV}$" "\n(direction-signed gap)",
fontsize=11.5, weight="bold", edgecolor=COL_ACCENT, lw=2.0)
arrow(ax, (10.65, 8.75), (11.60, 8.20), curved=True, rad=-0.20)
arrow(ax, (10.65, 6.75), (11.60, 7.20), curved=True, rad=0.20)
ax.text(11.00, 7.98, r"$\ell_{AV}$", fontsize=10, color=COL_TEXT)
ax.text(11.00, 6.95, r"$\ell_{VA}$", fontsize=10, color=COL_TEXT)
# ------ Feature vector for classifier ---------------------------------------
rbox(ax, 11.35, 4.60, 2.75, 1.10, COL_CREAM,
r"$[\,v_\text{pool},\,a_\text{pool},\,\mathrm{asym},\,\ell_{AV}\!+\!\ell_{VA}\,]$",
fontsize=10)
arrow(ax, (4.40, 8.40), (11.35, 5.55), curved=True, rad=-0.30, color=COL_MUTED)
arrow(ax, (4.40, 6.40), (11.35, 5.30), curved=True, rad=0.30, color=COL_MUTED)
arrow(ax, (12.72, 7.15), (12.72, 5.70))
# l_av / l_va (detached) via dashed
arrow(ax, (9.60, 8.35), (12.00, 5.70), curved=True, rad=-0.22, style=(0, (5, 4)), color="#333")
arrow(ax, (9.60, 6.35), (12.00, 5.70), curved=True, rad=0.22, style=(0, (5, 4)), color="#333")
snowflake(ax, 11.05, 7.55)
snowflake(ax, 11.05, 5.85)
ax.text(10.55, 7.80, ".detach()", fontsize=8, color="#2F6DB5")
ax.text(10.55, 5.60, ".detach()", fontsize=8, color="#2F6DB5")
# ------ Classifier ----------------------------------------------------------
rbox(ax, 11.55, 3.05, 2.30, 1.00, COL_PINK,
"Classifier head\n(Linear-GELU-Do-Linear)", fontsize=10)
arrow(ax, (12.72, 4.60), (12.72, 4.05))
rbox(ax, 14.15, 3.05, 1.85, 1.00, COL_BG,
r"$\hat y = \sigma(\mathrm{logit})$" "\nreal / fake",
fontsize=10, edgecolor="#333")
arrow(ax, (13.85, 3.55), (14.15, 3.55))
# ------ Losses row (mid-bottom) --------------------------------------------
rbox(ax, 0.35, 3.35, 3.00, 0.90, COL_MINT,
r"$\mathcal{L}_{AV}=\mathrm{mean}_{y=0}\,\ell_{AV}$"
"\n" r"$\mathcal{L}_{VA}=\mathrm{mean}_{y=0}\,\ell_{VA}$",
fontsize=10)
ax.text(1.85, 3.05, "trained ONLY on real",
ha="center", fontsize=8, color=COL_ACCENT, style="italic")
rbox(ax, 3.85, 3.35, 3.55, 0.90, COL_MINT,
r"$\mathcal{L}_\text{asym}=\mathrm{ReLU}"
r"(\bar{\mathrm{asym}}_\text{fake}-\bar{\mathrm{asym}}_\text{real})$",
fontsize=10)
ax.text(5.625, 3.05, "margin=0 -- keep signal weak (predictor no-collapse)",
ha="center", fontsize=8, color=COL_ACCENT, style="italic")
rbox(ax, 7.90, 3.35, 3.10, 0.90, COL_MINT,
r"$\mathcal{L}_\text{aux}=\mathrm{MSE}(\mathrm{asym},\mathrm{asym}')$",
fontsize=10)
ax.text(9.45, 3.05, "same (ref, audio) -> two fakes",
ha="center", fontsize=8, color=COL_ACCENT, style="italic")
rbox(ax, 14.05, 4.30, 1.95, 0.65, COL_MINT,
r"$\mathcal{L}_\text{cls}=\mathrm{BCE}(\hat y, y)$",
fontsize=10)
arrow(ax, (13.85, 3.55), (15.00, 4.30), curved=True, rad=-0.28)
# arrows: sources -> loss boxes
arrow(ax, (11.35, 7.35), (7.40, 4.25), curved=True, rad=0.25)
arrow(ax, (11.35, 7.10), (11.00, 4.25), curved=True, rad=0.10)
arrow(ax, (8.55, 8.35), (3.35, 4.25), curved=True, rad=0.35, color=COL_MUTED)
arrow(ax, (8.55, 6.35), (3.35, 4.25), curved=True, rad=-0.15, color=COL_MUTED)
# ------ Total loss ----------------------------------------------------------
rbox(ax, 5.10, 1.95, 6.20, 0.95, COL_CREAM,
r"$\mathcal{L}=w_{av}\mathcal{L}_{AV}"
r"+w_{va}\mathcal{L}_{VA}"
r"+w_\text{asym}\mathcal{L}_\text{asym}"
r"+w_\text{cls}\mathcal{L}_\text{cls}"
r"+w_\text{aux}\mathcal{L}_\text{aux}$",
fontsize=11.5, weight="bold")
for x_from in [1.85, 5.625, 9.45]:
arrow(ax, (x_from, 3.35), (8.20, 2.90), color=COL_MUTED)
arrow(ax, (15.00, 4.30), (11.30, 2.90), color=COL_MUTED, curved=True, rad=-0.15)
# ------ Callouts (design notes, moved to bottom) ----------------------------
callouts = [
r"$\bigstar$ predictor only sees real -> learns real audio-visual coupling as prior",
r"$\bigstar$ $\ell_{AV},\ell_{VA}$ detached into classifier -> BCE cannot distort predictors",
r"$\bigstar$ asym is signed (not $|\cdot|$) -> direction of coupling breakage IS the signature",
r"$\bigstar$ $\mathcal{L}_\text{aux}$ pulls asym across generators together -> generator-agnostic",
]
notes_x, notes_y = 0.35, 1.30
for i, t in enumerate(callouts):
row = i // 2
col = i % 2
ax.text(notes_x + col * 8.0,
notes_y - 0.42 * row,
t, fontsize=10, color=COL_ACCENT)
# ------ Legend --------------------------------------------------------------
legend_x, legend_y = 15.10, 10.20
ax.text(legend_x + 0.65, legend_y + 0.05, "Legend",
fontsize=10, weight="bold", ha="center")
# swatches
sw_specs = [
(COL_BLUE, "backbone"),
(COL_LAVEND, "cross-modal predictor"),
(COL_MINT, "loss / per-sample metric"),
(COL_PINK, "core module (asym / cls)"),
(COL_CREAM, "feature vector / total"),
]
for i, (c, lab) in enumerate(sw_specs):
y = legend_y - 0.25 - i * 0.28
ax.add_patch(Rectangle((legend_x, y), 0.28, 0.20,
facecolor=c, edgecolor="black", lw=0.8))
ax.text(legend_x + 0.38, y + 0.10, lab, fontsize=8.5, va="center")
# dashed = detach; snowflake = stop-grad
y = legend_y - 0.25 - 5 * 0.28
ax.plot([legend_x + 0.02, legend_x + 0.26], [y + 0.10, y + 0.10],
linestyle=(0, (5, 4)), color="black", lw=1.4)
ax.text(legend_x + 0.38, y + 0.10, "detach / stop-grad", fontsize=8.5, va="center")
y -= 0.28
snowflake(ax, legend_x + 0.14, y + 0.10, size=0.08)
ax.text(legend_x + 0.38, y + 0.10, "snowflake = .detach()", fontsize=8.5, va="center")
y -= 0.28
ax.annotate("", xy=(legend_x + 0.26, y + 0.10),
xytext=(legend_x + 0.02, y + 0.10),
arrowprops=dict(arrowstyle="-|>", color="black", lw=1.2))
ax.text(legend_x + 0.38, y + 0.10, "forward flow", fontsize=8.5, va="center")
# ------ Save ---------------------------------------------------------------
import os
OUT_DIR = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "images")
os.makedirs(OUT_DIR, exist_ok=True)
png_path = os.path.join(OUT_DIR, "cta_framework.png")
pdf_path = os.path.join(OUT_DIR, "cta_framework.pdf")
plt.savefig(png_path, dpi=220, bbox_inches="tight", facecolor=COL_BG)
plt.savefig(pdf_path, bbox_inches="tight", facecolor=COL_BG)
print("saved:", png_path)
print("saved:", pdf_path)