第一款真正实时的 AI 绘画The first genuinely real-time drawing AI

随手涂,
实时成图。
Scribble now,
see it render.

提示词再长,也说不清你脑子里那张画;图生图再准,也要等上几十秒才开始。人在画的时候需要的是笔尖一落,画面就跟着变——VibeDraw 把这条回路接上了:你随手涂几笔,成图层立刻变成一张完整的图,再涂几笔,它继续变。No prompt is long enough to describe the picture in your head, and image-to-image still makes you wait before it starts. Drawing is a loop: the stroke goes down and the picture answers. VibeDraw closes that loop - scribble a few strokes and the result layer becomes a finished image, add more strokes and it keeps up.

VibeDraw · 0.4.26 · 由 Hermit 安装VibeDraw · 0.4.26 · installed by Hermit

需要 Android 10 及以上手机,并先安装 Hermit。模型可以由你自己的电脑或云端服务提供。Needs Android 10 or later plus Hermit. Models can come from your own computer or a cloud service.

为什么现有方式都不够Why the existing options fall short

描绘画面的两种办法,都少了“手”。Both ways of describing a picture leave out your hand.

提示词是一种翻译A prompt is a translation

你先在脑子里想好画面,再把它翻译成词,然后祈祷模型和你用同一本词典。构图、留白、某一块颜色应该在哪,写成句子十有八九会走形。You compose the image in your head, translate it into words, then hope the model shares your dictionary. Composition, negative space and where one colour belongs rarely survive the sentence.

图生图是一次会议Image-to-image is one meeting

你画完、提交、等待、看结果,不对就回去改一点再提交。每次迭代都要几十秒,思路早就断了。它适合定稿,不适合想东西。Draw, submit, wait, look, adjust, submit again. Tens of seconds per loop is long enough to lose the idea. It is a way to finish an image, not to think one through.

涂鸦才是人的方式Doodling is how people work

人的创造力是信手拈来的:一条歪线、一个圈、随手补的一块暗部。VibeDraw 让你就一直这样画下去,画面在旁边实时跟上。Human creativity is offhand: a crooked line, a circle, a patch of shadow added without thinking. VibeDraw lets you keep drawing like that while the picture keeps up beside you.

实时是认真的Real-time, literally

几百毫秒一条笔触,
不是几十秒一张图。
A stroke per several hundred milliseconds, not a picture per minute.

实时不是把大模型加速,而是把它拆成两层:一个够快的小模型专门跟着你的笔走,一个高质量模型负责把满意的画面渲染成高清图。Real-time does not mean speeding up a big model. It means splitting the job in two: a fast model that follows your pencil, and a high-quality model that renders the frame you are happy with.

  1. 实时槽位固定 512×512、2/4/8 步,用 DreamShaper8 LCM,CFG 固定为 2。The live slot stays at 512×512 with 2, 4 or 8 steps on DreamShaper8 LCM and a fixed CFG of 2.
  2. 服务端先把画布做轻微模糊副本再与原图混合,你的笔迹因此不会变成刺噪点;滑块“绘制稿强度”决定模型多忠于原稿。The server blends a lightly blurred copy with your canvas so strokes do not turn into noise, and the strength slider decides how faithful the result stays to your drawing.
  3. 画布原样提交:颜色、透明度、构图都不改写。你看到的就是模型看到的。The canvas is submitted as it looks: no rewriting of colour, opacity or composition. What you see is what the model sees.
  4. 想重复一版就锁定随机种子;想保留上一版结果就打开“叠加生成”,它会作为历史快照的一部分保存。Lock the seed to repeat a result, or switch on composite generation to keep the previous frame; both are saved with the history snapshot.

成图层和元素层的透明度各自可调,画的时候可以一直看得见自己的笔。生成过程中你还能继续画,快速操作会自动排队合并。Result and element layers fade independently, so your own strokes stay visible while you work. You can keep drawing during a generation; rapid actions queue and coalesce.

两层模型,各干各的Two slots, two jobs

实时Live
DreamShaper8 LCM

512×512 · 2/4/8 步 · CFG 2。跟着笔走的那个,几乎任何独立显卡都能跑。512×512, 2/4/8 steps, CFG 2. The one that follows your pencil, comfortably within reach of almost any discrete GPU.

渲染Render
Flux.2 Klein 4B

distilled 版 2/4 步,8 步换 base。原生 1024×1024,四步就能出可用结果。Distilled for 2 or 4 steps, base at 8. Native 1024×1024 and usable output in four steps.

高保真Fidelity
QwenImage 2512

牺牲一些速度换细节与构图服从度,适合显存充足的台式机。Trades some speed for detail and instruction following; best on a desktop with headroom.

两个槽位完全独立:给实时槽位点一下 A1X 快速配置,只会把实时槽位设成上面的组合,不会改动你的渲染槽位。The two slots are independent: the one-tap quick preset only configures the live slot and never rewrites your render slot.

想怎么接,就怎么接Connect it your way

本地、云端、还是自己的协议。Local, cloud, or your own protocol.

VibeDraw 不自带 API Key,也不绑定任何平台。它按你的选择提交请求,模型和工作流仍然在你手上。VibeDraw ships no API key and locks you to no platform. It submits requests to whatever you configure, and the models and workflows stay yours.

本地 ComfyUILocal ComfyUI

装上 VibeDraw 插件后,在任意工作流里放一个 VibeDraw Input 和一个 VibeDraw Output:把提示词、参考图、蒙版、种子、强度、步数和尺寸接到你自己的采样或重绘节点上,最终图片接回输出节点。VibeDraw 把 API 格式的工作流和当前输入提交给插件,再直接从 ComfyUI 读取结果图。With the VibeDraw plugin installed, drop one VibeDraw Input and one VibeDraw Output into any workflow: wire prompt, reference, mask, seed, strength, steps and size into your own sampler or inpainting nodes, and route the final image back. VibeDraw submits the API-format workflow with the current inputs and reads the result straight from ComfyUI.

fal.ai

不想在本地跑模型,就把实时和渲染都交给云端。手机只负责画和看,电脑可以关着。If you would rather not run models locally, hand both slots to the cloud. The phone only draws and displays; your computer can stay off.

稳定任务接口A stable job API

也支持 A1X 风格的任务接口:提交任务、轮询状态、下载输出,客户端会校验返回图片的真实尺寸。An A1X-style job API is supported too: submit, poll, download, and the client verifies the real size of what comes back.

通用兼容层Compatible servers

OpenAI Images 兼容接口、Stable Diffusion WebUI / Forge、Stability AI v2beta 都能直接接。可信局域网允许明文 HTTP,公网服务必须 HTTPS。OpenAI Images compatible endpoints, Stable Diffusion WebUI and Forge, and Stability AI v2beta all connect directly. Plain HTTP is allowed on a trusted LAN; public services must use HTTPS.

按你的台式机选模型Pick models for the desktop you have

8 GB

实时跑 DreamShaper8 LCM,渲染用 Flux.2 Klein 4B distilled 的 2 步。速度和清晰度都在线。Run DreamShaper8 LCM live and Flux.2 Klein 4B distilled at 2 steps to render. Speed and clarity both hold up.

12 GB

实时同上,渲染可以上到 4 步,需要更细的细节时切 8 步的 base 版本。Same live slot, and rendering can go to 4 steps, switching to the base model at 8 steps when detail matters.

16 GB+

把 QwenImage 2512 接进渲染槽位,拿原生 1024×1024 的构图服从度和质感。Put QwenImage 2512 on the render slot for native 1024×1024 composition accuracy and texture.

VibeDraw 会检查服务返回图片的实际尺寸,512 的结果不会被放大冒充高清。你也可以先只连实时槽位,等画满意了再配渲染。VibeDraw checks the real dimensions of returned images, so a 512 result is never upscaled and passed off as high quality. You can also start with only the live slot and add rendering later.

画布上你能做的On the canvas

它首先是个能画画的工具。It has to be a drawing tool first.

AI 只是画布的一部分。笔、选区、图层、蒙版、撤销这些东西不顺手,实时也救不了。AI is only part of the canvas. If the pen, selection, layers, masks and undo are clumsy, real-time cannot save it.

笔、涂色与擦除Pen, fill and erase

三种工具直接显示粗细和透明度滑杆,参数栏右端一个圆形按钮切换背景。所有滑杆共用同一套大圆形触点,手指够得到。All three tools show width and opacity sliders right there, with a round button on the right of the bar to set the background. Every slider shares one large round handle that a thumb can actually hit.

对象与图层Objects and layers

碰触式框选,成组、解组、单选或整组移动、复制、上下调层。选中后四角出现缩放手柄,也能双指捏合缩放和移动。Tap to marquee-select, then group, ungroup, move a single object or the whole group, duplicate, or reorder layers. Handles appear on the corners and two fingers can pinch to zoom and pan.

局部蒙版Local masks

只想改一块地方,就只把那块交给模型。不用为了让一只手变得更好看而重画整张图。Hand only that region to the model. You should not have to redraw the whole picture to fix one hand.

历史与副本History and copies

作品自动保存并生成缩略图,可以搜索、继续编辑、创建副本或删除。恢复时会带回提示词、笔触、导入的图片、生成结果和当时的画面参数。Works save themselves with thumbnails, and can be searched, reopened, duplicated or deleted. Restoring brings back the prompt, strokes, imported images, generated results and the settings that were in force.

该出的图出得干净Exports that stay clean

工具栏的下载按钮保存画布的实际显示效果;渲染预览里的下载按钮单独保存模型返回的 1024 大图。服务以文件返回的图片按原格式导出。The toolbar download saves exactly what the canvas shows; the one inside the render preview saves the model's 1024 image on its own. Images returned as files export in their original format.

不卡手No stutter

绘制热路径用动画帧合并、离屏内容缓存和笔画边界缓存;历史缩略图按可视区域加载,图片解码与文件读写有并发上限和容量上限,画布没变就不会重复写入数据。The hot path coalesces on animation frames with offscreen content and stroke-bound caches. Thumbnails load by viewport, decoding and file writes are bounded, and an unchanged canvas is never written again.

深浅主题与中英双语Themes and two languages

浅色、深色、跟随系统;中文与英文。弹窗、确认、按钮和提示共用一个组件体系,风格不会各自为政。Light, dark or follow the system, in Chinese or English. Dialogs, confirmations, buttons and toasts share one component system, so nothing looks bolted on.

离线可用,不靠 CDNOffline, no CDN

界面使用 Hermit 内置的本地图标字体,不访问外网 CDN;源文件无需构建即可运行,画布和配置保存在应用自己的数据区,图片只存文件引用。The interface uses Hermit's bundled icon font and never reaches an outside CDN. Sources run without a build step, the canvas and settings live in the happ's own data partition, and images are stored as file references.

不需要另找破解No need to beg for a feature

不顺手?直接改成你的。If it does not fit, change it.

VibeDraw 是开源的 happ,装在你的 Hermit 里。在 Hermit 的开发页开启开发模式,把地址和密码交给你信任的智能体,就可以让它改画笔参数、换配色、加一个你天天要用的快捷功能。VibeDraw is an open-source happ running inside your own Hermit. Turn on development mode, hand the address and password to an agent you trust, and it can retune the brushes, restyle the interface, or add the shortcut you reach for every day.

把 VibeDraw 装进手机Install VibeDraw

先用 Hermit 扫码,确认后就能打开。Scan with Hermit, confirm, and open it.

VibeDraw 安装二维码:在 Hermit 中点扫码添加扫描;也可复制下方地址。
  1. 打开 Hermit,点“扫码添加”Open Hermit and choose Scan to add.
  2. 扫描这个二维码,确认安装。Scan this code and confirm the install.
  3. 点开 VibeDraw 就能画,接好模型后再开生成。Open VibeDraw and draw right away, then add a model to start generating.

不接模型也能画:笔、图层、蒙版和导出都能单独用。要生成时才需要在设置里填一个服务地址与密钥。It draws without a model: pen, layers, masks and export all work on their own. Generating is what needs a service address and key in settings.

同一位作者的上一个作品The previous one from the same workshop

画完了,找几个 AI 评一评。Done drawing? Let a room of AIs argue about it.

chataxi 是同一个 Hermit 里的多角色 AI 群聊:每个角色有自己的模型、人设和声音,可以让它们一起讨论你的设定、命名和文案。chataxi is the multi-role AI group chat for the same Hermit: each role has its own model, persona and voice, ready to argue over your settings, names and copy.

chataxi 应用图标

一定要有独立显卡吗?Do I need a discrete GPU?

本地跑才有这个要求,而且实时槽位用的是 512×512 的小模型,门槛不高。不想装环境就用云端服务,手机照样画。Only for local models, and the live slot runs a small 512×512 model, so the bar is low. If you would rather not set anything up, a cloud service leaves the phone doing all the drawing.

出的图不是我想要的怎么办?What if the result is not what I meant?

先调“绘制稿强度”:调高更忠于你的线稿,调低更多交给模型发挥。再锁定种子复现某一版,或者用局部蒙版只改一块。颜色如果是靠细线表达的,模型很难判断那是物体表面色还是标注线,这时候把颜色写进提示词更稳。Start with the strength slider: higher stays truer to your drawing, lower gives the model more room. Then lock a seed to reproduce a frame, or mask a region to change only that part. When colour is carried by thin lines the model cannot tell a surface from an annotation, so naming the colour in the prompt is more reliable.

和 Hermit 是什么关系?How does Hermit fit in?

VibeDraw 是一个 happ,需要装在 Hermit 里运行。Hermit 负责安装、数据、文件、权限和系统能力,VibeDraw 负责画布与生成。常见问题在这里VibeDraw is a happ and runs inside Hermit. Hermit owns installation, data, files, permissions and system capabilities; VibeDraw owns the canvas and generation. More answers in the FAQ.

你的手比提示词更会说话。Your hand describes better than a prompt.

装好 Hermit,扫码把 VibeDraw 加进来,随手涂两笔试试。Install Hermit, scan, add VibeDraw, and scribble something.