""" CFG (Classifier-Free Guidance) Animation: vector arithmetic in 2D score space @fminxyz Series 2, Post 5 — Финал серии 1080x1080 px, 2 fps, 25 сек (~50 frames) """ import numpy as np import matplotlib.pyplot as plt import matplotlib.animation as animation from matplotlib.patches import FancyArrowPatch OUTPUT = "/root/Strategy/content/drafts/cfg_vector_field_animation.mp4" FPS = 2 N_FRAMES = 50 rng = np.random.default_rng(42) # ── Color scheme (dark bg, same series palette) ─────────────────────────────── BG_COLOR = '#0a0a14' COLOR_UNCOND = '#666688' # grey — unconditional score COLOR_COND = '#4fc3f7' # blue — conditional score (cat) COLOR_CFG = '#ff4466' # red — CFG boosted COLOR_TRAJ = '#ffdd44' # yellow — sample trajectory # ── Gaussian mixture: 3 modes in 2D ────────────────────────────────────────── # Mode 0: "cat" — upper left # Mode 1: "dog" — upper right # Mode 2: "landscape" — bottom center MODES = { 'cat': {'mean': np.array([-1.8, 1.4]), 'std': 0.55, 'color': '#4fc3f7', 'label': 'cat'}, 'dog': {'mean': np.array([ 1.8, 1.4]), 'std': 0.55, 'color': '#f06292', 'label': 'dog'}, 'landscape': {'mean': np.array([ 0.0, -1.6]), 'std': 0.65, 'color': '#a5d6a7', 'label': 'landscape'}, } MODE_LIST = ['cat', 'dog', 'landscape'] MODE_WEIGHTS = np.array([1/3, 1/3, 1/3]) # Target class for conditioning: cat TARGET_CLASS = 'cat' TARGET_MEAN = MODES[TARGET_CLASS]['mean'] # ── Score functions ─────────────────────────────────────────────────────────── def gaussian_score(x, mean, std): """Score of a single Gaussian: -(x - mean) / std^2""" return -(x - mean) / (std ** 2) def mixture_score_uncond(x, weights=None): """Score of full Gaussian mixture (unconditional).""" if weights is None: weights = MODE_WEIGHTS # Compute responsibilities log_probs = [] for name in MODE_LIST: m = MODES[name] diff = x - m['mean'] log_p = -0.5 * np.sum(diff**2) / m['std']**2 log_probs.append(log_p + np.log(weights[MODE_LIST.index(name)])) log_probs = np.array(log_probs) log_probs -= log_probs.max() responsibilities = np.exp(log_probs) responsibilities /= responsibilities.sum() + 1e-12 score = np.zeros(2) for i, name in enumerate(MODE_LIST): m = MODES[name] score += responsibilities[i] * gaussian_score(x, m['mean'], m['std']) return score def mixture_score_cond(x): """Score conditioned on 'cat': upweights cat mode strongly.""" # p(x|cat) ∝ p(x) * p(cat|x), approximate as sharp focus on cat mode weights_cond = np.array([0.88, 0.08, 0.04]) # strong cat weighting return mixture_score_uncond(x, weights=weights_cond) def cfg_score(x, w): """ε_cfg = ε_uncond + w * (ε_cond - ε_uncond)""" s_u = mixture_score_uncond(x) s_c = mixture_score_cond(x) return s_u + w * (s_c - s_u) # ── Grid for vector fields ──────────────────────────────────────────────────── XLIM = (-3.2, 3.2) YLIM = (-3.0, 3.0) NX, NY = 13, 13 xg = np.linspace(XLIM[0], XLIM[1], NX) yg = np.linspace(YLIM[0], YLIM[1], NY) Xg, Yg = np.meshgrid(xg, yg) grid_pts = np.stack([Xg.ravel(), Yg.ravel()], axis=1) def compute_field(score_fn): vecs = np.array([score_fn(p) for p in grid_pts]) Vx = vecs[:, 0].reshape(Xg.shape) Vy = vecs[:, 1].reshape(Yg.shape) return Vx, Vy def normalize_field(Vx, Vy, scale=0.35): mag = np.sqrt(Vx**2 + Vy**2 + 1e-12) return Vx / mag * scale, Vy / mag * scale # Precompute fields Vx_u, Vy_u = compute_field(mixture_score_uncond) Vx_c, Vy_c = compute_field(mixture_score_cond) W_VALUES = [0.0, 1.0, 3.0, 7.0, 15.0, 30.0] W_LABELS = ['w=0 (unconditional)', 'w=1 (just conditional)', 'w=3 (light boost)', 'w=7 (Stable Diffusion default)', 'w=15 (strong)', 'w=30 (oversaturated)'] # ── Sample trajectory: follow CFG field from noise ─────────────────────────── # Compute trajectory for each w value N_TRAJ_STEPS = 60 TRAJ_START = np.array([-0.3, -0.1]) # near center (noise-like) def compute_trajectory(w, n_steps=N_TRAJ_STEPS, step_size=0.06): pos = TRAJ_START.copy() path = [pos.copy()] for _ in range(n_steps): grad = cfg_score(pos, w) mag = np.linalg.norm(grad) + 1e-8 pos = pos + step_size * grad / mag # Clamp to plot area pos[0] = np.clip(pos[0], XLIM[0]+0.1, XLIM[1]-0.1) pos[1] = np.clip(pos[1], YLIM[0]+0.1, YLIM[1]-0.1) path.append(pos.copy()) return np.array(path) trajectories = {w: compute_trajectory(w) for w in W_VALUES} # ── Frame schedule ──────────────────────────────────────────────────────────── # 50 frames total, 6 w values: each gets ~8 frames # Frame layout: # 0-1: title / intro # 2-9: w=0 (8 frames) # 10-17: w=1 (8 frames) # 18-25: w=3 (8 frames) # 26-33: w=7 (8 frames, "sweet spot" highlight) # 34-41: w=15 (8 frames) # 42-49: w=30 (8 frames, oversaturation) PHASE_STARTS = [2, 10, 18, 26, 34, 42] PHASE_ENDS = [9, 17, 25, 33, 41, 49] def get_phase(frame): """Returns (phase_idx, local_frame 0..7) or (-1, 0) for intro.""" if frame < 2: return -1, frame for i, (s, e) in enumerate(zip(PHASE_STARTS, PHASE_ENDS)): if s <= frame <= e: return i, frame - s return 5, frame - PHASE_STARTS[5] # ── Figure ──────────────────────────────────────────────────────────────────── DPI = 108 fig, ax = plt.subplots(figsize=(10, 10), dpi=DPI, facecolor=BG_COLOR) def draw_mode_blobs(ax, alpha=0.25): """Draw soft Gaussian blobs for each concept mode.""" for name in MODE_LIST: m = MODES[name] theta = np.linspace(0, 2*np.pi, 60) for r_scale, a in [(1.0, alpha), (0.6, alpha*1.4), (0.3, alpha*1.8)]: rx = m['mean'][0] + m['std'] * r_scale * np.cos(theta) ry = m['mean'][1] + m['std'] * r_scale * np.sin(theta) ax.fill(rx, ry, color=m['color'], alpha=a, zorder=1) # Mode label ax.text(m['mean'][0], m['mean'][1] + m['std'] + 0.22, m['label'], ha='center', fontsize=13, color=m['color'], fontfamily='monospace', alpha=0.85, zorder=6) def draw_frame(i): ax.clear() ax.set_facecolor(BG_COLOR) ax.set_xlim(XLIM) ax.set_ylim(YLIM) ax.set_aspect('equal') ax.axis('off') phase, local = get_phase(i) fade_in = min(1.0, (i + 1) / 4.0) # ── Intro frames ────────────────────────────────────────────────────────── if phase == -1: draw_mode_blobs(ax, alpha=0.18) ax.text(0, 2.6, 'Classifier-Free Guidance', ha='center', fontsize=19, fontweight='bold', color='white', fontfamily='monospace', alpha=fade_in) ax.text(0, 2.15, 'ε_cfg = ε_uncond + w·(ε_cond − ε_uncond)', ha='center', fontsize=13, color='#aaaaee', fontfamily='monospace', alpha=fade_in) ax.text(0, -2.6, 'три концепции в пространстве шума', ha='center', fontsize=12, color='#556677', fontfamily='monospace', alpha=fade_in) return # ── Phase frames ────────────────────────────────────────────────────────── w_val = W_VALUES[phase] w_label = W_LABELS[phase] local_frac = local / 7.0 # Fade in at start of each phase phase_alpha = min(1.0, (local + 1) / 3.0) # Draw concept blobs blob_alpha = 0.18 if phase >= 3: # w=7+: cat blob brighter draw_mode_blobs(ax, alpha=0.14) # Extra glow on cat for high w cat_m = MODES['cat'] cat_glow = min(0.5, 0.18 + (w_val - 7) * 0.008) theta = np.linspace(0, 2*np.pi, 60) ax.fill(cat_m['mean'][0] + cat_m['std']*1.2*np.cos(theta), cat_m['mean'][1] + cat_m['std']*1.2*np.sin(theta), color=cat_m['color'], alpha=cat_glow, zorder=1) else: draw_mode_blobs(ax, alpha=blob_alpha) # ── Unconditional field (always shown, grey, faint) ─────────────────────── Vx_un, Vy_un = normalize_field(Vx_u, Vy_u, scale=0.28) uncond_alpha = 0.30 if phase > 0 else 0.55 * phase_alpha ax.quiver(Xg, Yg, Vx_un, Vy_un, color=COLOR_UNCOND, alpha=uncond_alpha, scale=7.5, headwidth=3, headlength=4, width=0.0025, zorder=3) # ── Conditional field (shown from phase 1+, blue) ───────────────────────── if phase >= 1: cond_alpha = min(0.55, phase_alpha * 0.55) if phase == 1 else 0.35 Vx_cn, Vy_cn = normalize_field(Vx_c, Vy_c, scale=0.28) ax.quiver(Xg, Yg, Vx_cn, Vy_cn, color=COLOR_COND, alpha=cond_alpha, scale=7.5, headwidth=3, headlength=4, width=0.0025, zorder=3) # ── CFG field (main, red) ───────────────────────────────────────────────── if phase >= 1: Vx_cfg_raw, Vy_cfg_raw = compute_field(lambda x: cfg_score(x, w_val)) Vx_cfgn, Vy_cfgn = normalize_field(Vx_cfg_raw, Vy_cfg_raw, scale=0.30) cfg_alpha = min(0.85, phase_alpha * 0.85) lw = 0.0028 + phase * 0.0003 # slightly thicker for higher w ax.quiver(Xg, Yg, Vx_cfgn, Vy_cfgn, color=COLOR_CFG, alpha=cfg_alpha, scale=7.0, headwidth=4, headlength=5, width=lw, zorder=4) # ── Sample trajectory ───────────────────────────────────────────────────── traj = trajectories[w_val] # Show trajectory progress over local frames traj_steps = int(local_frac * len(traj)) traj_steps = max(2, traj_steps) traj_show = traj[:traj_steps] if len(traj_show) > 1: ax.plot(traj_show[:, 0], traj_show[:, 1], '-', color=COLOR_TRAJ, alpha=0.75, linewidth=2.0, zorder=7) # Current position dot cur = traj_show[-1] ax.plot(cur[0], cur[1], 'o', color=COLOR_TRAJ, markersize=10, alpha=0.95, zorder=8, markeredgecolor='white', markeredgewidth=1.0) # Start marker ax.plot(TRAJ_START[0], TRAJ_START[1], 's', color='white', markersize=7, alpha=0.6, zorder=7) # ── w value annotation (big) ────────────────────────────────────────────── # Color gradient: grey→blue→orange→red for increasing w w_colors = ['#888899', '#4fc3f7', '#ffa040', '#ff4466', '#ff1133', '#cc0022'] wc = w_colors[phase] ax.text(0, 2.62, w_label, ha='center', fontsize=15, fontweight='bold', color=wc, fontfamily='monospace', alpha=phase_alpha, zorder=9) # ── CFG formula line ────────────────────────────────────────────────────── ax.text(0, 2.28, 'ε_cfg = ε_uncond + w·(ε_cond − ε_uncond)', ha='center', fontsize=11, color='#8888aa', fontfamily='monospace', alpha=0.7, zorder=9) # ── Status annotation ───────────────────────────────────────────────────── status_msgs = { 0: 'только общее направление к данным', 1: 'guidance = conditional = нет усиления', 2: 'небольшое усиление "кошачести"', 3: 'типичный SD: баланс качество/разнообразие ✓', 4: 'сильное усиление — теряем разнообразие', 5: 'артефакты: "пережаренное" изображение ⚠️', } status_colors = ['#888899', '#4fc3f7', '#aaddff', '#44ff88', '#ffaa44', '#ff4444'] ax.text(0, -2.58, status_msgs[phase], ha='center', fontsize=12, color=status_colors[phase], fontfamily='monospace', alpha=phase_alpha, zorder=9) # ── Legend (compact, bottom right) ─────────────────────────────────────── leg_x, leg_y = 2.4, -1.6 leg_entries = [ (COLOR_UNCOND, 'ε_uncond (grey)'), (COLOR_COND, 'ε_cond (blue)'), (COLOR_CFG, 'ε_cfg (red) '), (COLOR_TRAJ, 'trajectory '), ] for k, (lc, lt) in enumerate(leg_entries): ax.text(leg_x, leg_y - k * 0.28, lt, ha='left', fontsize=9, color=lc, fontfamily='monospace', alpha=0.7, zorder=9) # ── Title ───────────────────────────────────────────────────────────────── ax.text(0, -2.85, 'CFG: экстраполяция в пространстве score', ha='center', fontsize=11, color='#445566', fontfamily='monospace', alpha=0.7, zorder=9) # ── Frame indicator ─────────────────────────────────────────────────────── ax.text(XLIM[1]-0.1, YLIM[0]+0.1, f'f={i:02d}', ha='right', fontsize=9, color='#334455', fontfamily='monospace', zorder=9) # ── Phase progress dots ─────────────────────────────────────────────────── dot_y = YLIM[0] + 0.35 dot_xs = np.linspace(-1.5, 1.5, len(W_VALUES)) for di, dx in enumerate(dot_xs): dc = '#ff4466' if di == phase else ('#445566' if di < phase else '#223344') ax.plot(dx, dot_y, 'o', color=dc, markersize=6, alpha=0.8, zorder=9) anim = animation.FuncAnimation(fig, draw_frame, frames=N_FRAMES, interval=1000 // FPS) anim.save(OUTPUT, writer='ffmpeg', fps=FPS, dpi=DPI, extra_args=['-vcodec', 'libx264', '-pix_fmt', 'yuv420p', '-crf', '22', '-preset', 'fast']) plt.close() print(f"Saved: {OUTPUT}")