Skye DeskBeta

本手冊目前提供英文和簡體中文版本。 英文版 · 簡體中文版

說明文件

Skye Desk manual

Everything Skye Desk does and how to use it – from shooting your photos to using the finished LoRA. Each feature shows its status; only features in the current build have steps.

01

入門

This chapter gets Skye Desk open on your Mac and explains projects and the workspace.

Download and first open

The first public build is not out yet; join the waitlist to get one email when it is. When it is out:

  1. Download Skye Desk from the Download page.
  2. Drag Skye Desk into Applications.
  3. Open it. If macOS says it cannot check the app, open System Settings › Privacy & Security.
  4. Scroll to Security and click Open Anyway next to Skye Desk, then confirm.
  5. From then on Skye Desk opens normally.

Until the build is signed by Apple, macOS asks once. For Apple silicon Macs; system requirements will be confirmed with the first build; Intel Macs are not tested yet.

Workspace and projects

A workspace is a folder on your Mac that holds your projects. A project is one subject – one product, one character, one style – with its images, captions and training runs. Everything stays as ordinary files in that folder.

A first walk-through

  1. Create a project and give it a trigger word (a made-up word such as prodshoot).
  2. Drop 15–30 photos on the Dataset page.
  3. Write a caption for each image, starting with the trigger word.
  4. Export a training ZIP, or try a run in demo mode.
  5. Compare the samples in a blind test.

02

Preparing your images

This chapter covers the photos a LoRA learns from: shooting them, choosing them and importing them.

Shoot with your phone

Most people do not need AI pictures to start – 20 to 30 good phone photos of your subject are enough.

How many photos

Product, Character15–30
服飾20–40
食物15–25
風格30 or more
A real person (only with their consent)20–40
Effect, Material, Interior, Packaging20–30

Vary everything except the subject

  • Angles: front, three-quarter, side, back, from above, and a few close-ups.
  • Backgrounds: change them 3–5 times. One background in every photo gets learnt as part of the subject.
  • Light: daylight by a window, shade, a lamp – not all the same.
  • Distance: wide, medium and detail shots.

Phone settings

  • Main camera at 1×; not the ultra-wide lens (it bends shapes).
  • Portrait mode off, filters off, beauty effects off.
  • Tap to focus on the subject; wipe the lens first.
  • Keep the highest resolution; HEIC is fine.

Avoid

  • Other brands’ logos and any text you do not want learnt.
  • Fingers over the subject; your reflection in shiny surfaces.
  • Near-identical shots – ten photos from the same spot teach less than three different ones.

Tips by kind

產品
Turn it a little between shots; include the logo side, the back and the base; one close-up of any engraving or label.
服飾
Use a dress form or a hanger, then a few worn shots (face out of frame is fine); include cuffs, collar, buttons and the fabric up close.
Character / figure
Front, three-quarter, side, back; the face and any markings close up; keep the same lighting.
食物
Shoot within minutes, before it changes; top-down, 45° and side; one shot cut open if the inside matters.
室內
Corners and straight-on walls at eye height; daylight and lamp light; avoid people walking through.
風格
Many different subjects in the same style – the subjects should vary, the style should not.
A real person
Written consent first. Varied expressions, light and clothes; no other people in frame.

Get them onto your Mac

  1. Select the photos on your iPhone and share with AirDrop to your Mac.
  2. They land in Downloads.
  3. Drag them onto the Dataset page.

A shooting checklist inside the app, which shows what is still missing after import, is 規劃中.

Use photos you already have

Pick sharp photos of the subject from different occasions; drop blurry ones, heavy filters and screenshots. 15–30 good photos beat 100 similar ones.

Use AI-generated or reference images

You can also build a dataset from pictures you generated, for example from a design sheet. Check each one for drift – a changed logo or pattern is learnt too.

Working with reference images directly, without training, is the Reference Kit: 規劃中.

Importing

如何匯入圖片？ 現已可用

  1. 先準備 15 到 30 張同一主體、不同角度的圖片。
  2. 按「匯入」，或把圖片、資料夾或 ZIP 拖進來。
  3. 搜尋檔名或字幕，找出任何圖片。
  4. 打開資料集總覽，找出缺口與失衡。
  5. 在右側面板輸入字幕（描述圖片的文字）。

Supported: PNG, JPG, WebP, BMP, HEIC and TIFF images, folders, ZIP files, and same-name TXT or JSON captions. iPhone HEIC photos import as they are.

在任何地方拖放或貼上 下一個測試版推出

拖放到格狀檢視、側邊欄中的專案、專案頁面或首頁，或按 ⌘V 貼上。按住 ⌥ 可改為移動而非複製。複製前就會找出重複的檔案，匯入也可以復原。

Not in the current build yet – no steps until it is.

03

Organising your dataset

This chapter keeps the dataset tidy so the LoRA learns the right things.

按 Space 預覽 現已可用

  1. Select an image and press Space.
  2. Use ← 和 → to move between images.
  3. Press Space again to close.

標出缺少的字幕 現已可用

  1. Look for the marked images in the grid.
  2. Choose the Missing captions filter to see only those.
  3. Write or import the captions.

用起來像 Finder 的資料集 下一個測試版推出

按 Return 重新命名，按 ⌘⌫ 移到垃圾桶（可復原），並使用和 Finder 相同的顏色標籤。你在 Finder 裡做的變更會立即顯示。

Not in the current build yet – no steps until it is.

批次重新命名 下一個測試版推出

選一個命名規則，套用前先看到每個新名稱。字幕也會跟著改名，一次復原就能還原整批。

Not in the current build yet – no steps until it is.

資料集健檢 下一個測試版推出

缺少觸發詞、太小無法訓練、比例奇怪或與其他圖幾乎相同的圖片都會被標出，旁邊附上修正方式。近似重複的圖可以用在你 Mac 上執行的模型依內容找出（經你同意才會下載）；圖片絕不會上傳。

Not in the current build yet – no steps until it is.

從首頁搜尋所有專案 下一個測試版推出

依字幕、檔名或提示詞搜尋所有專案。每個專案都會顯示下一步。

Not in the current build yet – no steps until it is.

04

Writing captions

This chapter explains what to write – and what to leave out.

Leave out what you want the LoRA to learn. Write what should stay changeable. Start every caption with your trigger word.

各類主體的字幕技巧

角色
要描述： 姿勢、視角（正面、側面）、背景、光線、站在什麼上面。
要省略： 角色本身的外觀：體型、顏色、臉、記號，讓觸發詞去承載它們。
產品
要描述： 角度、背景、表面、光線、旁邊的道具。
要省略： 產品的形狀、Logo 位置、圖案與材質。
服飾
要描述： 呈現方式（人台、衣架、平拍）、角度、背景。
要省略： 版型、布料、車線、鈕扣、顏色。
風格
要描述： 每張圖的主題（一棵樹、一間房子、一艘船）。
要省略： 風格本身：紙張層次、柔和陰影、白上加白。
效果
要描述： 底下的物件（松果、鑰匙）及其場景。
要省略： 效果本身：霜晶、光暈、磨損。
材質
要描述： 材質被做成什麼形狀、如何打光。
要省略： 材質的紋理與顏色。
室內
要描述： 空間配置、家具、相機位置、時段。
要省略： 牆面處理、地板、配色、光線質感。
食物
要描述： 盤子、桌面、角度、是否切開。
要省略： 料理本身：外皮、水果擺法、顏色。
包裝
要描述： 周圍有什麼、表面、角度、光線。
要省略： 瓶子或盒子的形狀、標籤位置、蓋子與顏色。

The trigger word

A short made-up word that calls up your subject, such as prodshoot. Spell it the same way in every caption.

觸發詞拼法檢查 下一個測試版推出

觸發詞（用來叫出你主體的詞）的不同拼法，會一次找出並修正。

Not in the current build yet – no steps until it is.

如何批次修正字幕？ 現已可用

  1. 打開字幕工具 › 尋找與取代。
  2. 輸入要尋找的文字。
  3. 輸入要改成的文字。
  4. 按「預覽」，逐列查看修改前後。
  5. 按「套用」。復原可還原整批。

查看標籤分布 現已可用

  1. Open Dataset overview.
  2. Read the share of images for each tag.
  3. Click a tag to see those images, and fix captions that lean one way.

如何從我自己的 AI 取得字幕？ 現已可用

  1. 匯出所有圖片，再請任何 AI 為每張圖寫一個 .txt。
  2. 選擇匯入 › 匯入字幕包，然後選取檔案。
  3. 確認「寫入至」顯示的是正確的專案。
  4. 檢查每一列：MATCH 會寫入，SKIP 不會。
  5. 按「寫入字幕」。

整個資料集的標籤面板 下一個測試版推出

查看每個標籤及其使用次數，再一次在多張圖片上新增、移除或重新命名，並可復原。

Not in the current build yet – no steps until it is.

逐則檢查 AI 字幕 下一個測試版推出

並排查看每則新舊字幕，只保留你勾選的。匯入前會先建立備份。

Not in the current build yet – no steps until it is.

05

Training

This chapter turns the dataset into a LoRA – and explains what is real and what is simulated today.

In the current build, training runs in demo mode: no GPU is rented and nothing is charged.

如何匯出訓練用 ZIP？ 現已可用

  1. 每張圖片都要有字幕；空白字幕會導致無法匯出。
  2. 打開「檔案」選單。
  3. 選擇「匯出訓練套件」，或按 ⌘⇧E.
  4. 解壓縮：01.png 會和 01.txt 放在一起。
  5. 用 kohya（另一款訓練工具）訓練？依 kohya 的要求，把檔案放進一個以重複次數命名的額外資料夾。

示範模式

如何開始一次訓練？ 現已可用

  1. 設定步數（學習的輪數）。
  2. 選擇雲端 GPU，也就是你在線上租用的高效能電腦。
  3. 確認時間與費用估算。
  4. 按「開始訓練」。在這個測試版中，訓練是模擬的，不會產生費用。

訓練預設集與檢查清單 下一個測試版推出

選擇 LoRA 種類即可取得現成設定，每次訓練前還有一份檢查清單。檢查清單提供建議，不打分數。

Not in the current build yet – no steps until it is.

在 RunPod 上一鍵訓練 測試期間推出

選擇最便宜或最快；Skye Desk 會挑一張可用的 GPU，並在租用任何東西之前顯示估算。

Not in the current build yet – no steps until it is.

花費上限與自動停止 測試期間推出

為每次訓練設定上限。訓練結束、閒置或達到上限時，GPU 就會停止。

Not in the current build yet – no steps until it is.

從中斷的地方接著跑 測試期間推出

如果發生這種情況，訓練會在新的 GPU 上從最後儲存的步數繼續。

Not in the current build yet – no steps until it is.

Connecting Runpod

Real training on Runpod is 測試期間推出. You will need a Runpod account with credit and an API key, pasted on the Connections page.

Simple and advanced modes

Simple mode asks for the kind, base model and budget; advanced mode shows every setting with an explanation. 規劃中

Using reference images

參考圖工作台（Reference Kit）：用參考圖取代訓練 規劃中

從設定圖建立一組參考圖、檢查它，再搭配支援參考圖的圖像工具使用，或讓 Skye Desk 比較兩種方式並建議其中一種。

Not in the current build yet – no steps until it is.

創作者模式

創作者模式 規劃中

一個圍繞角色與其系列打造的簡化工作區，給部落客、AI 網紅和短影音創作者使用。

Not in the current build yet – no steps until it is.

06

Generating and comparing

This chapter makes new pictures with your LoRA and helps you keep the best version.

如何把 ComfyUI 結果帶回來？ 現已可用

  1. 選一個圖片編號。
  2. 寫下它的提示詞與負面提示詞。
  3. 複製工作流程，在 ComfyUI 中執行。
  4. 將每個結果以圖片編號命名，再匯入。

如何比較檢查點？ 現已可用

  1. 從某次訓練的樣本開啟盲測。
  2. 在每一組 A/B 中選出較好的圖。
  3. 按「到此為止並查看結果」查看排名。
  4. 點錯了？復原上一票。

比較訓練並設定最終 LoRA 下一個測試版推出

在專案頁面並排比較兩次訓練，並把其中一個標記為最終 LoRA，同時顯示它的底模。

Not in the current build yet – no steps until it is.

從圖片中找出工作流程 下一個測試版推出

Skye Desk 會讀取圖片中儲存的 ComfyUI 工作流程，並依底模（你的 LoRA 所疊加的 AI 模型）分組。

Not in the current build yet – no steps until it is.

把提示詞當作字幕草稿 下一個測試版推出

把圖片中儲存的提示詞插入它的字幕，再加以編輯。

Not in the current build yet – no steps until it is.

把提示詞送到 ComfyUI 測試期間推出

一次把所有圖片編號送到 ComfyUI，並自動填入各自的提示詞。

Not in the current build yet – no steps until it is.

檢查每個結果是否走樣 規劃中

每張新圖都會和你的主體比較；走樣、出現多餘文字或 Logo 的圖片會被標出，由你決定保留或捨棄。

Not in the current build yet – no steps until it is.

工作流程頁面：放入圖片即可找出它的工作流程 規劃中

把圖片、JSON 檔或 ZIP 拖到工作流程頁面；每個工作流程都會被偵測、命名，並與你現有的工作流程比較。

Not in the current build yet – no steps until it is.

07

Exporting and using your LoRA

This chapter covers where the finished LoRA lives and how to use it.

Today you train with your own tool (AI Toolkit or OneTrainer) from the training ZIP; the LoRA file (.safetensors) appears in that tool’s output folder.

  1. Copy the .safetensors file into ComfyUI/models/loras.
  2. In ComfyUI, add a Load LoRA node after the model loader.
  3. Choose your file and start with a strength of 0.7–1.0.
  4. Put the trigger word at the start of your prompt.

Base model licences differ: some do not allow commercial use of what you make. Check the licence of the base model before you sell or publish.

比較訓練並設定最終 LoRA 下一個測試版推出

在專案頁面並排比較兩次訓練，並把其中一個標記為最終 LoRA，同時顯示它的底模。

Not in the current build yet – no steps until it is.

08

Keeping your files safe

This chapter explains where your files are and how changes can be undone.

  • Images and captions are ordinary files in your workspace folder; captions sit next to images as .txt files.
  • Bulk changes show a preview first and can be undone as a whole.
  • Back up the workspace folder like any other folder (Time Machine works).

09

Privacy and consent

This chapter covers what leaves your Mac and whose pictures you may train on.

  • Your images stay on your Mac. Nothing is uploaded unless you start a cloud run.
  • Train on a real person only with that person’s written consent, and never on minors.
  • Do not train on other people’s copyrighted work or likeness without permission.

隱私權政策

10

參考資料

Shortcuts, words and fixes in one place.

鍵盤快捷鍵

  • ⌘1–⌘8 由上到下開啟側邊欄的各個頁面。
  • ⌘K 找出任何頁面、動作或專案。
  • Space 預覽選取的圖片； ← 和 → 在圖片間切換。
  • ⌘, 開啟設定； ⌘⇧F 開啟字幕的尋找與取代。

詞彙表

LoRA
一個小型附加檔案，用來教 AI 圖像模型學會一個主體，例如你的角色、產品或風格。
字幕
描述一張圖片的文字。AI 會同時從圖片和字幕中學習。
標籤
字幕中的一小段詞語，例如「front view」或「white studio background」。
觸發詞
一個自創的詞，例如 wren，在生成新圖時用來叫出你的主體。
底模
你的 LoRA 所疊加的 AI 圖像模型。LoRA 只能搭配訓練時所用的底模系列。
檢查點
訓練途中儲存的一份 LoRA。比較檢查點可以看出哪一步的效果最好。
雲端 GPU
按小時在線上租用、用來執行訓練的高效能電腦。
ComfyUI
一款用 AI 模型生成圖片的免費 app。Skye Desk 負責規劃提示詞，ComfyUI 負責出圖。
AI Toolkit、OneTrainer
免費的訓練程式。Skye Desk 會準備好它們需要的檔案。

Troubleshooting

macOS says it cannot check the app
System Settings › Privacy & Security › Open Anyway (see chapter 01).
Export is blocked
Every image needs a caption; use the Missing captions filter.
Imported captions did not match
Caption files must share the image’s file name (01.png ↔ 01.txt).
The LoRA learnt the background
Re-shoot with 3–5 different backgrounds, or describe the background in each caption.

閱讀常見問題

圖片預覽