from pathlib import Path from PIL import Image, ImageDraw, ImageFont import textwrap ROOT = Path(__file__).resolve().parents[1] OUT_DIRS = [ ROOT / "doc" / "assets", ROOT / "out" / "proposal_materials" / "assets", ] FONT_CANDIDATES = [ "/System/Library/Fonts/STHeiti Medium.ttc", "/System/Library/Fonts/Hiragino Sans GB.ttc", "/System/Library/Fonts/Supplemental/Arial Unicode.ttf", "/System/Library/Fonts/Helvetica.ttc", ] def font(size, bold=False): for p in FONT_CANDIDATES: if Path(p).exists(): return ImageFont.truetype(p, size=size) return ImageFont.load_default() F_TITLE = font(42) F_SUBTITLE = font(25) F_HEAD = font(24) F_BODY = font(21) F_SMALL = font(18) def text_size(draw, text, fnt): box = draw.textbbox((0, 0), text, font=fnt) return box[2] - box[0], box[3] - box[1] def wrap_text(text, width_chars): lines = [] for paragraph in text.split("\n"): if not paragraph: lines.append("") continue lines.extend(textwrap.wrap(paragraph, width=width_chars, break_long_words=False)) return lines def draw_wrapped(draw, xy, text, fnt, fill, width_chars, line_gap=7): x, y = xy for line in wrap_text(text, width_chars): draw.text((x, y), line, font=fnt, fill=fill) y += text_size(draw, line or " ", fnt)[1] + line_gap return y def round_rect(draw, box, radius, fill, outline=None, width=2): draw.rounded_rectangle(box, radius=radius, fill=fill, outline=outline, width=width) def arrow(draw, start, end, fill=(60, 70, 90), width=4): draw.line([start, end], fill=fill, width=width) x1, y1 = start x2, y2 = end if abs(x2 - x1) >= abs(y2 - y1): direction = 1 if x2 >= x1 else -1 pts = [(x2, y2), (x2 - 16 * direction, y2 - 9), (x2 - 16 * direction, y2 + 9)] else: direction = 1 if y2 >= y1 else -1 pts = [(x2, y2), (x2 - 9, y2 - 16 * direction), (x2 + 9, y2 - 16 * direction)] draw.polygon(pts, fill=fill) def save_all(img, name): for d in OUT_DIRS: d.mkdir(parents=True, exist_ok=True) img.save(d / name) def build_literature_bridge(): W, H = 1800, 1320 img = Image.new("RGB", (W, H), "white") d = ImageDraw.Draw(img) bg = (247, 249, 252) img.paste(bg, [0, 0, W, H]) d.text((70, 50), "文献方法如何支撑本项目方法设计", font=F_TITLE, fill=(20, 30, 45)) d.text((72, 105), "不直接复制论文原图；根据方法部分重画为 proposal 专用方法映射图", font=F_SUBTITLE, fill=(90, 100, 115)) cols = [70, 635, 1210] widths = [500, 500, 500] y0 = 175 headers = ["已阅读/核对的文献线索", "论文中的方法思想", "本项目采用的模块"] colors = [(232, 241, 255), (235, 248, 241), (255, 244, 229)] for x, w, h, c in zip(cols, widths, headers, colors): round_rect(d, (x, y0, x + w, y0 + 70), 18, c, outline=(195, 205, 220)) tw, th = text_size(d, h, F_HEAD) d.text((x + (w - tw) / 2, y0 + 21), h, font=F_HEAD, fill=(25, 40, 60)) rows = [ ( "T2V-CompBench\nGeckoNum\nT2ICountBench\nDemystifying Numerosity", "数值/计数能力可单独评测；prompt refinement 不可靠；显式 layout/noise guidance 和自动 reward 有效。", "把普通数字从 prompt 中拆出来，设计 quantity-specific numerical token vocabulary，并用可测代理量评估。", ), ( "Textual Inversion", "冻结生成模型，用少量图像学习新的 pseudo-word embedding，使新概念可被 prompt 调用。", "学习物理量 token / anchor token；token 表示数值区间、倍率和单位类型，而不是人物/物体/风格。", ), ( "QuantiPhy\nPhysQuantAgent\nGuidedSceneGen", "物理量估计需要尺度、几何和视觉锚点；VLM 数值判断会有 qualitative 与 quantitative gap。", "第二阶段做 unit-anchor calibration；用液面、轨迹、亮度、OCR、几何测量替代纯 VLM judge。", ), ( "DenseDPO\nHuViDPO\nGPO", "DPO/LoRA 可用于视频后训练；segment-level preference 更适合动态视频；groupwise ranking 信息密度高。", "用 groupwise preference 学 1<2<4<8；动态量采用 segment reward；LoRA/adapter 降低训练成本。", ), ( "DiffPhy\nPhysVideoGenerator\nWan/Hunyuan/LongCat", "物理上下文、latent physical tokens 和开放 video foundation models 可作为训练与推理基础。", "以 Wan/Hunyuan/LongCat 为 baseline；引入物理上下文和 quantity token 条件，但核心创新聚焦单位量 grounding。", ), ] row_h = 164 y = y0 + 95 for idx, row in enumerate(rows): fill = (255, 255, 255) if idx % 2 == 0 else (252, 253, 255) for x, w in zip(cols, widths): round_rect(d, (x, y, x + w, y + row_h), 18, fill, outline=(215, 222, 232)) for col_idx, text in enumerate(row): x = cols[col_idx] + 25 if col_idx == 0: draw_wrapped(d, (x, y + 28), text, F_BODY, (35, 62, 112), 26, 8) else: draw_wrapped(d, (x, y + 24), text, F_BODY, (45, 55, 70), 32, 8) arrow(d, (cols[0] + widths[0] + 12, y + row_h // 2), (cols[1] - 12, y + row_h // 2)) arrow(d, (cols[1] + widths[1] + 12, y + row_h // 2), (cols[2] - 12, y + row_h // 2)) y += row_h + 24 d.text((70, H - 70), "结论：这些论文不是直接解决 physical unit quantity video generation，但分别支撑“数值 token 化、相对偏好训练、单位锚定、自动评测”的设计。", font=F_SMALL, fill=(80, 90, 105)) save_all(img, "literature_method_bridge_zh.png") def build_pipeline(): W, H = 1800, 1200 img = Image.new("RGB", (W, H), (248, 250, 252)) d = ImageDraw.Draw(img) d.text((70, 50), "本项目方法：Quantity Token → Relative Training → Unit Anchor", font=F_TITLE, fill=(20, 30, 45)) d.text((72, 106), "目标：把物理数值从普通 prompt 中拆出，使模型先学相对量，再学单位 1 的视觉/物理锚点", font=F_SUBTITLE, fill=(90, 100, 115)) # Top input split y = 180 boxes = [ (90, y, 500, y + 130, "原始 prompt", "A cup with 1.0 L water\nA car moves at 2 m/s"), (650, y, 1060, y + 130, "场景 prompt", "杯子/小车/相机/背景\n保持稳定"), (1210, y, 1620, y + 130, "物理量字段", "quantity type + value + unit\nvolume: 1.0 L"), ] for x1, y1, x2, y2, h, b in boxes: round_rect(d, (x1, y1, x2, y2), 22, (255, 255, 255), outline=(202, 213, 226)) d.text((x1 + 25, y1 + 20), h, font=F_HEAD, fill=(25, 40, 60)) draw_wrapped(d, (x1 + 25, y1 + 60), b, F_BODY, (70, 80, 95), 30, 6) arrow(d, (500, y + 65), (650, y + 65)) arrow(d, (1060, y + 65), (1210, y + 65)) # Vocabulary and embedding y2 = 390 round_rect(d, (90, y2, 690, y2 + 210), 24, (235, 248, 241), outline=(175, 210, 190)) d.text((120, y2 + 25), "QuantityUnitBench / Token Schema", font=F_HEAD, fill=(20, 80, 55)) draw_wrapped(d, (120, y2 + 70), "统计物理量类型、单位范围、分桶方式和可测代理信号。\n例：pH=0-14；volume=0.1-2.0L；speed=log bins。", F_BODY, (45, 65, 55), 42, 7) round_rect(d, (780, y2, 1560, y2 + 210), 24, (232, 241, 255), outline=(175, 195, 230)) d.text((810, y2 + 25), "Quantity-specific Token Embedding", font=F_HEAD, fill=(35, 70, 125)) draw_wrapped(d, (810, y2 + 70), ", , , \n借鉴 Textual Inversion：学习少量 token embedding，但含义是物理量数值/倍率。", F_BODY, (45, 60, 85), 56, 7) arrow(d, (690, y2 + 105), (780, y2 + 105)) # Model core y3 = 690 round_rect(d, (90, y3, 520, y3 + 240), 24, (255, 244, 229), outline=(225, 190, 145)) d.text((120, y3 + 25), "视频基座模型", font=F_HEAD, fill=(110, 70, 20)) draw_wrapped(d, (120, y3 + 70), "Wan / Hunyuan / LongCat\n冻结 backbone，训练 LoRA / adapter / token embedding。", F_BODY, (75, 65, 50), 32, 7) round_rect(d, (650, y3, 1030, y3 + 240), 24, (255, 255, 255), outline=(205, 215, 225)) d.text((680, y3 + 25), "阶段一：相对量", font=F_HEAD, fill=(45, 55, 70)) draw_wrapped(d, (680, y3 + 70), "同一场景生成：\n1x, 2x, 4x, 8x, 16x, 32x\nGPO/DPO/GRPO 学排序和倍率。", F_BODY, (55, 65, 80), 27, 7) round_rect(d, (1160, y3, 1600, y3 + 240), 24, (255, 255, 255), outline=(205, 215, 225)) d.text((1190, y3 + 25), "阶段二：单位锚定", font=F_HEAD, fill=(45, 55, 70)) draw_wrapped(d, (1190, y3 + 70), "给定 1 unit reference：\n刻度杯/标尺/弹簧/灯泡。\n学习 0.5x, 1x, 2x, 4x。", F_BODY, (55, 65, 80), 31, 7) arrow(d, (520, y3 + 120), (650, y3 + 120)) arrow(d, (1030, y3 + 120), (1160, y3 + 120)) arrow(d, (1170, y2 + 210), (920, y3), fill=(55, 95, 145)) # Bottom reward/eval y4 = 1010 round_rect(d, (90, y4, 790, y4 + 120), 22, (240, 244, 248), outline=(205, 215, 225)) d.text((120, y4 + 22), "自动/半自动 reward", font=F_HEAD, fill=(45, 55, 70)) draw_wrapped(d, (120, y4 + 60), "液面高度、像素长度、轨迹斜率、亮度、OCR、tracking、人工抽检", F_BODY, (65, 75, 88), 55, 6) round_rect(d, (920, y4, 1600, y4 + 120), 22, (240, 244, 248), outline=(205, 215, 225)) d.text((950, y4 + 22), "Reference-conditioned editing", font=F_HEAD, fill=(45, 55, 70)) draw_wrapped(d, (950, y4 + 60), "先固定 0 物理量/初始视频，再用数值 token 像 edit prompt 一样改变目标物理量。", F_BODY, (65, 75, 88), 52, 6) arrow(d, (790, y4 + 60), (920, y4 + 60)) save_all(img, "quantity_token_pipeline_zh.png") if __name__ == "__main__": build_literature_bridge() build_pipeline()