Generating Media
The Generate wizard is the one place to make something with AI. In the Writer view it writes pages and reports into the Library. In the Director view it makes media: pictures, textures and 3D models, and with the sequencer on, video takes, sound and voices. AI is part of Pro (or the 14-day trial).
Opening it
- The Library's ⋯ menu (More Library actions) → Generate, the first row. It offers what fits the view you're in: pages in the Writer, media in the Director.
- Media's + in the Director: choose Generate ("Make it with AI") next to From this device and From links.
- Add media in a page editor → Generate. What is made is also placed on that page.
- Editors that make one kind of thing, such as the 3D editor's sky, open the wizard straight on that kind.
On a desktop the wizard is a dialog; on a phone it is a full-height sheet. Closing it stops anything still running.
Choose
The first step is a card for each kind. A kind that can't run yet stays on the list and says why on its card:
- Part of Pro — AI needs Pro or the trial.
- Needs a model — nothing is set up for it. Set one up in Settings → AI (see Models & Connections).
- Turn on the sequencer in Labs — video takes, sound and voices need Settings → Labs → Sequencer.
- Open the script first — scene synopses are written into the script, so it has to be open.
In the Writer view
"Choose what to write. It's added to the Library."
| Kind | What it makes |
|---|---|
| Character | A character page, from what the story has. Give it a name and notes if you like. |
| Location | A location page, the same way. |
| Outline | A beat sheet, act by act. |
| Scene synopses | A one-line synopsis under each scene heading, for This scene or the Whole script. See AI Breakdown & Synopses. |
| Coverage report | How a reader might score this draft. See Reading & Coverage. |
| Breakdown report | What each scene needs, as notes tagged in the script. See AI Breakdown & Synopses. |
Pages are written from what the project already has: the script and its pages. When a page is done the wizard says "Added … to the Library", with Open page and Make another. These all use the model for Analysis under Which model does what.
In the Director view
"Choose what to make. It's added to Media."
| Kind | What it makes |
|---|---|
| Image | Concept art, a frame, a portrait. |
| Texture | A Surface or a Sky for a 3D set. |
| 3D model | A prop or a piece of set, in 3D. |
| Video take | A shot as video, one take at a time. Sequencer only. |
| Sound & music | An effect, an ambience or music. Sequencer only. |
| Voice | A line read in a character's voice. Sequencer only. |
With nothing set up to make pictures or sound yet, the wizard says so first and points you to Settings → AI. For video takes, see Timeline & Sequencer.
Describe
Write what you want in the box. Opened from a character, location or other page, the box is Using this page: the page supplies who or what it is, so "portrait, natural light" is enough.
Referring to things with @
Type @ in the box to pick one of the project's scripts, pages or media. Each one you pick shows as a chip at the top of the box, saying what will be sent along; remove one with its ×. Add media and Add 3D asset in the same row pick pictures, clips, sounds or models from the project without typing.
Not every model can take every attachment. A chip drawn dashed is sent by name only: the model can't take pictures, or it takes fewer than you added, or it can't take a clip and only the first frame is used. A note under the chips says which.
Make it
Make it says what the generator starts from: From words, From a picture, or for a video take From a clip. Auto decides from what you attached. Starting from a picture or a clip shows a Strength slider: low stays close to what you gave it, high follows the words more. If the step needs a picture or a clip you haven't added, the wizard says so ("Add a picture with @ or Add media to start from it.").
The model
The Model line shows which model will do it and whether it runs On this device or on a service. Change it in Settings → AI → Which model does what. A line under Generate says where your words go: "Generated on-device — nothing is uploaded", or that your prompt is sent to the provider with your API key. A hosted service is only ever used for pictures if you picked it there yourself.
LoRAs
When this device has LoRAs for the image model you're using, a LoRAs row shows them as chips. Switch one on to use it. A LoRA set on a page you @-mentioned is already on, marked "On because its page is in the prompt". LoRAs work with the on-device image models in the Mac app. See Models & Connections.
Progress
Press Generate. One progress bar shows how it is going: "Waiting in the queue: 2 ahead…" when other jobs are first, then "Making it…", the time elapsed and "about 3 min left" once it can estimate. Stop ends it. A picture made on this device can take a while.
Review
When it's done, the wizard shows what it made:
- Save keeps it. It is added to Media → Generated, or to Media, and this page when you started from a page. 3D models and skies go to 3D assets.
- Change the description lets it go and goes back to Describe, with your words as they were, to try again.
Video takes and recorded voices cost money each time, so they are kept as soon as they are made. For those the step shows Done, and Try again makes another beside it.
Everything you make is kept on your device and reaches collaborators directly, like any media you add. EpicDraft doesn't host it.
Training a LoRA
Teach the on-device FLUX.2 [klein] image models to draw one character, place or look from your own pictures.
- In Media, select pictures of the same subject and press Train a LoRA. Or, in the Mac app, open the Set LoRA menu on a page and choose Train a new LoRA…. It's there on characters, locations, and notes of a type with a look: cast, extras, props, wardrobe, makeup & hair, sets, set dressing, vehicles, animals, and special and visual effects.
- Give it a Name and a one-word Trigger word. The trigger word is added to each picture's caption and to your prompts when the LoRA is on.
- Pick 4 to 50 pictures. 10 or more of the same subject work best. Opened from a page, the page's own pictures (its lookbook and the pictures in it) are listed first. Search finds pictures by name or tag.
- Under Captions, tag what changes from picture to picture, like the place or the clothes. Each caption is the trigger word and the picture's tags, and tags added here stay on the picture.
- Choose Where to train:
- fal.ai — on your fal.ai key. The pictures are uploaded to fal, which deletes them after a day. The dialog shows the cost ("About $… at fal.ai").
- This Mac — the Mac app on Apple Silicon, with 32 GB of memory for the 4B model or 64 GB for the 9B. The pictures stay on this Mac. It takes hours, and the first run downloads the base model once (over 10 GB).
- Adjust Steps if you like (100 to 3,000), and press Train.
When it's ready the LoRA is in Settings → AI and in the LoRA chips. LoRAs are trained on FLUX.2 [klein] base and used with the faster klein the app draws with, so the likeness can come out looser.
Related
- Models & Connections — on-device image and 3D models, ComfyUI, fal.ai, Meshy and Tripo keys.
- Media & 3D Assets — where generated media lives.
- Storyboards — panels and frames.