Skip to content

v1.63.0 (AI Ways and Skills, Saves, Official Example Workflows, Privacy Mode)

Choose a tag to compare

@DemonGatanjieu DemonGatanjieu released this 10 Nov 07:42
· 10 commits to main since this release
2cc8dad
  • 画布左上角的立方体入口不再压在 ComfyUI 自己的菜单上面:以前它在最顶层,打开左上角的 C 菜单时会盖住“新建”那一行。现在它在画布上面、顶栏和菜单、对话框下面。README 的安装第 2 步也配上了图,写明装好后入口在哪、看不到时怎么办。

  • The cube entry at the canvas's top left no longer sits over ComfyUI's own menus: it was on the very top layer and covered the "New" row of the C menu. It is now over the canvas and under the top bar, menus and dialogs. Step 2 of the README's install has a picture of where the entry is after installing and what to do when it can't be seen.

  • 在设置里切换语言后,AI 工作台右边的展台、留言框和画布下方的存档条也马上换成新的语言。以前它们要刷新页面才变。

  • After switching the language in Settings, the AI workbench's stage and message box and the saves bar under the canvas change at once too. They used to wait for a page reload.

  • 浏览器里同时开着两个 ComfyUI 页面时,AI 干活不再时好时坏:以前两个页面会各抢一半动作,一个页面开了任务的标签页,另一个页面接着去跑却找不到它("标签页关了")。现在任务的动作先交给上次干活的那个页面,它关掉了别的页面才接手。

  • With ComfyUI open in two browser tabs the AI no longer fails now and then: the two pages took half the actions each, one opened the task's tab and the other then could not find it ("the tab is closed"). Now the page that did the last action gets the next first, and another takes over only when it is gone.

  • AI 用 run_workflow 时可以直接写图库里一张图的名字:一步打开这张图的工作流、改好提示词或种子、跑出来;同一张图再跑就接着用那个标签页。打开图的工作流时换一个新的工作流编号,原来那个标签页还开着也不会串;打开后的标记在加载之前就写进去,不会被 ComfyUI 收尾时冲掉(以前偶尔会说"标签页关了"或"没打开")。

  • With run_workflow the AI can name a picture of the gallery: its workflow opens, gets the new prompt or seed and runs, in one call, and a run of the same picture again reuses that tab. A picture's workflow opens with an id of its own, so the tab it ran on, still open, is not mixed up with it, and the tab's mark goes in before it loads, so ComfyUI finishing the tab does not drop it (it said now and then that the tab was closed or did not open).

  • 偶尔打开 ComfyUI 时 AMB 整个没出来(刷新一下又好了)的问题修好了。ComfyUI 会同时加载插件的全部 js 文件,碰巧某个文件先跑时,会用到另一个还没准备好的变量,整个插件就加载失败;重启或更新后第一次打开最容易碰上。新加的测试会把每个文件都当成第一个加载试一遍。

  • AMB is no longer missing now and then when ComfyUI opens (a reload brought it back). ComfyUI loads every js file of the plugin at once, and when one file happened to run first it used a value of another file not made yet, and the whole plugin failed; the first open after a restart or update met it most. A new test runs each file first.

  • AI 看得到工具的回答了:AI 软件收到太长的回答(Antigravity 大约 4,000 字节以上)会存成文件不看,AI 就什么也没拿到,只好缩小范围重查(考试里为挑一个 LoRA 搜了 8 次)。现在搜模型的回答只留有用的(同类模型的加载节点只写一次,触发词列前四个,LoRA 带上作者示例图的权重),读大画布时后面的节点写成一行,其他列表放不下的列出名字,告诉 AI 怎么查剩下的。

  • AI apps get to read the tools' answers: an app files an answer past its size (about 4,000 bytes in Antigravity) away unread, so the AI had nothing and searched again with smaller limits (eight searches for one LoRA in the exam). Model search now keeps what is used (a type's loaders said once, four trigger words, a LoRA's weight from its author's pictures), a big canvas gives its last nodes a line each, and any other list names what did not fit and says how to get it.

  • 画布是空的时候,AI 跑图或读画布会被告知先打开一个工作流(存的、示例,或者图库里某张图自带的),不再让它去"加一个保存节点"。AI 查一张图的参数时,也会知道这张图的工作流能直接打开。

  • On an empty canvas the AI is told to open a workflow first (a saved one, an example, or a gallery picture's own) instead of being told to add a save node. Reading a picture's settings, the AI learns that its workflow can be opened as it was.

  • AI 发来只写着"test"这类试探的问题,不再放到你面前等回答,AI 会收到正确的问法。任务上的"出图 N 张"只数真正跑出来的图,AI 找到给你看的旧图不算。

  • A question that only says "test" or the like no longer waits in front of you: the AI is told how to ask. A task's picture count counts the pictures its runs made, not old ones the AI found to show you.

  • AI 想改图库里的某一张图时,可以直接打开那张图自带的工作流(open_workflow 写图片名),不用再打开原来的工作流卡,把换过的模型、LoRA 和提示词一样一样设回去。打开的是新标签页,你开着的那页不动。Gemini 实测,这类题从 28 次调用降到 21 次。

  • When the AI changes a picture from the gallery, it can open the workflow that picture carries (open_workflow with the picture's name), instead of opening the saved card and putting back the model, LoRA and prompt that had been changed one by one. It opens in a new tab and leaves yours as it was. In a real Gemini run such a task went from 28 calls to 21.

  • 在设置里切换语言后,AI 工作台的任务列表和时间线马上跟着换。以前有一部分要刷新页面才变。

  • Switching the language in Settings now changes the AI workbench's task list and timeline at once. Part of them used to wait for a page reload.

  • 任务上的“出图 N 张”不再把同一张图算两次。以前 AI 汇报时附上刚跑出的那张图,会被多数一次。

  • A task's picture count no longer counts one picture twice. When the AI's report showed the picture its run had made, it was counted again.

  • AI 看工作流清楚多了(Gemini 实测发现的):列表以前一次两万多字,AI 软件直接存成文件不看了,现在卡片多时只给名字、路线和备注。每张卡的路线说明以前一直是空的,现在有了;按模型找卡片也恢复了。官方示例总会列出名字,用中文(图生图、两段放大)也搜得到。

  • AI apps see your workflows better (found in a real Gemini run): the list was over 20,000 characters, which an AI app files away unread, and is now names, routes and notes when there are many. Each card's route line was always empty and now says the route, and finding cards by model works again. Official examples are always named, and found by their Chinese names too.

  • 新任务里先写了留言的话,任务列表就用这句话当标题,不再都叫“接手任务”。

  • A new task you wrote to takes its name from what you wrote, instead of the AI's "take the task".

  • 两个 Anima 两段放大示例写明了:要多张图就多跑几次,别调批次。Anima 的 VAE 会把一批图当成视频帧,第二段只剩一张。

  • The two Anima two-pass examples now say to run several times for several pictures, not to raise the batch: Anima's VAE takes a batch for video frames, and the second pass is left with one.

  • 当前节点的提示词卡片现在都标着“正面提示词”或“负面提示词”:从图片存下来的也按它原来接在采样器哪个口上来认,以前只有工作流来源的才标,图片和工作流同时有时标注还会丢。

  • Current node's prompt cards all say positive or negative now: values kept from a picture are told by what they were wired to in that workflow. Before, only a workflow's said so, and the label was lost when a picture held the same text.

  • 回到一个存档时,输出节点上的图也跟着换:画布是哪个存档,节点上就是哪个存档跑出来的图。跑了几次就都排在节点里,和批量出图一样,新的在前,最多 16 张。这个存档没有图时节点空着,不留别的存档的图,免得看错。只对这个版本之后跑的图有效。

  • 回到存档时数字有时不亮、撤销多出一步的问题修好了:画布上有 Preview Image 这类节点时,每隔一次就会这样。

  • Going back to a save brings its pictures back too: the output nodes show what that save made. Several runs sit on the node together as a batch does, the newest first, up to 16. A save with no picture leaves the node empty rather than showing another save's. This works for runs from this version on.

  • Going back to a save no longer loses the lit number or leaves an empty undo step every other time, as it did with a node like Preview Image on the canvas.

  • 发版前把每个页面按中英文、深浅色和 480 窄宽走了一遍,修了一批小问题。浅色主题下,模型整理页的标题和文件名、任务上“接着做”的标签原来几乎看不见。窄窗口下,任务的“接着哪个任务”标签伸出窗口、× 点不到,设置页的主题按钮把页面撑出横向滚动,模型页的搜索框压住标题,工作流详情的“新预设”被切掉。

  • 英文界面的列表和冒号不再是中文标点,“1 saves”这类单复数对了,左边栏的 AI 工作台写作 “AI”,一条漏翻的提示补上了。深色主题下浅紫色按钮和角标上的字改成深色,看得更清楚。

  • A round through every page in both languages, all themes and at 480 fixed small things. In the light theme, the Tidy page's headings and file names and a task's "goes on in" chip could hardly be read. In a narrow window, a task's "goes on from" chip ran out of the window with its × out of reach, the theme buttons in Settings pushed the page sideways, the Models search sat over the title, and "New preset" was cut off.

  • The English window no longer shows Chinese punctuation in lists and after labels, "1 saves" and the like have the right plural, the rail says "AI" for the AI workbench, and one line that was left in Chinese is translated. In the dark themes, text on pale accent buttons and badges is dark now, so it reads better.

  • 侧栏下面的工具按常用程度排了:当前节点在最上面,然后是模型检查,模型扫描在最下面。

  • The rail's tools go by how often they are used: Current node first, then Model Check, with the scan last.

  • 当前节点的参数卡清爽了:来源不再一条一行,改成按工作流、搭配、图片、这里存的分开的小标签,同一渠道几条合成一个,点标签直接过去,几条的话先列出来。

  • 当前节点现在也列出搭配里同类节点的数值,可以对比和套用;普通搭配给提示词框,节点结构搭配给每个框。

  • Current node's cards are easier to read: where a set came from is a chip per channel (workflow, combo, picture, saved here) instead of a line per source, a press goes there, and several in a channel are listed first.

  • Current node now lists the values combos hold for that node type, to compare and put in: a plain combo's prompt on a prompt box, every box of a node structure.

  • 给 AI 挑模型时和模型页一样好找了:类型用平常的叫法,有基础模型的分类条,版本和名字分开写,同一模型的几个版本合成一张卡,点版本就能换。

  • Picking models for the AI works like the Models page: types by their usual names, base model chips, the version apart from the name, and a model's versions on one card, changed from its version.

  • 工作流页多了“官方示例”:文生图、两段放大、图生图、局部重绘,打开时自动选一个你已有的模型。两段放大有 SDXL / SD 1.5、Anima 和推荐的 Anima · Tile 锁图三个版本,缺节点包或模型的版本会写明缺什么。新装的人和 AI 都有现成路线可用;AI 没有合适的已存工作流时会用它们。

  • The Workflows page has "Official examples": text to picture, two passes, picture to picture and repainting an area, each opening on a model you already have. Two passes comes for SDXL / SD 1.5, for Anima, and as the recommended Anima held by Tile; a version says what pack or model it lacks. New installs and the AI have routes to start from; the AI gets them when no saved workflow fits.

  • 可以自己选做法了:在“做法”里点“这次用它”,只对当前任务生效。AI 会收到,这一类活 AMB 直接按你选的搭,AI 想用别的也不行。

  • You can pick the way yourself: "Use it this time" in Ways, for the current task only. The AI is told, and Anomalous builds your pick for that kind of job, whichever one the AI asks for.

  • 点“做法”马上出来,不再每次重画:打开 AI 工作台时就提前读好,再点时停在你上次看的位置,展开的详情也还在;开关一个做法也不会跳回顶部。

  • Ways opens at once and no longer draws itself twice: it is read when the AI workbench opens, and comes back where you left it, details still open; turning a way on or off no longer jumps to the top.

  • 新任务的页面清爽了:只留一句发给 AI 的话和一个复制按钮。要先交代的话和图都放下面的留言框,不再有重复的两行。连过 AI 以后也不再看到“第一次用?”;从没连过时,先告诉你怎么连。

  • The new task page is plain now: one sentence for the AI and a Copy button. What to say first and pictures go in the message box below, with no second place for them. Once an AI has connected, "First time?" is gone; until then, how to connect comes first.

  • AI 工作台的任务列表:一条线上的任务排在一起,后续任务缩进挂在最早的任务下面,一眼看出谁接着谁。每个任务行上多了 ⋯,改名、隐私模式、从列表移除不用先打开任务。移除中间的任务时,线不会断。

  • The AI workbench's task list keeps a line of tasks together: follow-ups hang under the first task, so you see what goes on from what. Each task row has a ⋯ for rename, privacy mode and remove, without opening the task first. Removing a task in the middle does not break the line.

  • 存档条和存档页上的按钮按一下就管用:刚在提示词框里打完字、刚改完次数,或者运行还在跑的时候,第一下不再被吃掉;运行途中正在输入的次数也不会被清掉。改完次数马上点运行,也一定按新的次数跑。次数没存上时会提示,框里退回实际的次数。按住存档页的图时刚好出了新图,这一下也算数,新图等松手再排进来。

  • The controls of the saves bar and the Saves page answer the first click: right after typing in a prompt box or changing a count, or while runs go, the click is no longer lost, and a count being typed stays. Run right after changing a count goes by the new count. A count that could not be saved says so, and the box goes back to the real one. A press on a picture on the Saves page counts even if a new picture arrives during it; the new one joins after you let go.

  • 插入 LoRA 可以一次选多个:选择窗口里点卡片加入,再点一下取消,卡片上的数字就是接上去的顺序,线自动连好,Ctrl+Z 一次全部退回。默认用 ComfyUI 自带的节点串成一串;装了 rgthree 或 Easy-Use 时可以改放进一个 Power Lora Loader 或 LoRA 堆叠节点,会记住上次的选择。

  • 在采样器前面插入 LoRA 时,如果前面已经有一个 LoRA,新的 LoRA 也能接上 CLIP 线了。

  • Insert LoRA takes several at a time: in the picker, click cards to add them and again to drop one; the number on a card is the order they go in. They are wired for you, and one Ctrl+Z takes them all out. They chain as ComfyUI's own nodes by default; with rgthree or Easy-Use installed they can go in one Power Lora Loader or LoRA stack, and the choice is remembered.

  • A LoRA inserted before a sampler now gets the CLIP line even when another LoRA already stands there.

  • 操作手感更顺了:打开和关闭窗口、切换页面都有短暂的过渡,左边栏的高亮会滑到新页面,切换侧边停靠和全屏时窗口平滑过去,不再每一帧重排页面。系统开了“减少动态效果”时这些都关掉。

  • Moving around feels smoother: the window and its pages come and go with a short transition, the rail's highlight slides to the new page, and going between docked, floating and full window glides there instead of laying the page out on every frame. All of it stays off when the system asks for reduced motion.

  • 侧边停靠的宽度可以拖了:按住窗口右边缘左右拖,最窄 480,最宽给画布留 240,宽度会记住,双击边缘恢复原来的 660。窗口变小时也会给画布留位置。各页面都按 480 宽走过一遍:模型页的关闭按钮不再被挤出去,工作流的搜索框放不下时换到第二行,图库拖宽后自动加列。

  • 窗口窄时提示词页不再只占上半截,词卡和提示词框上下排好。

  • The docked window's width can be dragged: hold its right edge and drag, from 480 up to leaving 240 of the canvas. The width is remembered, and a double-click on the edge returns to the usual 660. A smaller browser window leaves room for the canvas too. Every page was gone through at 480: the Models page keeps its close button, the Workflows search moves to a line of its own when there is no room, and the gallery adds columns as it gets wider.

  • In a narrow window the Prompts page no longer fills only its top half: the cards and the prompt boxes stack.

  • 工作流卡的“参数方案”改叫“预设”,详情页也重新排了:左边那一栏去掉,从上往下是 流程图、一排 预设标签、选中那个预设的内容。流程图只画结构,格子都很小:最下面一排按出图顺序画出每一段(出底图、接着画、放大重画),后面接 ColorMatch 这类后处理和保存;上面几排是底模、LoRA 和 Tile 锁图这类控制模型,每个横跨在它管的那几段上面,一眼看出哪个模型给哪几段用。这次不跑的一段画成虚线,写着“不跑”;没用上的模型也是虚线。步数、强度这些数字放在鼠标停上去的提示里;点一格,图下面列出这个节点的全部参数,改了的写成“默认 → 这个预设”,虚线的那格会告诉你哪个预设会用上它,点一下就切过去,× 收起。图下面的标签:默认参数、每个预设(写着和默认有几处不同)、“+ 新预设”(把当前画布存成新预设,或照着正在看的这个新建),最右边是全部出图和版本;选中的标签连着下面的内容,往下翻时这排标签一直停在顶上。选一个预设,它和默认参数不一样的格子在图上标成 橙色(只是不跑的那段换了模型不算改动)。“交给 AI”挪到了预设旁边,交的就是你正看着的那个预设。AI 也能用预设了:list_workflows 会列出每张卡的预设,写成“工作流名 / 预设名”;run_workflow 带上它,AMB 先打开工作流,再像点“应用到当前画布”一样套上那个预设,然后填题材、种子出图。AI 只要选哪个预设,改数值的事由 AMB 做。

  • A workflow card's parameter sets are now presets, and its details are laid out anew: no left column; from the top, the flow, a row of preset tabs, and the preset chosen. The flow is small and shows only the structure: a bottom row of the stages in the order they run (base picture, goes on, enlarge and redraw), then finishing nodes such as ColorMatch and the saves; above them a row each for the models, the LoRAs and control models such as a Tile LLLite, each a bar over the stages it feeds, so which model works on which stage shows at a glance. A stage these values do not run is dashed and says "not run"; an unused model is dashed too. Steps, strength and the like are in a box's tooltip; a click lists the node's values under the flow, changed ones as "default → this preset", and a dashed box names the presets that use it, one click away (× closes it). The tabs: the default values, each preset (with how many values it changes), "+ New preset" (save the canvas as one, or copy the one shown), and at the end all pictures and versions; the tab chosen joins the panel under it, and the row stays in view as you scroll. Pick a preset and what it changes from the default turns orange in the flow (a model swapped only in a stage that does not run is no change). "Hand it to the AI" now sits beside the preset and hands the one shown. AI apps can use presets too: list_workflows lists each card's as "workflow / preset"; with that name run_workflow opens the workflow, applies the preset as its Apply button does, then fills in the subject and seed. The AI only picks the preset; Anomalous changes the values.

  • 放大做法从 6 个合成 3 个:放大(1.25 倍轻轻重画)、Tile 锁图放大、分块放大(2 倍)。Tile 锁图放大合并了原来的平铺放大、Anima Tile 放大细化和双模型放大:放大那一段有自己的一份模型,标题是 “Upscale model”,默认和底图同一个模型、LoRA 一起带上(AMB 把画布上的模型和 LoRA 那一串复制一份);想换第二个模型(更稳、换皮肤质感)只要换这一格,LoRA 照样在。Anima 上走你验证过的那条路线:潜空间放大 1.22 倍,Tile LLLite 锁图,降噪 0.85。迭代放大去掉了(要 Impact Pack,和上面几个重叠)。局部重绘对 AI 只算一项,AMB 按你装的节点包挑裁切缝合、LanPaint 或官方节点;AI 也不再看到全部做法的名字,只按用途要。怎么用写在按需 skill 分阶段出图的思路 里(见下面)。

  • The upscale ways went from six to three: Upscale (1.25× with a light redraw), Upscale held by Tile and Tiled upscale (2×). Upscale held by Tile takes in the tile upscale, the Anima Tile detail upscale and the second-model upscale: the redraw has its own copy of the model, titled "Upscale model", the base's model with its LoRAs by default (Anomalous copies the canvas's model line); a second model (steadier, other skin or texture) is one swap of that loader, the LoRAs staying. On Anima it is your tested route: a 1.22× latent enlargement, a Tile LLLite holding it, denoise 0.85. The iterative upscale is gone (it needed Impact Pack and overlapped the others). Repaint counts as one for the AI, Anomalous picking crop-and-stitch, LanPaint or the built-in nodes by the packs you have; the AI no longer sees every way's name, only the uses. How to use them is in the on-demand skill Drawing a picture in stages (below).

  • 新增 半途换模型接着画 做法:AI 把画布上的采样器拆成两段,前面的步数用原模型定构图,后面换另一个模型画完(默认第 21 步换,共 35 步的 60%)。Anima 上直接用验证过的采样设置(dpmpp_2m_sde_gpu + beta);er_sde 拆成两段在两个 Anima 模型上都出了颗粒,AI 会避开(在一个 SDXL 模型上试过是干净的)。SDXL 上也实际跑通过(Nova Anime XL 换 MiaoMiao Harem v2.0)。换模型精修 也能用在 Anima 上了:Anima 只换模型文件,文本编码器和 VAE 沿用画布上的。两个都在 6 GB 显卡上实际出过图。

  • New way Switch models mid-way: the AI splits the canvas's sampler in two, the first steps on the canvas's model for the composition, the rest on another model (by default from step 21, 60% of 35). On Anima it uses the tested sampling (dpmpp_2m_sde_gpu, beta); split er_sde came out grainy on two Anima models, so the AI avoids it there (on one SDXL model it split cleanly). It ran on SDXL too (Nova Anime XL to MiaoMiao Harem v2.0). Refine with another model now works on Anima too: only the model file changes, the text encoder and VAE stay the canvas's. Both made pictures on a 6 GB card.

  • 新的按需 skill 分阶段出图的思路 取代原来的「Anima 插画·底图与 Tile 细化」:只讲一张图分几步画的几种路子(同一张图前后段换模型;锁住原图放大重画,可以在这一步顺手换成同类的第二个模型,不用放大完再精修一轮;画完换模型接着修,接几个都行),什么时候必须是同一类模型(沿用前一步提示词和 VAE 的都要同类;自带加载器的单文件换模型精修可以跨类,但要按新模型的习惯改写提示词),以及你选的画风 LoRA 要不要一起挂到第二个模型上(想保持画风就挂,想换画风才不挂)。先出底图给你挑时,AI 只跑第一段,你挑中的那张就是第二步用的底图(见下面)。不写具体数值,数值在做法里;你存过的路线卡和搭配优先原样用,你没选的画风和 LoRA 不加。

  • New on-demand skill Drawing a picture in stages replaces "Anima illustration: base and Tile detail": the ways to make one picture in passes (different models for its early and late steps; a locked redraw while enlarging, which may switch to a second model of the same family in that same redraw, with no extra refine pass after it; model after model, as many as wanted), when the models must be of one family (any switch reusing the first stage's prompts and VAE; the single-file refine with a loader of its own may cross families, its prompts rewritten for the new one), and whether your style LoRA goes onto the second model too (kept look: yes; a new look: no). The AI shows you the first pass alone, and the base you pick is the one the second pass grows from (below). No values: those live in the ways. A route card or combo you saved comes first, as it is; no style or LoRA you did not pick.

  • 修手·先放大再修 多了一轮 SAM3:放大重画之后,先用 SAM3 按文字找手(一张图最多 8 只),只在贴着手形的轮廓里重画,再用手部检测框补修一遍(参考作者就是这样两轮)。SAM3 的节点是 ComfyUI 自带的,模型(sam3.1,1.7 GB)在做法面板里一点就能下载,下载完核对 SHA-256;没有它时这个做法照旧只跑手部检测那一轮。

  • Fix hands: enlarge, then fix gains a SAM3 pass: after the light redraw, SAM3 finds the hands by the word (up to 8 in a picture) and redraws each inside its own outline, then the hand detector mends them in their boxes, as the reference author does in two passes. SAM3's nodes come with ComfyUI; its model (sam3.1, 1.7 GB) is one click in the Ways panel, checked by its SHA-256, and without it the way runs the detector's pass alone as before.

  • 半途换模型接着画 可以在换模型那一步顺手放大(1.5 倍):后面的步数在大尺寸上画细节,省掉画完再放大重画的那一轮;AI 加一个放大节点、改两个设置就行,前后用同一个模型也可以。放大前,前一段会先把画面收干净:带着剩下的噪声直接放大,整张图会变成彩色方块。skill“分阶段出图的思路”补了一句。

  • Switch models mid-way can enlarge at the switch (1.5×): the later steps draw the detail at the larger size, in place of an upscale pass after; the AI adds one node and changes two settings, with the same model or another. The first part finishes its picture before it is enlarged: enlarged with its leftover noise, the picture turned into coloured blocks. The skill "Drawing a picture in stages" says so in one sentence.

  • 新增 存档:在画布上改任何东西(参数、模型、节点连接),右键节点或画布空白处选“存档当前工作流”,就把整个工作流存成一个存档,画布下方出现存档条,之后按条上的 + 接着存。存档条放在画布区域里,不会盖住底部日志,收起来是贴在画布底边的小把手;条上只放最新的 6 个,更早的在存档页。存档跟着工作流走,切标签就换成那个工作流的;刷新、重启都还在,备份也带上它们。点存档的数字回到那一步,就在当前标签页,Ctrl+Z 能回到回退之前,也可以接着改出新的分支。每个存档设好跑几次,种子选 随机(每次换新,不管节点原来是固定还是随机)或 锁种(每轮一个种子,所有存档共用,公平对比),按 运行:AMB 先让 ComfyUI 检查每个存档,有问题(断线、模型没了)就一次都不排,并说明是哪个存档;没问题就按顺序全部排进队列。跑的是副本,你的画布和上面的种子都不动,可以接着改。条上的 看图 打开 存档页(“工作流”页顶上的 存档):所有存档按画布分块,标签关了、文件删了的也在,默认只展开当前画布。每个存档一张卡,左边是和上一个比改了什么(改了参数还是节点变了),右边是它出的图,放不下的折叠成最后一格 +N,点开是这个存档的图库,操作和图库页一样。点图看大图,查看参数 打开参数页,左右键在这个存档的图里切换。块顶上的 运行 按顺序跑这个画布的存档,再点别的块就排在后面,画布关了也能跑;每个存档也能单独 运行,多跑几张,每次换新种子。收藏为工作流 把存档存成一张新的工作流卡,之后互不影响;⋯ 里能回到这里、在新标签页打开(画布开着算一条新分支,画布没了这组存档接着用)或删掉(进回收站,图不受影响)。

  • New saves: change anything on the canvas (values, models, how nodes connect), right-click a node or the empty canvas and choose "Save this state", and the whole workflow is kept as a save in a bar under the canvas; the bar's + keeps the next ones. The bar sits in the canvas's area, clear of the logs under it, and folds to a handle on its bottom edge; it holds the newest six, the rest are on the Saves page. Saves go with their workflow: another tab shows its own, and they stay through a reload or a restart and go into backups. A save's number goes back to that state on the same tab, Ctrl+Z returns to where you were, and you can branch from it. Set each save's runs, pick Random seeds (new every run, whatever the nodes are set to) or Locked (one seed a round, shared by all saves, for a fair comparison), and Run: ComfyUI checks every save first, and a broken link or a missing model stops it before anything is queued, naming the save; else all of it is queued in order. Copies run, so your canvas and its seeds stay as they are and you can work on. Pictures on the bar opens the Saves page (Saves at the top of the Workflows page): every save, by canvas, those of closed tabs and deleted files too, only the open canvas's unfolded. Each save is a card, what changed from the save before (values or nodes) on the left, what it made on the right, those that do not fit folded into a last +N tile that opens the save's pictures as gallery cards, working as on the Gallery page. A picture opens large, and View Parameters opens its parameters, the arrow keys stepping through that save's pictures. Run at the top of a part runs that canvas's saves in order, and another part's goes in after it, closed canvases too; a save can also Run alone for more pictures, a new seed each. Keep as a workflow makes a new workflow card of a save, and the two go their own ways; ⋯ goes back to it, opens it on a new tab (a branch while its canvas is open; with the canvas gone the saves go on there) or removes it (to the Recycle Bin; its pictures stay).

  • 侧边停靠和窄窗口下,AI 工作台不再是一排看不懂的图标:任务上方是一排带字的标签,时间线、展台、节点结构、决策、素材、成果、做法,一点就切。“在右侧看”这类放不下侧栏时说不通的字也改了。任务列表是从左边盖上来的抽屉,后面露出当前任务,一看就知道是会话记录;按 Esc 先收起它,再按才关窗口。

  • 提示词页在宽窗口下铺满:左边词卡更宽,正面和负面两个框并排,占满高度,不再是两条细框加一大片空白。工作流的存档页在宽窗口下卡片并排,顶上写着有几个画布、几个存档。

  • With the browser docked or in a narrow window, the AI workbench no longer hides behind a row of icons: labelled tabs over the task lead to the timeline, Stage, Node map, Decisions, Materials, Results and Ways, and words such as "on the right" went where there is no right side. The task list is a drawer over the task shown, so it reads as the records of conversations; Esc puts it away first, and closes the window after that.

  • The Prompts page fills a wide window: wider cards on the left, the positive and negative boxes side by side and full height, instead of two thin boxes over empty space. The Saves page puts its cards side by side there, and says how many canvases and saves there are.

  • AI 工作台的标题栏清爽了:固定两行,常用的只留运行状态、改副本和切到副本,改名、隐私模式、接着哪个任务、移除都收进 ⋯;侧栏打开时也不再挤成四五行。时间线上面的阶段和本次成果并成一行,时间线多露出一截。出过错的任务在列表里不再整行标红,AI 试错的那一步也不再是一片红字。新任务 一点就开,不再先弹一个输入框。修好“做法”面板里显示成内部代号的标题。

  • The AI workbench's task head is tidier: two lines however wide, keeping the run status, the canvas choice and the way to the copy; rename, privacy mode, what it goes on from and remove are in ⋯, and an open side column no longer squeezes it into four or five lines. The stages and what the task made share one line over the timeline. A task with a failed step is no longer red in the list, and a step the AI got wrong is a mark, not a wall of red. New task opens at once, with no box to fill in first. A heading in the Ways panel no longer shows its internal name.

  • 你给任务挑了两个主模型时,AI 不用自己琢磨怎么分工:AMB 准备好几种用法问你(比如“第一个出底图,第二个用 Tile 放大重画”“第一个画前面,第二个接着画完”“只用第一个一次出图”),你存过同时用这两个模型的卡时,它排第一并标推荐。你选好后,AMB 在新标签页里把整条路线搭好,LoRA 也挂到每个模型上,AI 只写提示词。AI 看模型时还能看到这个模型在你哪些卡里用过、每张卡是什么路线。

  • 关卡里新增 新路线:AI 不用现成的卡和做法、自己往画布上放节点时,默认提醒一次“这条路线还没验证过,跑完看看效果,好的话存成卡”;选“先问我”,每个任务要先等你点头。

  • With two main models picked for a task, the AI need not work out who does what: Anomalous prepares the ways to use them and asks you ("the first draws the base, the second redraws it enlarged under Tile", "the first draws the early steps, the second finishes", "only the first, in one pass"), a card you saved with both on top and recommended. Once you choose, it builds the whole route on a new tab, the LoRAs on each model, and the AI only writes the prompt. The AI also sees which of your cards use a model and each card's route.

  • New routes in the Gates window: when the AI places nodes itself instead of using a card or a way, it reminds you once that the route is untried: see how the run turns out and save a card if you like it; with "Ask me first" it waits for your OK once a task.

  • 新增做法 出图:空画布也能起步。AI 说要出图,AMB 就用你给任务挑的模型(没挑就用 AI 指定的)搭好出图的节点:SDXL 一类是一个 checkpoint;Anima 是模型、文本编码器、VAE 三个文件,AMB 自己在你的模型文件夹里找到后两个。设置用验证过的那套(Anima:35 步、dpmpp_2m_sde_gpu、beta、CFG 4.4、832×1216,就是你两段放大卡的底图那一段)。画布上已经有出图流程时不会再叠一套。之后要放大或换模型,再按用途加做法。

  • New way Make a picture: an empty canvas is a start too. When the AI wants a picture, Anomalous builds its nodes on the model you picked for the task (else the one the AI names): one checkpoint for SDXL and its kin; for Anima the model, its text encoder and VAE, the last two found in your model folders. With the tested settings (Anima: 35 steps, dpmpp_2m_sde_gpu, beta, CFG 4.4, 832×1216, the base pass of your two-pass card). A canvas already making a picture gets no second one. Enlarging or another model after it are ways by their use.

  • AI 看你存的工作流卡时,每张卡多一行“路线”:分几段、每段用哪个模型、在哪放大、有没有 Tile 这类控制、最后接了什么(比如“2 段:出图 35 步(Harem 1.5)→ 放大 1.22 倍,用 Tile 在 Realskin 1.3 上重画 → ColorMatch”),不打开卡也知道它是一段出图还是两段放大,不会拿单段卡去做两段的活。

  • When an AI lists your saved workflow cards, each now has a route line: its passes, the model of each, where it enlarges, control such as a Tile, and what comes after ("2 passes: base 35 steps on Harem 1.5 -> enlarge 1.22x, redraw on Realskin 1.3 with Tile -> then ColorMatch"), so it tells a one-pass card from a two-pass one without opening it.

  • 先出几张底图给你挑:分几段出图时,AI 可以只跑第一段(底图),一次几张,每张换个种子。这些底图只是预览,不存进输出文件夹,你画布上的工作流也一点不动。跑完 AI 工作台里自动弹出 挑一张底图,你点 就选这张,只有这一张存进输出文件夹(和卡平时存底图的地方一样),种子也记住。AI 接着跑整张卡时:像“两段放大”这种底图画完再交下去的卡,第二段直接读你挑的那张图,底图不再重画(6 GB 显卡上整张从约 190 秒变成约 110 秒);“半途换模型”交接的是画到一半的东西,就按记住的种子重画,跑完 AMB 再拿它和你挑的那张比一比,对不上就告诉你和 AI。预览存在 ComfyUI 的临时文件夹,ComfyUI 重启就清掉:没挑之前重启了,点的时候会告诉你那张的种子,让 AI 重画。

  • Bases to pick from: for a picture in passes, the AI can run the first pass (the base) alone, a few at once, each with its own seed. They are previews only, nothing goes into the output folder, and the workflow on your canvas stays as it is. When they are done, Pick a base opens in the AI workbench; Take this one keeps only that one in the output folder (where the card saves its bases) and holds its seed. When the AI then runs the whole card, a card that hands on the finished base (the two-pass upscale) has its second pass read the picture you picked, and the base is not drawn again (on a 6 GB card the run went from about 190 s to about 110 s); the mid-way switch hands on an unfinished picture, so its base is drawn again from the held seed, and when that run saves its base Anomalous compares it with yours and tells you and the AI if they differ. Previews live in ComfyUI's temp folder, emptied when ComfyUI restarts: picked after a restart, the panel tells you the seed to draw it again.

  • 跑完的图下面新增 效果好 / 不太行 两个按钮,点了直接告诉 AI(点“不太行”它就换做法、设置或提示词再跑),再点一下收回。关卡 里新增 第三方节点:选“只用做法里的”,AI 自己加节点时只能用做法里搭好的组合和你给任务挑的,ComfyUI 自带的节点照常能用。等不到回答时设成按推荐的继续,挑图面板里 AI 推荐的那张也会替你选上。

  • Under a finished run's pictures, Turned out well / Not quite tells the AI at once (with "Not quite" it changes the way, a setting or the prompt before running again); a click again takes it back. The Gates window has a new Custom nodes row: with "Only the ways'" the AI may add only the nodes of the ready-made ways and what you picked for the task; ComfyUI's built-in nodes stay allowed. With "go on with the recommended" for unanswered questions, a picks panel takes the picture the AI recommended too.

  • 新增 隐私模式:任务标题栏的 🔒 一点,这个任务的图和提示词原文都不会发给 AI。图一张都不给(放大看图、图片信息、跑完的预览、你递给它的图、AI 软件里任务卡片上的图都一样),它只知道出了几张、成没成功;提示词和图片文件名换成 [P1]、[img1] 这样的代号:画布上、图片信息里、提示词库、搭配、模型的触发词和介绍、你自己改画布的记录里都一样(只有节点的“文字输入”被换掉,模型名、数字、选项照常给,第三方节点也一样)。AI 拿代号去跑图、写提示词、放挑图面板,AMB 再换回原文,所以它照样能干活,你那边看到的也都是原样。要新的提示词时,AI 会放一个 提示词框:你在里面写几条提示词、反向提示词,可以从图库选或拖进参考图,原文只存在 AMB 里;框上方还会按画布主模型的大小提醒你(小模型手容易画坏,可以选简单动作或藏手,只是可选)。新增 skill“隐私模式·怎么干活”,隐私任务里的 AI 开工时会被指过去;隐私任务里 AI 只拿得到这一个 skill,做法 面板的 skill 一栏也只显示它,关掉隐私后才换回普通的 skill。关卡 窗口里设新任务默认开不开;打开、关掉都记在时间线上并告诉 AI。所有工具的结果出门前都过同一道关,测试会在每个工具的结果里埋“暗号”提示词和图片名,确认一个都漏不出去。

  • New: privacy mode. The 🔒 in a task's header keeps that task's pictures and prompt text from the AI. No picture goes (a closer look, picture info, a run's previews, pictures you hand over, the task card in the AI app alike): it hears how many came out and whether they did. Prompts and picture file names become tokens like [P1] and [img1]: on the canvas, in picture info, saved prompts, combos, models' trigger words and descriptions, and your own canvas changes (only the nodes' text inputs are swapped; model names, numbers and choices stay, third-party nodes too). The AI runs, writes prompts and shows pick panels with the tokens and Anomalous puts the texts back, so it still works and you see everything as it is. For new prompts the AI shows a prompt box: you write prompts and a negative one and add reference pictures from the gallery or by dropping them; the texts stay in Anomalous, and above the box a word on what suits the canvas's model (a smaller model often gets hands wrong: a simple pose, or hands hidden, as an option). A new skill, "Privacy mode: how to work", is where a private task's AI is pointed; it is the only skill a private task's AI gets, and the only one the Ways panel lists for it, the usual skills coming back once privacy mode is off. The Gates window sets it for new tasks; turning it on or off is on the timeline and told to the AI. Every tool's result passes one check on its way out, and a test plants canary prompts and picture names in every tool's result to make sure none gets through.

  • AI 可以一次排一串提示词:跑图时给几条提示词(最多 8 条),AMB 按顺序一条条填进去排队,其余设置(比如同一个种子)每条都一样,算同一次跑图,图都在时间线的同一步里。以前要一条条跑、每次一个来回。哪一条排不上就停在那里,前面排上的照常跑完。

  • An AI can queue a list of prompts: a run given several prompts (up to 8) fills them in and queues them in order, the other settings (the same seed, say) the same for each, as one run whose pictures share one step of the timeline, instead of a round trip per prompt. One that cannot be queued stops the rest; those queued before it run as usual.

  • 同一个模型的几个版本放在一起:扫描认出来自同一个 Civitai 模型页的文件,在模型列表里收成一张卡片,卡片右边有一条带数字的箭头,点一下其余版本就从它右边滑出来、并排挨着,同一个模型的卡片顶上是同一道橙色边;最后一张右边的箭头点一下又收回去。展开后每张卡片的版本名旁边有个图钉,点哪个,收起时就显示哪个版本(封面、名字、插入节点都跟着换)。展开过的模型下次打开还是展开的;版本多的话,类型下面那排最右边的 展开全部版本 一下全展开,再点一下全收起。只认 Civitai 的模型编号,文件名再像也不会合在一起;没扫描、Civitai 上找不到的模型照旧各自一张。模型详情里名字下面一排列出这里有的几个版本,点一下就过去;再下面的 Civitai 上的全部版本 点了才联网,列出这个模型页的每个版本(底模、日期、大小),这里有的可以直接打开,没有的点 下载 放到这个模型旁边,下载完按 SHA-256 核对;名字和已有文件重了,就在名字后面加上版本名,不覆盖。

  • Versions of one model sit together: files a scan found on the same Civitai model page fold into one card in the model list with a numbered arrow on its right edge: a click slides the other versions out beside it, side by side, the cards of one model under one orange top edge, and the arrow on the last folds them back. Open, each card has a pin by its version name: the pinned version is the one the folded card shows (its cover, name and Add node too). A model left open stays open; Open all versions, at the right end of the row under the types, opens them all, and again folds them all. Only Civitai's model id groups them, never a similar file name; models not scanned or not on Civitai keep a card each. A model's details list the versions here under its name, one click each, and All versions on Civitai, asked only when pressed, lists every version of the model page (base model, date, size): the ones here open, the others Download next to this one, checked by their SHA-256; a taken file name gets the version's name added, nothing is replaced.

  • 新增 skill,和做法并列:做法是 AMB 替 AI 搭好的节点,skill 是 AI 遇到某类活时自己去取的建议,不搭节点。第一个是“写提示词·动作适合模型”,按画布主模型的大小给建议:AMB 读模型文件头数出画图部分的参数量(压缩过的 GGUF 也一样,不看文件大小),100 亿参数以上的大模型(Flux.1、Flux.2、千问图像这类)手和复杂动作画得好,AI 照画面需要写,不藏手;小模型(SD1.5、SDXL、Illustrious、Anima 这类)没指定姿势就选画得清楚的动作,手不是重点时可以自然遮挡或构图不露手,但这只是可选,不是规定;认不出模型时两种情况都告诉它。你指定的姿势和你给的提示词照原样。AI 开工时只看到 skill 的名字和一句话,要写提示词时才去取全文,工具说明里只留一句提示,给 AI 的工具清单不变长。AI 工作台右边的做法面板最下面列着 skill,每个有开关,关掉的 AI 就拿不到。AI 取了哪个 skill、按哪一档给的建议,时间线上都有一行 Skill;之后那次跑图下面也写着“这次跑图前参照了:……”。

  • New: skills, beside the ways. A way is nodes AMB builds for the AI; a skill is advice the AI fetches when a kind of job comes up, and it builds nothing. The first: "Writing prompts: poses the model can draw", by the size of the canvas's main model: AMB counts its drawing model's parameters from the file's header (a compressed GGUF alike; the file size is not the yardstick). A large one, 10 billion or more (Flux.1, Flux.2, Qwen-Image and the like), draws hands and hard poses well, so the AI writes what the picture wants with no hiding; on a smaller one (SD 1.5, SDXL, Illustrious, Anima…), with no pose asked for, it picks one the model draws cleanly, and hands that are not the point may be hidden naturally or framed out, as an option, not a rule; a model it cannot tell gets both cases. Poses you ask for and prompts you give stay as they are. At the start of a task the AI sees only each skill's name and a line, and fetches the advice when it writes a prompt; the tool descriptions only point at it, so the tool list AI apps hold does not grow. The skills are at the bottom of the Ways panel on the right of the AI workbench, each with a switch: the AI can't fetch the ones turned off. Each skill the AI fetches is a Skill row on the timeline, with the form its advice came in, and the run after it says "Went by before this run: …".

  • 当前节点里的 ⇄ 对比 只比这个节点了:以前会把整个工作流的参数都列出来。现在只列这个节点存下的值和画布上现在的值哪里不一样(提示词按标签对比),每张参数卡片都能对比。

  • ⇄ Compare in the current-node panel now compares that node alone: it used to list the whole workflow's values. It shows how the saved values differ from the node's values on the canvas (prompts tag by tag), on every saved set.

  • 更换模型 时弹出的模型卡片和模型页的一样了:封面(视频封面先显示第一帧,指着才播放)、底模角标、版本、好认的名字和文件名。以前大多是空白卡片、只有文件名。同一模型的几个版本仍然各占一张卡,不折叠。

  • The model cards in Change model are now the models page's: the cover (a video's first frame until pointed at), the base model badge, the version, the readable name and the file name. Most used to be blank with the file name only. Versions of one model still get a card each, not folded.

  • 修正点击穿透:当前节点 →「查看全部」的历史生成图库往下滚以后,点“关闭图库”会点到滚上去的图片。

  • Fixed a click passing through: in the generated-pictures gallery (current node → View all), once scrolled, “Close gallery” opened the picture scrolled up under it.

  • 修正 AI 工作台在侧边栏模式下标题栏放不下:右上角的放大、停靠、关闭按钮被挤出窗口。现在按 AMB 窗口自己的宽度(不再按整个浏览器的宽度)收起标签文字,只留图标。

  • Fixed the AI workbench's title bar when docked: the maximize, dock and close buttons were pushed out of the window. Its tabs now fold to icons by the width of the Anomalous window, not of the whole browser.

  • 修正模型页 整理 的几处误判:装了 ComfyUI-Anima-LLLite 节点包时,放在 controlnet 里的 Anima LLLite 不再算放错(这个节点包就从 controlnet 读,挪走反而用不了;没装才建议挪到模型补丁);带 LoRA 层的 IP-Adapter(比如 Anima IP-Adapter)不再被认成 LoRA;同一个文件在 controlnet 和 model_patches 各一份、两边各有节点在读时,不再算“重复、白占硬盘”,也不再建议删一份,只在下面写明两处都在用。

  • 模型类型识别又查了一遍:拿 ComfyUI-Manager 模型清单里 315 个模型的真实文件头逐个核对,修掉了会认错的几种:混元 DiT 整合包被当成 VAE、LBM 打光模型被当成 VAE、SUPIR 被当成 ControlNet、XLabs 的 Flux LoRA 和 IP-Adapter 被当成扩散模型、完整 CLIP(文字+图像)被当成文本编码器、PhotoMaker 被当成 LoRA、只存差值的 LoRA 被当成扩散模型、去掉 VAE 的整合包被当成扩散模型。两种加载器从不同文件夹读的文件(SUPIR、完整 CLIP)只标“不确定”,放哪边都不报放错。只有模型主体的文件放在 checkpoints 里也不再算放错:ComfyUI 的大模型加载器会把它当扩散模型读进来。新认得:diffusers 格式的扩散模型(IC-Light、可图、Lotus)、Stable Cascade、SDXL 的 T2I-Adapter、RCAN / SPAN / 小型 Real-ESRGAN 放大模型。核对后 315 个里 301 个和清单一致,剩下的要么是有意标的“不确定”,要么认不出(LTX 的潜空间放大、Stable Cascade 的 VAE 这类),认不出的照旧不下结论。扫描页和模型卡片上原来写“看起来不是绘图模型”,其实只是认不出,现在改成 认不出类型;规则改了以后,存下的识别结果会自动重读一遍。

  • Model types were checked again, against the real headers of 315 models in ComfyUI-Manager's list, and the layouts it misread are fixed: a Hunyuan-DiT checkpoint and LBM's relighting model taken for VAEs, SUPIR for a ControlNet, XLabs' Flux LoRAs and IP-Adapter for diffusion models, a whole CLIP (text and vision) for a text encoder, PhotoMaker for a LoRA, a LoRA of weight differences and a checkpoint saved without its VAE for diffusion models. Files two loaders read from different folders (SUPIR, a whole CLIP) are only "not sure", never misplaced in either. A lone drawing model in checkpoints is no longer misplaced: ComfyUI's checkpoint loader reads it as a diffusion model. Newly told: diffusers diffusion models (IC-Light, Kolors, Lotus), Stable Cascade, SDXL T2I-Adapters, RCAN / SPAN / compact Real-ESRGAN upscalers. Checked again, 301 of the 315 agree with the list; the rest are "not sure" on purpose or not told (LTX latent upscalers, Stable Cascade's VAEs), which still get no verdict rather than a wrong one. The scan page and model cards said "looks like no drawing model" when it only could not tell: now Kind not told. Answers kept from older rules are read again.

  • The models page's Tidy misjudged a few files: an Anima LLLite in controlnet is no longer misplaced while the ComfyUI-Anima-LLLite pack is loaded (it reads them from there; moving one broke it; without the pack it still suggests model patches); an IP-Adapter carrying LoRA layers (the Anima IP-Adapter) is no longer taken for a LoRA; one file in both controlnet and model_patches, each read by a node, is no longer "identical, wasting space" with a copy to delete: a note says both are in use.

  • 跑图前的尺寸提醒放宽一档:6GB 显卡采样到约 420 万像素才提醒(以前 210 万),12GB 约 840 万。今天在 6GB 的笔记本显卡上实测,400 万像素一次约 100 秒,并不吃紧;以前的门槛会让 AI 动不动就问你要不要停。解码大图的提醒不变。

  • The size warning before a run is less eager: on a 6 GB card from about 4.2 megapixels (2.1 before), 8.4 on 12 GB. A 6 GB laptop card sampled 4 megapixels in about 100 s today; the old line had the AI ask you to stop runs that were fine. The warning for decoding a large picture is as before.

  • 新做法 修手·先放大再修(备选):你递给 AI 一张手画坏的图,AMB 先把整张图放大到约 400 万像素(原图越小放得越大:1280×832 这种放 2 倍,已经很大的图不放,最多 2 倍)轻轻重画一遍,让手变大、模型有地方画手指,再自动找出每只手、放大重画后贴回去。这是照着参考作者的做法,在 6GB 的笔记本显卡上一组组试出来的:只修手不先放大,大约一半能成;先放大 1.5 倍再修,大约 4 张里 3 张能成;放大 2 倍,同样的种子手指更清楚。Anima 模型上有 Anima Turbo LoRA(做法面板里可以下载)时只跑 10 步,1280×832 的图约 100 秒;没有也能用,慢三倍左右。AI 会一次出 4 张,先自己放大看手,把手指数不对、粘在一起、大小不对的去掉,都不行就换种子再来,最多两轮,再让你在只放大手那一块的面板里挑。整张图会有些细小变化(脸、水花),构图不变。实测不如 Codex 用 FLUX-Fill 每只手单独裁一块重画的办法,所以只作备选:AI 手里有那个办法时先用那个。

  • New way Fix hands: enlarge, then fix (fallback): from a picture with bad hands, Anomalous enlarges the whole picture to about 4 megapixels (the smaller it is, the more: 2× for 1280×832, not at all for a large one, at most 2×) and redraws it lightly, so the hands grow and the model has room for fingers, then finds every hand and redraws it larger. Taken from the reference author's workflows and tried step by step on a 6 GB laptop card: the hand pass alone fixed about half the tries, enlarging 1.5× first about 3 in 4, and 2× gave sharper fingers in the same seeds. On an Anima model with the Anima Turbo LoRA (downloadable in the ways panel) it is 10 steps, about 100 s for 1280×832; without it, about three times slower. The AI makes 4 tries, checks the hands itself first, drops wrong finger counts, fused fingers and wrong sizes, tries new seeds at most twice more, and lets you pick in a panel zoomed on the hand. Small things change over the picture (faces, splashes); the layout stays. In the user's tests it did worse than Codex's FLUX-Fill redraw of one crop per hand, so it is a fallback: the AI uses that one when it has it.

  • 修一处(比如一只手)更容易成:从 Codex 这几天修手的试验看,重画的力度太大时,新画的手会撑满你圈的范围,于是手越修越大。现在“局部重绘”默认力度降到修东西用的 0.45(官方节点那块 0.5),AI 也会被告知:圈得贴着手、保持低力度、一次出 4 张,再用一个只放大手那一块的挑选面板给你挑,挑完用同样放大的前后对比给你看。展台上的“挑一张”和前后对比滑块都能只显示每张图的同一小块。

  • Fixing one spot (a hand) works more often: Codex's hand tests showed that with a strong repaint the new hand fills the whole marked area and grows each time. The repaint block's strength is now 0.45 by default, a fix (0.5 for the built-in one), and the AI is told to mark tightly, keep it low, make 4 tries and let you pick in a picks panel zoomed on the hand, then show before and after zoomed the same. The stage's picks and slider can show the same small part of every picture.

  • AI 交给“加载图像”的图不再被误报“没选图片”:AI 拿输出文件夹里的图去重画时,AMB 以前在节点里写“名字 [output]”,ComfyUI 重启、运行记录清空后,新版前端会把它标成“缺少图片”,其实图还在。现在和图库右键“放到画布”一样,先把图复制一份进 input 文件夹(同一张图只存一份,透明部分保留)再填进去。ComfyUI 页面上标红的“缺少图片/视频”,AI 用 get_situation 看现状时也能看到,并知道怎么补。

  • Pictures the AI gives a Load Image are no longer flagged missing: for an output picture Anomalous used to write "name [output]", which ComfyUI's newer page marks as a missing file once its run history is gone (after a restart), though the file is there. Now it is copied into the input folder first (one copy per picture, transparency kept), as the gallery's "Put on the canvas" does. Files ComfyUI's page marks missing also reach the AI in get_situation, with how to fix them.

  • 让 AI 一次出几张时,每张都不一样:以前种子设成“固定”时,AI 说一次出 4 张,ComfyUI 会把 4 次当成同一个任务,只出 1 张。现在一次出几张时,固定的种子会一张一张往上加(123、124、125…),出完画布上的种子还是原来的数、还是“固定”。修手、修脸这种看运气的活,可以一次出 4 张让你挑。

  • A batch of tries from the AI gives different pictures: with a seed set to "fixed", asking for 4 used to queue the same prompt 4 times and ComfyUI made 1 picture. Now a fixed seed counts up for the batch (123, 124, 125…) and is as it was afterwards, still fixed. Hands and faces, which depend on luck, can come 4 at a time for you to pick from.

  • 跑图前的尺寸提醒多算了一步:放大模型之后最后要存多大。要存出超过约 6400 万像素的图(比如 5120×3328 再放大 4 倍,存出来 20480×13312、PNG 约 300 MB),时间线上会出现黄色提醒,AI 也会收到,建议放大倍数小一点或者从小一点的图开始放大。

  • The size check before a run now also counts what is finally saved after an upscale model: a picture past about 64 megapixels (5120×3328 upscaled 4x is 20480×13312, about 300 MB as PNG) gets a yellow warning on the timeline, and the AI is told, with the advice to upscale less or start from a smaller picture.

  • 超大图不再卡住图库:几亿像素的图(比如把大图再放大 4 倍,存出来几百 MB),以前缩略图做不出来时,AMB 会把整张原图当缩略图发给浏览器,图库和 AI 工作台一翻到它就卡。现在缩略图在后台单独的进程里做一次就存下来(2.7 亿像素的图约 6 秒),不占 ComfyUI 出图的内存;大到连这样也做不了的,卡片上显示尺寸和大小,不再加载原图;截不出第一帧的视频(文件坏了、格式解不开)也一样,卡片显示文件大小,不再把整个视频发给浏览器。卡片角上标“超大图 · 324 MB”。点开看大图、图片详情、AI 工作台里放大看的,都是缩到 4096 像素的预览,下面写着原图多大,要看原图点“打开原图”(在浏览器新标签页里打开)。图片详情读参数也不再把整张图下载下来。

  • Very large pictures no longer freeze the gallery: for a picture of hundreds of megapixels (a big one upscaled 4x, hundreds of MB), a card image that could not be made used to be the whole original, sent to the browser every time the gallery or the AI workbench showed it. Now it is made once in a separate background process and kept (about 6 s for 272 megapixels), outside ComfyUI's own memory; one too large even for that shows its size instead, never the original, and so does a video whose first frame cannot be read (a broken file, no decoder). Its card says "Very large · 324 MB". The large view, the image details and the AI workbench's larger pictures show a 4096 px preview with a note saying how large the original is, and "Open the original" opens it in a new browser tab. Reading a picture's settings in its details no longer downloads the whole file.

  • 做法按用途合并成 7 个:局部重画、换背景、放大、换模型精修、照着画、修脸修手、图片处理。AI 只说"用途 + 想要哪种"(比如"放大,2x"、"照着画,姿势"),AMB 按你装了什么、画布上是什么模型,自动挑能跑的、效果最好的那一块;没装的照样按节点包列出,AI 能告诉你该装什么。开新任务时 AI 只看到 7 行,做法再多也不会变长。

  • The ways are grouped into 7 uses: repaint, new background, upscale, redraw with another model, follow a picture, faces and hands, picture work. The AI names a use and the kind it wants ("upscale, 2x", "follow, pose") and Anomalous picks the best block this ComfyUI can run, by what is installed and the canvas's model; the missing ones are still listed by pack, so the AI can say what to install. A new task's directory is 7 lines, however many ways there are.

  • 新做法 换画风·构图不变:你递给 AI 一张图,AMB 读出它的轮廓线(ComfyUI 自带的 Canny),用 Union 控制网钉住构图,再按提示词里的新画风整张重画:人物、姿势、位置不变,画法换掉(比如水彩、厚涂、平涂)。实测强度 0.8 只是变亮一点,1.0 才真正换了画风,轮廓线照样守住构图,所以默认 1.0。RTX 4050 6GB、SDXL 832×1216 约 41 秒。

  • New way New style, same layout: from a picture handed to the AI, its outlines (ComfyUI's own Canny) hold the layout through the union ControlNet while the whole picture is redrawn in the prompt's new style (watercolour, painterly, flat colours): who, pose and place stay. Tried: at denoise 0.8 it only got lighter, at 1.0 the style changed and the outlines still held the layout, so 1.0 it is. About 41 s for SDXL 832×1216 on an RTX 4050 6 GB.

  • 新做法 换模型精修:出完图以后,换另一个单文件模型(SD1.5、SDXL 各家族之间都能换,配它自己的 LoRA;Anima 这种分三个文件的还不行)把整张图轻轻重画一遍(强度 0.4),另存一张(文件名以 refined 开头)。一个模型的构图加上另一个模型的画风和细节。新模型有自己的提示词节点,先照抄原来的提示词。实测 RTX 4050 6GB、Illustrious 出图 → NoobAI 精修,832×1216 整次约 66 秒。

  • New way Refine with another model: after a picture is made, another single-file checkpoint (any SD 1.5 or SDXL family, with its own LoRAs; not yet a three-file model like Anima) redraws it lightly (0.4) and saves it beside it (prefix refined): one model's composition with another's style and detail. The new model has prompts of its own, copied from the first. About 66 s for a whole run, Illustrious then NoobAI at 832×1216, on an RTX 4050 6 GB.

  • 新做法 放进另一张图:你递给 AI 两张图,一张有人物、一张是背景。AMB 抠出人物,把背景缩放裁切成同样大小,贴上人物,再把人物以外的地方轻轻重画一遍(强度 0.5),让脚下的影子、地面和光线接上,背景的样子基本不变。实测 RTX 4050 6GB、SDXL 832×1216 约 40 秒。做法现在可以接第二张图了(add_nodes 的 scene)。

  • New way Put a character into a picture: from two pictures handed to the AI, a character and a place, the character is cut out, the place scaled and cropped to its size, the character pasted on, and everything else lightly redrawn (0.5) so shadow, ground and light meet it while the place keeps its look. About 40 s for SDXL 832×1216 on an RTX 4050 6 GB. A block can take a second picture now (add_nodes' scene).

  • 新做法 换场景·人物不变:你递给 AI 一张图,AMB 自动抠出人物(inspyrenet-rembg),只重画人物以外的地方,AI 在提示词里写新场景;人物边缘留 6 像素一起重画,接缝更自然。实测 RTX 4050 6GB、SDXL 832×1216 约 30 秒。人物身上的光还是原图的光,换到光线差别很大的场景时看得出来(以后可以加重新打光)。AI 递给做法的图也认 名字.png [output] 这种写法了。

  • New way New scene, same character: from a picture handed to the AI, the character is cut out (inspyrenet-rembg) and only the rest is drawn anew from the prompt's new scene; 6 px of its edge are redrawn too for the seam. About 30 s for SDXL 832×1216 on an RTX 4050 6 GB. The character keeps the old picture's light, which shows in a very differently lit scene (relighting may come later). A picture handed to a block may also be written name.png [output].

  • AI 有了现成的做法可以照着用(参考做法):从擅长这些活的工作流里提炼出来的几块节点,AI 说一声用哪块,AMB 就在画布上搭好、接进你的工作流里,一次调用,不用 AI 一个个节点地搭。AI 开工时会收到这份清单,每块一行。

    • 局部重绘·裁切缝合:把圈出的地方连同周围一圈一起裁出来,放大到 1024 重画,再缝回原图,别处一点不动。AMB 顺手把采样强度调到 0.45(修东西的力度:再高,新画的东西会撑满你圈的范围,手会变大),并提醒 AI 圈得贴着要修的东西、把提示词改成只写这块地方(比如“一只手,五根手指”),不然它会把整张图的内容(比如一整个人)画进这一小块里;手和脸一次出 4 张让你挑。需要节点包 ComfyUI-Inpaint-CropAndStitch。
    • 局部重绘(官方节点):没装上面那个包时用它,只用 ComfyUI 自带的节点。
    • 部位细化(脸和手):出完图以后找出脸和手,各自放大重画再贴回去,细化后的图另存一张(文件名以 detailed 开头)。需要 Impact Pack 和 Impact Subpack,以及 models/ultralytics/bbox 里的检测模型。
    • 局部重绘·LanPaint:把采样器整个换成 LanPaint 的(步数、CFG、种子照搬原来的),它每一步都多想几遍,新画的部分和周围融合得最好,任何模型都能用,不用专门的重绘模型;代价是慢好几倍。需要节点包 LanPaint。
    • 放大:用放大模型放大到 1.25 倍,再用同样的设置轻轻重画一遍(强度 0.35),细节是真画出来的而不是糊的,另存一张(文件名以 upscaled 开头)。只用 ComfyUI 自带的节点,自动挑你装的放大模型。参考工作流作者也只用 1.25 倍,6GB 显存的卡上 1.5 倍就太吃力了。
    • 分块放大:放大 2 倍,按 1024 的小块逐块重画,多大的图都放得下,比上一个慢。需要节点包 ComfyUI_UltimateSDUpscale。
    • 照着姿势画、照着构图画:你递给 AI 一张图,AMB 读出它的姿势(OpenPose)或深度(Depth Anything V2),用控制网引导出图。控制网模型由 AMB 按画布上底模的类型自动挑(SDXL 的配 SDXL 的,同底模的优先),参考图自动复制一份到输入文件夹(只新建,不覆盖)。需要节点包 comfyui_controlnet_aux。
    • 迭代放大:分 3 小步放大到 1.3 倍,每步轻轻重画,比一步到位更稳。用 Impact Pack。
    • 双模型放大:和"放大"一样,但重画换一个更稳的模型(底模放大时容易走样就用它);新加的模型加载器先是同一个模型,AI 按你的意思换。只用自带节点。
    • 背景移除:把一张图的背景抠掉,存成透明 PNG(文件名以 nobg 开头)。需要节点包 comfyui-inspyrenet-rembg。
    • 读图出标签:你递给 AI 一张图,WD14 打标器读出它的 Danbooru 标签,AI 照着写提示词。需要节点包 ComfyUI-WD14-Tagger,第一次用时会下载打标模型(约 1.2GB)。AI 运行后,节点输出的文字(比如这些标签)现在也会带回给它。
    • 装了 KJNodes 时,几种放大都会在最后把颜色对齐到原图(参考工作流的做法),防止放大后颜色跑偏;没装也照常能用,面板上写着"装了 KJNodes 会更好"。
    • 照着姿势画改用 DWPose 读姿势:二次元人物也认得准,连脸和手指都标(原来的 OpenPose 认二次元常认错)。人物挤成一团、画风很浓的图可能认不出,读出的骨架图或深度图会作为预览给 AI 看,空的它就知道换照着构图画。第一次用时会下载 DWPose 的模型(约 350MB)。
    • 平铺放大:放大到 1.25 倍,重画时用平铺控制网把画面钉住(强度 0.6,到 80% 的步数,参考作者的参数),所以重画强度可以调到 0.5,细节多得多,构图也不跑。SDXL 用 Union 控制网,NoobAI/Illustrious 优先用 NoobAI 的平铺控制网,Anima 用 Anima 的修复 LLLite(较慢)。
    • 线稿上色:你递给 AI 一张黑白线稿,AMB 把线条交给线稿控制网(Union 的线稿模式),AI 在提示词里写颜色和画风,按线稿画出彩色图。
    • Union(万能)控制网用之前,AMB 会告诉它这次是姿势、深度、平铺还是线稿。还没扫描过的新控制网按它所在的文件夹(SDXL_1.0、NoobAI……)认底模。做法说缺某个模型时,AMB 先把 ComfyUI 的文件列表刷新一次再找:刚下载到子文件夹里的模型,ComfyUI 自己不重启是看不见的。
    • 放大类的块优先用"够用的最小倍数"的放大模型(放大到 1.25 倍用 2 倍模型,比 4 倍模型再缩回来省事)。
    • 做法会按画布上的底模换搭法:画布上是 Anima 模型时,照着构图画、照着姿势画改用 Anima 专用的 LLLite 控制网,局部重绘·裁切缝合会加上 Anima 的重绘 LLLite。AI 照样只说用哪块。需要节点包 ComfyUI-Anima-LLLite 和 models/controlnet 里的 anima-lllite 模型。Anima 的 LLLite 比 SDXL 的控制网温和,提示词要和参考图说的一致(几个人、在干什么),AI 会被这样提醒。
    • 画布上有好几个采样器时,AI 要指明用哪一个,AMB 不乱猜。
  • AI 工作台右边多了做法面板:按"要做的事"分组列出所有做法,每块写着能干什么、从哪学来的。缺节点包或缺模型的变灰,写明缺什么,点缺的节点包会打开 ComfyUI-Manager 去搜它。每块有个开关:关掉的,AI 就看不到,硬要用也会被拒绝。AI 开工时只收到开着又能用的做法,其余的只附一行"装了什么还能用什么"。

  • 做法面板上每块写着怎么测过的:在哪张卡上(目前都是 RTX 4050 笔记本,6GB 显存)、多大的图、整次运行用了多久;参数也是按这张卡定的(比如放大只到 1.25 倍)。AI 用某块时也会收到这句,能提前告诉你大概要等多久。

  • AMB 维护一份做法用到的模型清单(api/block_models.json):每个模型的下载链接(HuggingFace、Civitai、GitHub)、大小、SHA256、许可证和该放的文件夹。每块做法列出它推荐的模型,面板上已有的打勾(改过名的也认得出),缺的点一下就用 AMB 的下载器下载,下完核对哈希。把做法分享给别人时,对方也能照着这份清单拿到一模一样的文件。

  • 线稿上色在 Anima 模型上改用 Anima 的线稿 LLLite。

  • 每块做法现在是一个单独的数据文件(api/blocks/),里面有它的节点、要的节点包和模型、给 AI 的话、面板上的中英文说明。以后收集到新做法,加一个文件就行。

  • AI has ready-made ways of doing things (reference blocks): a few groups of nodes distilled from workflows that do one job well; the AI names one and Anomalous builds it on the canvas, spliced into your workflow, in one call instead of node by node. The AI gets their list, a line each, when it begins a task.

    • Repaint, crop and stitch: crops the marked area with a margin around it, repaints it at 1024 and stitches it back, the rest untouched. Anomalous sets the sampler's denoise to 0.45 (a fix: higher, the new thing fills the whole mark and a hand comes out too big) and tells the AI to mark tightly, write the prompt for that area only ("a hand, five fingers"), or it paints the whole picture's subject (a whole person) into it, and to make 4 of a hand or face for you to pick. Needs ComfyUI-Inpaint-CropAndStitch.
    • Repaint (built-in nodes): for when that pack is not installed, ComfyUI's own nodes only.
    • Face and hand detailing: after a picture is made, finds faces and hands, repaints each larger and puts it back, saving the detailed picture beside it (prefix detailed). Needs Impact Pack and Impact Subpack and detector models in models/ultralytics/bbox.
    • Repaint with LanPaint: the sampler becomes LanPaint's (same steps, cfg and seed), which thinks several times a step, so the new part blends in best, with any model and no inpainting model; several times slower. Needs LanPaint.
    • Upscale: an upscale model enlarges the picture to 1.25×, then a light redraw with the same settings (denoise 0.35) makes the new detail real rather than smeared; saved beside it (prefix upscaled). ComfyUI's own nodes, with an upscale model you have. 1.25× is what the source workflow used; 1.5× is too much for a 6 GB card.
    • Tiled upscale: 2×, redrawn in 1024 tiles, so any size fits; slower. Needs ComfyUI_UltimateSDUpscale.
    • Follow a pose, follow a layout: from a picture you hand the AI, Anomalous reads the pose (OpenPose) or depth (Depth Anything V2) and steers the picture with a ControlNet it picks to fit the canvas's model (SDXL with SDXL, the very same base first); the picture is copied into the input folder, never over a file. Needs comfyui_controlnet_aux.
    • Iterative upscale: 1.3× in 3 small steps, a light redraw each, steadier than one jump. Uses Impact Pack.
    • Upscale with a second model: like Upscale, but a second, steadier model redraws it (for a base model that goes wrong enlarged); its new loader starts as the same model and the AI changes it as you say. Built-in nodes only.
    • Remove the background: saves a picture with its background cut away as a transparent PNG (prefix nobg). Needs comfyui-inspyrenet-rembg.
    • Tags from a picture: from a picture you hand the AI, the WD14 tagger reads its Danbooru tags for the AI to write the prompt from. Needs ComfyUI-WD14-Tagger; the first time it downloads its model (about 1.2 GB). What a run's nodes show as words (these tags) now reaches the AI with the run.
    • With KJNodes, the upscales match their colours to the original at the end (as the source workflow does), so they don't drift; without it they work as before and the panel says they're better with KJNodes.
    • Follow a pose reads the pose with DWPose: anime characters too, face and fingers included (OpenPose often misread them). A crowded or very stylised picture may come out empty; the pose skeleton or depth map is left as a preview for the AI, which then turns to Follow a layout. The first time, DWPose downloads its models (about 350 MB).
    • Upscale held by tile: 1.25×, a tile ControlNet holding the picture while it is redrawn (strength 0.6 to 80% of the steps, the source's), so the redraw can be stronger (0.5): much more detail, the layout kept. SDXL uses the union ControlNet, NoobAI and Illustrious NoobAI's tile one first, Anima its tile-repair LLLite (slower).
    • Colour a line drawing: from a black-on-white drawing you hand the AI, its lines steer a lineart ControlNet (the union one's lineart mode) while the prompt says the colours and style.
    • A union ControlNet is told what it follows (pose, depth, tile or lineart). A ControlNet not scanned yet goes by the folder it is in (SDXL_1.0, NoobAI…). When a block lacks a model, Anomalous reads ComfyUI's file lists anew once before saying so: ComfyUI does not see a model just downloaded into a subfolder until it restarts.
    • The upscales take the least enlarging upscale model that reaches their size (a 2× one for 1.25×, less work than a 4× one scaled back down).
    • A way is built for the canvas's model: on an Anima model, Follow a layout and Follow a pose use Anima's ControlNet-LLLite, and Repaint: crop and stitch adds Anima's repaint LLLite; the AI still only names the way. Needs ComfyUI-Anima-LLLite and anima-lllite models in models/controlnet. Anima's LLLite steers more gently than an SDXL ControlNet, so the prompt has to agree with the picture, and the AI is told so.
    • With several samplers on the canvas the AI has to say which one; nothing is guessed.
  • A Ways panel on the right of the AI workbench lists every way by job, what it does and where it came from. Ways missing a node pack or a model are greyed with what they lack; a missing pack opens ComfyUI-Manager searched for it. Each has a switch: one turned off is not offered to the AI, and building it is refused. When it begins, the AI gets only the ways that are on and can run here, with one line on what installing which pack would add.

  • The Ways panel says how each way was tried: the card (so far an RTX 4050 laptop, 6 GB), the size, a whole run's time; its settings are for that card (the upscale stops at 1.25×). The AI hears it too when it uses one, so it can say how long to wait.

  • Anomalous keeps a list of the models the ways use (api/block_models.json): each one's links (Hugging Face, Civitai, GitHub), size, SHA256, licence and folder. Each way names the models it recommends; the panel ticks the ones here (renamed ones too) and downloads a missing one with Anomalous's downloader in a click, checking its hash. Someone given a way gets the very same files from the list.

  • Colouring a line drawing uses Anima's lineart LLLite on an Anima model.

  • Each way is now a data file of its own (api/blocks/): its nodes, the packs and models it needs, what the AI is told and the panel's words in both languages, so a way collected later is one more file.