博客

  • 居家健身房伸展的亚裔女性人像

    居家健身房伸展的亚裔女性人像

    居家健身房伸展的亚裔女性人像



    🤖 ChatGPT

    🇺🇸 English Prompt

    A lifestyle fitness photograph of a young East Asian woman in her early 20s with soft, youthful features, standing in a bright home gym. Her black hair is pulled up in a messy high ponytail with loose wispy bangs and a few strands framing her face. Both arms are raised behind her head with hands clasped at the back of her neck, as if she is stretching after a workout — elbows out wide, opening up her chest and shoulders. Her eyes are gently closed, lips slightly parted, cheeks flushed pink, the serene expression of someone savoring a deep post-exercise stretch. Small pearl stud earrings. She has a dramatically curvaceous hourglass figure — very full, heavy bust straining against the fabric, an exaggeratedly narrow cinched waist creating a deep inward curve at the sides, and wide round hips flaring out sharply below, with thick toned thighs — the kind of extreme waist-to-hip ratio that makes the one-piece cling and stretch in different tensions across bust, waist, and hips. She wears a dark charcoal form-fitting one-piece athletic romper with a scoop neckline, cap sleeves, and bike-short length hem that vacuum-seals to every contour of her body, a small white brand logo on the upper left chest. She stands facing the camera straight on, weight evenly distributed, legs slightly apart. To her left, partially in frame, is a dark metal cable weight machine with an orange adjustment knob. To her right, a large bird-of-paradise plant in a white ceramic pot. The background is a warm cream wall with natural sunlight streaming in from the right, casting her soft shadow on the wall behind her. Shot from a slightly low frontal angle at about hip height looking gently upward, medium framing from mid-thigh to just above her raised elbows. Soft natural window light, warm and even, 50mm lens, shallow depth of field with the gym equipment slightly blurred, clean editorial fitness aesthetic with natural skin texture, warm neutral tones.

    🇨🇳 中文提示词

    一张生活方式健身照片,捕捉了一位20岁出头的年轻东亚女性,面容柔和且富有青春气息,站在明亮的居家健身房中。她的黑发扎成凌乱的高马尾,留着轻盈的碎刘海,几缕发丝垂在脸庞。双臂抬起放在头后,双手扣在颈后,仿佛在运动后拉伸——手肘向两侧张开,舒展胸部和肩膀。她双眼微闭,嘴唇微张,脸颊泛着粉红色的红晕,表情宁静,仿佛在享受运动后的深度拉伸。佩戴着小小的珍珠耳钉。她拥有极具曲线感的沙漏型身材——丰满且厚重的胸部紧绷着面料,极其纤细的收腰在两侧形成深凹的曲线,宽阔圆润的双胯向下方大幅张开,大腿结实有力——这种极端的腰臀比使得连体衣在胸部、腰部和臀部呈现出不同的拉伸张力。她穿着一件深炭灰色的修身运动连体裤,采用圆领、短袖和骑行短裤长度的剪裁,像真空密封般贴合身体的每一处轮廓,左上胸部有一个白色的小品牌标志。她正对着镜头站立,重心分布均匀,双腿略微分开。在她的左侧,部分出现在画面中的是一台带有橙色调节旋钮的深色金属拉力器。在她的右侧,是一盆栽在白色陶瓷花盆里的巨大天堂鸟植物。背景是温暖的奶油色墙壁,自然阳光从右侧洒入,在身后的墙上投下她柔和的影子。从略低的正面角度拍摄,约在胯部高度轻轻向上看,中景构图,范围从大腿中部到略高于抬起的手肘。柔和的自然窗光,温暖且均匀,50mm镜头,浅景深使健身器材略显模糊,干净的社论健身美学,具有自然的皮肤纹理,温暖的中性色调。
  • 真实世界杯足球宝贝日系甜妹高清人像

    真实世界杯足球宝贝日系甜妹高清人像

    真实世界杯足球宝贝日系甜妹高清人像



    🤖 ChatGPT

    🇺🇸 English Prompt

    Generate a vertical 9:16 high-definition realistic photographic portrait with the theme of a real-world World Cup soccer girl. The overall style is youthful, relaxed, and very memorable. Close-up mobile phone shooting perspective, low angle slightly shooting upwards, the upper body of the person leans significantly towards the camera, hands holding a small football high above the head, the arms, football, and body forming a stable triangular composition, with an exaggerated and cute perspective. The head is slightly tilted up and then looking down at the camera, the chin is slightly raised, a bright smile, full of sweet girl vibes. An adult young Japanese woman, with lively and sweet facial features, real and delicate skin, preserving natural skin textures. Slender and well-proportioned body, clear shoulders and neck, natural collarbones. Hairstyle is high twin tails, matched with football bows, colorful beads, and small trophy stud earrings. The makeup is full of energy, with small football and heart-shaped face paintings on the cheeks, and fine sequins under the eyes. Wearing a blue, white, and pink three-color spliced cropped cheering top, paired with a high-waisted pleated mini skirt and a light transparent baseball jacket. The clothing is covered with footballs, numbers, team slogans, trophies, and lawn patterns. Shoelaces, wristbands, and waist accessories all incorporate football elements. The background is a real World Cup outdoor viewing square or fan cheering area, including giant screens, flags, light boards, fences, and cheering crowds. The background is slightly blurred, with vivid colors and a strong sense of viral spread.

    🇨🇳 中文提示词

    生成一张竖版 9:16 高清真实摄影人像,主题为真实世界的世界杯足球宝贝,整体风格青春、轻松、非常有记忆
    点。近距离手机拍摄视角,低机位轻微仰拍,人物上半身明显向镜头前探,双手把一个小型足球高高举在头顶上
    方,手臂、足球和身体形成稳定三角构图,透视夸张而可爱。头部微微仰起再向下看镜头,下巴略扬,笑容灿
    烂,甜妹感十足。
    
    成年日本年轻女性,五官灵动清甜,皮肤真实细腻,保留自然皮肤纹理。身材纤细匀称,肩颈清晰,锁骨自然。
    发型为高双马尾,搭配足球蝴蝶结、彩色串珠和小奖杯耳钉。妆容元气十足,脸颊有小型足球和爱心彩绘,眼下
    带细亮片。
    
    穿蓝白粉三色拼接短款应援上衣,搭配高腰百褶迷你裙和轻薄透明棒球服外套,服装上布满足球、号码、球队标
    语、奖杯和草坪纹样。鞋带、腕带和腰饰都融入足球元素。
    
    背景为真实世界杯露天观赛广场或球迷应援区,包含巨幕、旗帜、灯牌、围栏和欢呼人群。背景轻微虚化,色彩
    鲜明,极具爆款传播感。
  • 极简浴缸时尚人像摄影

    极简浴缸时尚人像摄影

    极简浴缸时尚人像摄影



    🤖 ChatGPT

    🇺🇸 English Prompt

    Create a vertical 9:16 ultra-realistic empty bathtub fashion portrait.
    
    An adult young woman is curled up inside a clean white bathtub, wearing [OUTFIT]. Side-lying curled pose, naturally bent legs, soft window light, quiet expression, simple styling, realistic skin texture, natural hair, real fabric details, elegant upper-body shape, slim waist, graceful curves, clean minimal bathroom, private fashion editorial mood.
    
    Use a soft cinematic look with low-saturation colors, natural shadows, clear details, and tasteful composition.
    
    Avoid water, bubbles, bathing scene, oily skin, heavy filters, plastic skin, exaggerated body, cheap sexy pose, messy background, awkward hands, distorted legs, and obvious AI face.

    🇨🇳 中文提示词

    创建一个垂直 9:16 的超现实空浴缸时尚人像。
    
    一名年轻成年女性蜷缩在干净的白色浴缸内,穿着 [OUTFIT]。侧卧蜷缩姿势,自然弯曲的双腿,柔和的窗光,恬静的表情,简约造型,真实的皮肤纹理,自然的头发,真实的织物细节,优雅的上半身轮廓,纤细的腰部,优美的曲线,干净简约的浴室,私密时尚社论氛围。
    
    使用带有低饱和度色彩、自然阴影、清晰细节和高雅构图的柔和电影感外观。
    
    避免出现水、泡沫、沐浴场景、油腻皮肤、厚重滤镜、塑料皮肤、夸张身体、廉价性感姿势、混乱背景、尴尬的手部、扭曲的双腿以及明显的 AI 脸。
  • 咖啡材料来源与生命周期信息图

    咖啡材料来源与生命周期信息图

    咖啡材料来源与生命周期信息图



    🤖 ChatGPT

    🇺🇸 English Prompt

    {
      "title": "COFFEE — MATERIALS ORIGIN & LIFECYCLE",
      "type": "Ultra-detailed circular sustainability infographic",
      "style": {
        "palette": [
          "coffee brown",
          "kraft beige",
          "cream",
          "sage green",
          "charcoal"
        ],
        "look": "premium editorial infographic, industrial design report, photorealistic, technical diagrams, clean typography"
      },
      "center": {
        "coffee_bag": "kraft pouch labeled SPECIALTY, SINGLE ORIGIN, COLOMBIA, MEDIUM ROAST",
        "props": [
          "ceramic mug with black coffee",
          "bowl of roasted coffee beans",
          "scattered beans"
        ],
        "material_diagram": [
          "beans",
          "paper bag",
          "valve",
          "foil liner",
          "label",
          "adhesive",
          "ink"
        ]
      },
      "sections": [
        {
          "id": "02",
          "title": "Raw Origin",
          "items": [
            "coffee cherries",
            "green beans",
            "soil",
            "water",
            "sunlight",
            "jute sack",
            "cardboard",
            "polymer pellets"
          ]
        },
        {
          "id": "03",
          "title": "Manufacturing",
          "steps": [
            "harvest",
            "ferment",
            "wash",
            "dry",
            "sort",
            "roast",
            "grind",
            "package",
            "ship",
            "brew"
          ]
        },
        {
          "id": "04",
          "title": "Use & Aging",
          "items": [
            "aroma loss",
            "oxidation",
            "staling",
            "oil bloom",
            "moisture",
            "package wear"
          ]
        },
        {
          "id": "05",
          "title": "Repair & Maintenance",
          "items": [
            "airtight storage",
            "reseal clips",
            "cleaning",
            "reusable filters",
            "grinder repair"
          ]
        },
        {
          "id": "06",
          "title": "End of Life",
          "items": [
            "composting",
            "recycling",
            "landfill risks",
            "parts harvesting"
          ]
        },
        {
          "id": "07",
          "title": "Sustainable Redesign",
          "items": [
            "recyclable pouch",
            "refill system",
            "compostable filters",
            "soy inks",
            "regenerative farming"
          ]
        }
      ],
      "eco_scorecard": {
        "Durability": 6.5,
        "Recyclability": 4,
        "Repairability": 6,
        "Material Honesty": 7.5,
        "Packaging Efficiency": 6.5,
        "Upgrade Potential": 7
      },
      "footer": "ORIGIN → MAKE → USE → MAINTAIN → END → REDESIGN → BETTER FUTURE",
      "quality": "8K, photorealistic, print-quality poster"
    }

    🇨🇳 中文提示词

    {
      "title": "咖啡 — 材料来源与生命周期",
      "type": "超详细圆形可持续性信息图",
      "style": {
        "palette": [
          "咖啡棕",
          "牛皮纸米色",
          "奶油色",
          "鼠尾草绿",
          "木炭色"
        ],
        "look": "高级社论信息图,工业设计报告,写实,技术图表,整洁排版"
      },
      "center": {
        "coffee_bag": "贴有精品、单一产地、哥伦比亚、中度烘焙标签的牛皮纸袋",
        "props": [
          "盛有黑咖啡的陶瓷马克杯",
          "一碗烘焙过的咖啡豆",
          "散落的豆子"
        ],
        "material_diagram": [
          "咖啡豆",
          "纸袋",
          "气阀",
          "铝箔内衬",
          "标签",
          "粘合剂",
          "油墨"
        ]
      },
      "sections": [
        {
          "id": "02",
          "title": "原始产地",
          "items": [
            "咖啡樱桃",
            "生豆",
            "土壤",
            "水",
            "阳光",
            "麻袋",
            "纸板",
            "聚合物颗粒"
          ]
        },
        {
          "id": "03",
          "title": "生产制造",
          "steps": [
            "收获",
            "发酵",
            "清洗",
            "干燥",
            "分拣",
            "烘焙",
            "研磨",
            "包装",
            "运送",
            "冲泡"
          ]
        },
        {
          "id": "04",
          "title": "使用与老化",
          "items": [
            "香气流失",
            "氧化",
            "陈化",
            "油脂析出",
            "水分",
            "包装磨损"
          ]
        },
        {
          "id": "05",
          "title": "维修与保养",
          "items": [
            "密封存储",
            "密封夹",
            "清洁",
            "可重复使用滤网",
            "研磨机维修"
          ]
        },
        {
          "id": "06",
          "title": "生命周期终点",
          "items": [
            "堆肥",
            "回收",
            "填埋风险",
            "零件回收"
          ]
        },
        {
          "id": "07",
          "title": "可持续重设计",
          "items": [
            "可回收包装袋",
            "补充系统",
            "可堆肥滤纸",
            "大豆油墨",
            "再生农业"
          ]
        }
      ],
      "eco_scorecard": {
        "耐用性": 6.5,
        "可回收性": 4,
        "可维修性": 6,
        "材料真实性": 7.5,
        "包装效率": 6.5,
        "升级潜力": 7
      },
      "footer": "来源 → 制造 → 使用 → 维护 → 终点 → 重设计 → 更美好的未来",
      "quality": "8K, 写实, 印刷质量海报"
    }
  • 像素织纹端午龙舟节庆海报

    像素织纹端午龙舟节庆海报

    像素织纹端午龙舟节庆海报



    🤖 ChatGPT

    🇨🇳 中文提示词

    Generate a bright and clear festive graphic visual around any theme object: The visual first reads as a coarse-grained pixel weave graphic translated from the core outline of the theme. The graphic does not aim for realistic depiction but resembles a symbolic dark pattern woven into a light, bright background. Edges appear intermittently as stepped blocks. The subject and background maintain a monochromatic value difference, ensuring the image has texture when viewed up close and forms a complete memorable point from a distance. Below, continuous undulating horizontal ripples or theme-derived rhythmic patterns are laid out as a supporting layer and reading buffer; the lines stack layer by layer but maintain a sense of ventilation. The information layer uses extremely thin, elongated, and elegant high-contrast fonts for the main title. The font size can overpower the texture layer, with spacious letter spacing, supplemented by a small amount of small-font dates, blessings, seal-style phrases, and miniature brand information at the bottom. All text floats gently on the screen like transparent white light. Colors are extracted from the material, season, and cultural mood of the theme itself, mapped into bright high-value background colors, slightly deeper theme dark patterns, clean structural shadows, a few clear information colors, and slight grainy shadows, maintaining a mood that is clear and quiet, fresh and clean, low-contrast but distinct in layers; Textures can include fine paper fibers, vertical fabric lines, soft grains, and a slight printed feel. The overall style is restrained, with ample whitespace and precise information order, avoiding a realistic illustration feel, heavy or muddy colors, complex decorations, and crowded layouts. —————— The visual is to be replaced with: Dragon Boat racing, retaining the direction of a festive poster with a pixel-weave dragon boat and water wave support. For WeChat Moments use: As a Dragon Boat Festival greeting poster, the main title can be written as DRAGON BOAT FESTIVAL or 端午安康. Size requirements: Vertical 9:16, overall still clear light green, with a small amount of fine white text.
  • 高端图形化时尚摄影海报设计

    高端图形化时尚摄影海报设计

    高端图形化时尚摄影海报设计



    🤖 ChatGPT

    🇺🇸 English Prompt

    Create a high-end graphic fashion poster with strong color, typography, fashion photography and brand mood.
    
    Theme: [streetwear / dance / fashion film / designer brand / lifestyle brand]
    Direction: [color block street poster / floral motion typography / type mask portrait / geometric space fashion poster]
    Brand / project name: [name]
    Main title: [short headline]
    Subtitle: [short subtitle]
    Model / subject: [model, dancer, portrait, product, fashion look]
    Color palette: [blue-yellow / orange-black-white / monochrome / white-navy-grey]
    Mood: [fresh / graphic / minimal / playful / artistic / modern / editorial]
    Aspect ratio: [9:16]
    
    The poster should feel like a premium fashion design piece, brand campaign poster, lookbook cover or editorial visual.
    
    Use bold typography, clean layout, strong negative space, realistic fashion photography, graphic color blocks, geometric structure, floral or abstract elements when needed, and a clear hierarchy between image, type and space.
    
    Choose one direction:
    
    1. Color block street poster
       Use a large brand wordmark, side-walking model, bold color blocks, small color palette swatches, flower accents and a fresh streetwear mood.
    
    2. Floral motion typography
       Use a fashion model or dancer with elegant body movement, oversized letterforms, fabric motion, floral / orange accent elements and refined editorial typography.
    
    3. Type mask portrait
       Use a black-and-white portrait cut through huge vertical typography. Let the face appear inside the letterforms, like a fashion film poster or design magazine cover.
    
    4. Geometric space fashion poster
       Place the model inside a clean white geometric space, cube, doorway or architectural structure. Use minimal typography, soft shadows and designer-brand styling.
    
    Keep the result high-end, graphic, clean, fashionable and visually designed.
    
    Avoid cheap template design, messy stickers, random decorations, low-end poster style, weak typography, distorted body, bad hands, fake plastic skin, cluttered layout, bad text rendering and real copyrighted brand logos.

    🇨🇳 中文提示词

    创建一个具有强烈色彩、排版、时尚摄影和品牌氛围的高端图形时尚海报。
    
    主题:[街头服饰 / 舞蹈 / 时尚电影 / 设计师品牌 / 生活方式品牌]
    方向:[色块街头海报 / 花卉动态排版 / 字体遮罩肖像 / 几何空间时尚海报]
    品牌 / 项目名称:[名称]
    主标题:[短标题]
    副标题:[短副标题]
    模特 / 主体:[模特、舞者、肖像、产品、时尚造型]
    调色板:[蓝黄 / 橙黑白 / 单色 / 白海军蓝灰]
    氛围:[清新 / 图形化 / 极简 / 俏皮 / 艺术 / 现代 / 编辑感]
    纵横比:[9:16]
    
    海报应该感觉像一件高端时尚设计作品、品牌活动海报、画册封面或编辑视觉效果。
    
    使用大胆的排版、干净的布局、强烈的留白、真实的时尚摄影、图形色块、几何结构,必要时使用花卉或抽象元素,并在图像、字体和空间之间建立清晰的层级关系。
    
    选择一个方向:
    
    1. 色块街头海报
       使用大型品牌文字标记、侧身行走的模特、大胆的色块、小色块色板、花卉装饰和清新的街头服饰氛围。
    
    2. 花卉动态排版
       使用具有优雅肢体动作的时尚模特或舞者、超大字母形状、织物运动、花卉 / 橙色装饰元素和精致的编辑排版。
    
    3. 字体遮罩肖像
       使用黑白肖像切割巨大的垂直排版。让面部出现在字母形状内部,就像时尚电影海报或设计杂志封面。
    
    4. 几何空间时尚海报
       将模特放置在一个干净的白色几何空间、立方体、门口或建筑结构中。使用极简排版、柔和阴影和设计师品牌造型。
    
    保持结果高端、图形化、干净、时尚且具有视觉设计感。
    
    避免廉价模板设计、杂乱的贴纸、随意的装饰、低端海报风格、弱排版、扭曲的身体、糟糕的手部、假塑料皮肤、杂乱的布局、糟糕的文本渲染和真实的版权品牌标志。
  • OpenAI砍掉Pulse每日摘要,ChatGPT的「早报」功能为啥只活了9个月

    OpenAI砍掉Pulse每日摘要,ChatGPT的「早报」功能为啥只活了9个月

    6月18日,OpenAI在宣布升级ChatGPT”定时任务”功能的同一条推文里,顺带说了一件事:Pulse要关了,14天之内正式下线。Pulse这个东西,如果你没用过,简单说就是ChatGPT给你的”每日早报”——它根据你的兴趣,每天自动生成一份新闻摘要,推送到你面前。

    OpenAI砍掉Pulse功能
    OpenAI宣布砍掉Pulse每日摘要功能,建议用户改用定时任务

    Pulse是2025年9月才上线的,算下来只活了大约9个月。在AI产品的生命里,这不算特别短——但也不长。OpenAI给的替代方案是用”定时任务”来自定义每日简报。换句话说,以前Pulse是主动推给你,现在是你要自己设一个定时任务,让ChatGPT每天定时帮你生成一份。

    为什么要砍掉Pulse

    OpenAI没细说原因,但可以猜一猜。一种可能是用的人不够多——主动推送的每日摘要,听起来很酷,但实际用起来,很多人可能看了两天就腻了。另一种可能是运营成本。Pulse需要每次都调用模型生成内容,还要做个性化,这都是真金白银的算力。如果留存率不够好,砍掉也合理。

    换成”定时任务”来代替,其实更符合OpenAI现在的产品逻辑。他们想让ChatGPT变成一个”什么都能做”的平台,而不是一个只能被动接收推送的App。你设一个定时任务,每天自动跑,效果和Pulse差不多,但更灵活,还能做更多定制。

    Pulse的死亡,某种程度上反映了AI产品的一个普遍困境:概念很性感,但实际使用频率很难做高。每日摘要这种东西,用两次觉得新鲜,用一个月后还能坚持的人,不多。

    这件事还有一个值得玩味的地方。OpenAI在推文里说Pulse会在”接下来的14天内”下线,但没有给出一个精确的最后日期。这种模糊的告别方式,在某种程度上也挺符合OpenAI一贯的作风——先宣布,再慢慢退,给用户一点时间适应。

    不管怎样,如果你还在用Pulse,抓紧时间,还有不到两周。之后要想在ChatGPT里看每日摘要,就得自己动手设定时任务了。


  • 诺贝尔化学奖得主跳槽Anthropic,Google DeepMind失去一位核心科学家

    诺贝尔化学奖得主跳槽Anthropic,Google DeepMind失去一位核心科学家

    6月19日,量子化学和AI蛋白预测领域最耀眼的科学家之一John Jumper在X上发了一条帖子,宣布离开Google DeepMind,加入Anthropic。这条消息在AI圈子里炸了锅——一个诺贝尔奖得主,从一方跳到另一方的直接竞争对手,这种事并不常见。

    John Jumper加入Anthropic
    诺贝尔化学奖得主John Jumper从Google DeepMind跳槽至Anthropic

    如果你对AlphaFold没概念,这么说吧——2024年的诺贝尔化学奖,一半颁给了Jumper和DeepMind CEO Demis Hassabis,理由就是他们用AI预测蛋白质三维结构的工作”从根本上改变了化学”。在这之前,科学家要搞清楚一个蛋白质的精确结构,得花几个月甚至几年做X射线晶体衍射。AlphaFold出来之后,输入氨基酸序列,几分钟就能拿到可信的结构模型。

    一个诺贝尔奖得主,为什么要换东家

    Jumper的离开,表面上看是一个科学家换了份工作,但在这个时间点,这件事的意义要复杂得多。首先,Anthropic和Google DeepMind是直接竞争对手。Anthropic出的Claude,和Google的Gemini在同一个赛道上竞争。一个刚拿了诺贝尔奖的顶级研究者,从一方跳到另一方,这不只是在简历上多一行——这是在说,AI领域最顶尖的人才,也在重新选边站。

    有意思的是,Jumper的研究背景和Anthropic的方向其实很搭。AlphaFold本质上是用深度学习解决科学问题,而Anthropic一直在强调”可解释性”和”对齐”——让AI系统的推理过程可以被人类理解。Jumper在蛋白结构预测上积累的方法论,某种程度上就是在做一个”可解释的AI”:你看得见模型预测的结构,你可以拿去和实验数据比对。

    有人说,这次人才流动背后,反映的是AI研究重心的一次转移——从”能力竞赛”到”可靠性竞赛”。Jumper的选择,也许是一个信号。

    当然,Google DeepMind失去Jumper,不等于AlphaFold项目会停摆。DeepMind里还有一大批顶尖研究者。但一个诺贝尔奖得主公开跳槽到竞争对手那里,这件事本身,就已经说明了很多问题。AI人才战争的烈度,可能比我们想象的还要高。


  • Hugging Face Transformers – 161K+ Stars,现代AI开发的基石框架,统一百万预训练模型调用标准

    Hugging Face Transformers – 161K+ Stars,现代AI开发的基石框架,统一百万预训练模型调用标准

    Hugging Face Transformers

    GitHub OpenGraph Preview

    🤗 项目简介

    Hugging Face Transformers 是现代AI开发的基石框架,为文本、视觉、音频和多模态模型提供统一的模型定义标准,支持推理与训练全流程。只需掌握3个核心类,即可调用Hugging Face Hub上超过100万个预训练模型检查点。

    161K+
    GitHub Stars
    1M+
    预训练模型
    500+
    模型架构
    3
    核心类

    ⚙️ 安装要求与过程

    环境要求:

    • Python 3.10+
    • PyTorch 2.4+(或 JAX/TensorFlow 2.0+)
    • CUDA(可选,用于GPU加速)

    快速安装:

    # 使用 pip 安装(推荐)
    pip install "transformers[torch]"
    
    # 使用 uv 安装(更快)
    uv pip install "transformers[torch]"
    
    # 从源码安装最新版
    git clone https://github.com/huggingface/transformers.git
    cd transformers
    pip install '.[torch]'

    验证安装:

    import transformers
    print(transformers.__version__)  # 输出版本号即成功

    🚀 核心功能

    1. 统一的模型定义框架 🏗️

    Transformers 提供了统一的模型定义标准,只要模型被支持,就能兼容绝大多数训练框架(Axolotl、Unsloth、DeepSpeed等)、推理引擎(vLLM、SGLang、TGI等)以及相邻建模库(llama.cpp、mlx等)。

    2. 极简API设计 🎯

    只需学习 AutoModelAutoTokenizerPipeline 三个核心类,即可使用Hugging Face Hub上超过100万个预训练模型,覆盖NLP、CV、音频、多模态等任务。

    3. 跨框架无缝切换 🔄

    支持在PyTorch、JAX、TensorFlow 2.0之间随意切换模型,可针对不同阶段(训练、评估、生产)选择最合适的框架,3行代码即可训练前沿模型。

    4. 全模态Pipeline支持 🎨

    提供高层推理API Pipeline,支持文本生成、图像分类、音频识别、视频理解等多模态任务,自动处理输入预处理并返回对应输出。

    5. 高度可定制与可复现 🔬

    提供每个架构的官方结果复现示例,模型内部结构尽可能统一暴露,模型文件可脱离库独立使用,方便快速实验和研究。

    💡 典型使用场景

    场景一:快速原型开发 🏃

    使用Pipeline API,3行代码即可完成文本分类、命名实体识别、文本生成、图像分类等任务。无需关心模型架构细节,直接调用SOTA模型进行推理。

    from transformers import pipeline
    
    # 文本生成
    generator = pipeline("text-generation", model="gpt2")
    result = generator("AI will", max_length=50)
    
    # 图像分类
    classifier = pipeline("image-classification", model="google/vit-base-patch16-224")
    result = classifier("cat.jpg")

    场景二:模型微调与训练 🎓

    基于预训练模型进行下游任务微调,利用Hugging Face Trainer API或Accelerate库,轻松实现分布式训练和混合精度训练,大幅降低计算成本。

    from transformers import AutoModelForSequenceClassification, Trainer, TrainingArguments
    
    model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
    training_args = TrainingArguments(output_dir="./results", num_train_epochs=3)
    trainer = Trainer(model=model, args=training_args, train_dataset=dataset)
    trainer.train()

    🌟 推荐理由

    Hugging Face Transformers 是现代AI开发的”基础设施”,几乎每个AI开发者都或多或少用过它。它不仅提供了统一的模型定义标准,更重要的是构建了一个庞大的生态系统——Hugging Face Hub上超过100万个预训练模型,覆盖了从BERT到GPT、从ViT到CLIP、从Whisper到Llama的所有主流模型。

    作为一个AI开发者,掌握Transformers库是必修课。它的API设计非常优雅,只需3个核心类就能玩转绝大多数SOTA模型。而且,它对PyTorch/JAX/TF的跨框架支持,让你可以在不同场景下灵活选择最合适的后端。无论你是做研究还是做应用,Transformers都是不可或缺的工具。

    📄 许可证:Apache License 2.0(允许商业使用)
    💻 主要语言:Python
    🏢 开发方:Hugging Face Team
  • 机器人训练数据成了AI新战场,这家公司刚拿了7000万美元

    机器人训练数据成了AI新战场,这家公司刚拿了7000万美元

    做语言模型的时候,你可以把整个互联网的文本喂给AI;但教机器人学会在真实世界里动手,数据就没那么好找了。这正是XDOF(念作”eks-doff”)看到的空隙——6月17日,这家公司走出隐身模式,宣布拿到Thrive Capital、a16z等机构7000万美元融资。

    机器人训练数据概念图
    机器人训练需要大量真实物理交互数据

    创始人的亲身踩坑经历

    联合创始人Philipp Wu在UC Berkeley读博的时候就被这个问题卡住了——想训练机器人基础模型,手里却没有足够的高质量数据。他和后来成为CTO的Fred Shentu做了个叫GELLO的遥操作系统,让人类操作员可以控制机械臂生成训练数据。

    这个项目后来成了机器人圈的高引论文——因为很多实验室都面临同样的瓶颈,大家开始用GELLO这类设备来收集数据。Wu意识到,这背后是个还没人认真做的生意。

    “我们很早就看到语言模型竞赛里落后的后果……你不会想在这个技术领域也追太晚,物理AI是下一个前沿。”Wu对TechCrunch说。

    三层数据金字塔

    XDOF的业务分三个层级来搭建数据管道:最顶层是针对具体部署机器人采集的遥操作数据;中间层是用GELLO这类设备收集更通用的机器人操作数据;底层则是”自我中心视角”数据——让人类戴着穿戴式传感器完成日常任务,记录第一人称视角的动作。

    公司计划在全球招募和培训大批遥操作员和数据采集员。Wu说,这件事需要几十万平方英尺的仓库、成百上千台机器人,还要持续校准和维护设备——大部分AI实验室宁愿外包,也不想自己建这套东西。

    和学术界一起开放数据

    XDOF和UC Berkeley的AI实验室合作,发布了目前规模最大的机器人训练数据集之一,叫ABC——包含13万条机械臂操作轨迹、300小时仿真数据和100小时评估数据。这个发布让学术界第一次能拿到这么大量的预处理好的机器人数据。

    OpenAI在2021年关掉了它的机器人团队,但两周前宣布要重启这个方向。所有前沿AI实验室都在往物理世界发力,而XDOF赌的就是:谁先搞定数据管道,谁就占住了下一个赛道的入口。