Production — attach assets and generate images and videos

Production, on the left of your project, is the screen where you generate images and videos. Instead of producing many clips automatically, you choose everything yourself — bring in the assets and scenes you want, edit the prompt, pick the model and settings, and generate.

The screen has two panes. On the left you set the prompt, assets and options; on the right you see the result. A single row above the result gathers what you can do with it right now, such as [Edit result] and [Make a video from this image], and which result it is and which model made it are noted at the bottom left of the result. To see everything you've made side by side, click [Show results] at the top right — a results pane opens on the right, and when you close it again with [Hide results] or the ✕ at the top of the pane, that space goes back to showing the result large.

The unit of this screen is the job — one job = one result you're trying to make. When you click Generate, the prompt, assets and settings at that moment are saved as a job, and attempts pile up under it (→ section 4). The field under "Production" at the top of the screen is the name of the current job.

Project look & rules — generation standards for the whole project

Click [Project look & rules] in the settings row below the prompt to edit the project style and production rules in two tabs. New screenplay projects come pre-filled with a live-action look suited to contemporary Korean short-drama production, natural skin texture, a ban on subtitles, logos and watermarks, and a location-sound standard. You can edit every value, and an empty field sends no instruction to the generation model.

What you enter is used in later generations whether or not it's approved. [Approve this style] and [Approve these rules] mark the current version as the project's production standard. If you edit values after approving, they're marked Changed since approval, so you can see which version you used as the standard.

• The filled-in project style fields, plus the materials and global bans in the production rules, are included in images,

sequence sheets and videos alike.

• Sound policies such as ambience and subtitles are included only in videos, never sent for images.
• No background music is added to any video (since 2026-09-21 — previously this was a checkbox in the production

rules, and the ban sometimes dropped out when the prompt was long). The sounds generated are dialogue, location sound and sound effects. Music that plays inside the scene (a radio, club speakers, a performance) is staged as written if you put it in the prompt.

• The shooting feel (direction) is set by the Project default direction in Project look & rules and is included in every

image and video generation (since 2026-09-13 — previously you picked it each time you generated). If you want just one shot to look different, write it into the prompt.

• Up to 3 style moodboard images are attached automatically as style references. If the model's image limit isn't enough,

assets you picked and images you uploaded are kept first, and the moodboard is dropped first.

1. Choose image or video

Switch between Image / Video with the toggle just above the prompt field. Attached assets stay where they are — you can check with an image first and, if you like it, switch to video and generate.

2. Attach assets

The drawer on the left has five sections: Characters · Locations · Props · Scenes · Archive.

When you write a prompt or attach assets or frames, the current screen is saved automatically as a draft job. Even if you haven't clicked Generate yet, you can reopen it from the job list at the top; empty jobs aren't kept.

• Characters, locations, props — assets you've registered in the project. Click to attach, click again to remove.

You can attach several, and the order you attached them in is kept.

• The character list is per person. Expression and motion sheets and looks are chosen by expanding "Sheets & looks N"

under the character row — by default (clicking the row), the master sheet is attached. Looks show the physical state too, like "Look: hospital gown (bandages)." If your search term matches a sheet or look name, that character appears expanded. Only one image per character is attached — the master, a sheet or a look — and clicking another variant replaces the previous reference in place.

• Scenes — choose an episode to load only that episode's scenes. Click a scene's [Import] and the first line gets the

location and time, like 낮, @카페. (in English, DAY, @Cafe.), while ready character names in the action lines and dialogue speakers come in as asset calls, like @민수 (@Minsu). Unknown speakers stay as plain text and dialogue isn't changed. Text you've already written isn't deleted. Ready assets among the characters and locations in the scene are attached too, and you can remove the ones you don't need from the asset chips. Props registered by dragging a name in the script are attached the same way (only props with a master image appear as chips, and removing them works the same).

• Archive — attach images made in this project directly. Each image has [Reference] · [Start] · [End] buttons for

choosing where it goes (Start and End appear only for video).

Direction, which sets the shooting feel, isn't in this drawer but in Project look & rules (the button at the top of the screen) — because it's a property of the whole project, not of a single generation.

Assets without an image don't appear in the list. A character with only a name has nothing to show the model. How many were left out is noted below the list — generate an image on the planning screen and it appears in the list.

Attached assets gather as chips just below the settings row — under the row with the upload and frame buttons, so you can see what you've attached right there. Remove them one at a time with the chip's ✕. If nothing is attached, that row doesn't appear.

You can also upload files directly

Use [Upload image] below the prompt to upload PNG, JPG or WebP images (up to 20MB) and mp4 videos (up to 24MB). These are reference materials used only for this generation, without registering them in the project. Images become references, and videos become reference videos (for motion).

The [Start frame] · [End frame] buttons next to it open a picker ([End frame] appears only on models that accept it — not on Veo Fast or Veo Lite, → section 6). Veo Lite doesn't take a [Start frame] either, so neither button appears. There you can pick an image made in this project as is, or upload a file — no need to download and re-upload. The purpose is different from a reference — a reference means "keep this face," while a frame means "start (end) on this shot" (→ section 6). Once added, its name appears on the button; click again to remove it.

What you add lines up as chips next to [Start frame] · [End frame], one each. Remove them with the chip's ✕.

3. Write the prompt

The sentences you write go in as is. They aren't summarized or changed.

Along with them, the project context (title, genre, logline), the project look and production rules, what each attached asset is, and the shooting contract of the chosen direction are assembled into one long instruction. You can write in Korean — the server translates it into English just before generating.

If the prompt exceeds the length the model accepts, the project look, production rules and asset details are dropped first. If the result looks different from your settings, shorten your sentences or attach fewer assets.

For images, a separate exclusion instruction is sent to keep certain things out (watermarks, subtitles, finger errors and so on). Text inside the scene isn't blocked — if you write a sign or signboard in the prompt, it's attempted as is (though models rarely draw text accurately).

You split the shots

If you want the picture to change within one clip, split it yourself in the prompt.

Shot 1 — 민수가 창고 문을 연다. 어깨 너머 미디엄.
Shot 2 — 민수의 얼굴 클로즈업. 눈이 커진다.

In English:

Shot 1 — Minsu opens the warehouse door. Over-the-shoulder medium.
Shot 2 — Close-up on Minsu's face. His eyes widen.

Write Shot 1 · Shot 2 like this and both refining and generating keep that split. Write the label in English as Shot — it's the only form recognized for now.

When refining a single clip, the AI doesn't add shots for you. Refining only keeps the original number of shots — it doesn't add or merge them (→ Refine with AI). So text written as one sentence comes out as one shot. When the text is too long and is refined into several clips, though, cuts and angles are designed for each clip (→ When it's too long). If the result feels flat, choose [It's flat — no cuts] in the Edit result dialog, or split it yourself as above.

When you bring a scene from the script, the first split is already there. Send to production moves the scene over in the order it's written in the script and marks each place where the action line changes with Shot 1 · Shot 2 — the writer's "Minsu sits → his line → Jiyoung sits → her line" becomes two shots as is. If you don't like it, delete or move those lines to split it again. A scene with only one action line has nothing to split, so no marks are added. Parenthetical acting directions before dialogue, like (조심스럽게) ((carefully)), are carried over too. When you generate with a Seedance model, that direction is sent as a dash aside on the dialogue line instead of in parentheses (Seedance reads parentheses as a background-music cue).

Many scenes work fine as one shot. Splitting isn't always better.

Calling assets inside a sentence — @

There are three ways to call one — ① type @ while writing, ② the [@] button next to a row in the asset drawer on the left (inserted at the cursor), ③ drag an asset row into the prompt field (inserted where you drop it). Whichever way, the asset is attached too.

Type @ while writing and the asset list appears. Pick one and a short call like @민수 (@Minsu) goes into the sentence, shown in green text — if it's colored, the asset is recognized correctly; an uncolored @text is just text. The called asset is attached as a reference too (if it's already attached, it's left as is), and the details you saved when creating it are included.

• Characters — appearance (age range, face, hair, build, outfit), props, dominant hand, permanent features,

and items noted as "Do not change"

• Locations — spatial layout, fixed landmarks, entrances, continuity axis, time of day and weather, lighting and color tone
• Props — shape and size, condition in use, holder, surface text, wear and damage

This is different from just typing the name. Calling it sends the asset's image and details together as one, so the model knows who "Minsu" is and a different person appearing in each shot becomes less likely.

Only what you've saved is included. The more fields you fill in for characters, locations and props on the planning screen, the more gets included when you call them. Empty fields are left out, not made up.

An @ that isn't an asset (for example, @2배속, "@2x speed") is left alone. It's a call only when it matches a name in the list.

Refine with AI

[Refine with AI] below the prompt field refines the current text for image or video generation only when you click it. Importing scenes and editing directly remain free and never call the AI automatically.

• In video mode, it refines around the action and camera within the length you've set, and adds voice, expression and

gesture acting directions that fit the scene to each line of dialogue.

• Refining a single clip preserves the original — it tidies it into shooting directions without increasing the number of actions,

camera moves or shots. If the result adds shots, it's rewritten once, and if it still adds them, that part keeps the original text. The "Actions N→M · Shots N→M" badge above the result shows the before and after counts.

• Splitting shots is up to the writer. If you write Shot 1 · Shot 2 in the prompt, refining keeps that split

(→ section 3, "You split the shots"). If you don't, it comes out as one shot.

• If you've written production rules for the project (camera moves to avoid, no background music and so on), refining

reads those rules too — it doesn't add direction that conflicts with them. Those rules are never trimmed from the generation prompt, however long it gets (2026-10-13).

• On Seedance models, refining adds direction (2026-10-13): one lighting phrase per shot (a light source inside the

scene, its direction and colour temperature), one sound line paired with what's visible, one camera move written through to its end point if the text has none, emotions as one visible action, and a final phrase saying what's on screen at the end. Events, dialogue, characters and shot counts aren't increased. When splitting into several clips, a lock line at the end of each clip restates the light, each character's look and position, so light and looks drift less between cuts. Nothing is added on Veo, Gemini, MiniMax or Kling.

• Video refining is fitted to the prompt grammar of the model you've selected — the same text gives different results on

different models. The dialog names which model it's refining for, and if you changed the model after refining, it's best to refine again.

• In image mode, you choose the goal first.
• Single scene (still) — tidies it around the composition, character placement, expressions and lighting of one moment.
• Sequence sheet (several shots in one image) — the sheet is always 16:9, 6 shots (3×2) by default and 4 shots (2×2) for

short scenes, tidied into a grid prompt read left to right in order within one image. Dialogue isn't put in as text in the picture; it turns into the expressions, gestures and gaze of the speaker and the listener.

• The result doesn't overwrite right away. The dialog shows the current prompt and the refined suggestion side by side, and

it changes only when you click [Apply]. Click [Close] to discard it.

• In video, if the current text doesn't fit the chosen length, refining doesn't start and suggests a longer length

(for example, "This text doesn't fit in 4 seconds. Choose 6 seconds or longer and refine again."). No credits are used in this case — change the length in the settings row below and click again.

#### When it's too long, it proposes splitting into several clips

In video mode, if "Up to M sec (long dialogue may be cut off)" shows to the left of [Generate], the current text exceeds this model's single-clip length. Click [Refine with AI] in that state and it proposes the text split into several clips. How many clips depends on the text length and the model.

• When refining into several clips, cuts are designed for each clip. Each clip is split into Shot 1 · Shot 2 to fit

its length, and consecutive shots change shot size or angle (wide, medium, close-up, insert; over-the-shoulder, low or high angle and so on). Dialogue is cut between the speaker and the listener. The number of shots per clip follows its length — 2 shots for 5 sec, 3 for 8 sec, up to 4 from 10 sec. Events, dialogue and characters aren't added, and Shot boundaries written in the original are kept. That's why the shot-count badge above the result can grow without a warning color.

• If another model would need fewer clips than the current one, a line at the top of the dialog tells you

(for example, "With this model (Veo, up to 8 sec) it's 5 clips. Switch to Seedance (up to 30 sec) for 2 clips."). Change the model with Model in the settings row below.

• You check the proposal one clip at a time with numbered chips. The chevron between chips tells you how they relate

at that point; hover over the chevron to see which kind it is.

• Continuous shot — continuous action in the same space and time. It continues from the last frame of the previous clip.
• Cut — the location or time changes. It doesn't continue from the previous frame.
• Click [Apply] and a clip switcher row appears above the prompt field. Click the 1 2 3 chips to move between

clips; the screen shows only the prompt of the selected clip. Clips with a result get a check mark, and the undo (↺) at the front of the row turns the split back into one prompt.

• Asset calls (@) work in the selected clip's field. The length is set by Length in the settings row below, and that

value is the length of the currently selected clip.

• To go back to one clip, click [Revert to one clip] in the ⋯ menu on the right of the switcher row.

The split text isn't deleted; it's joined back into one.

• It doesn't stop because of the number of clips or a 90-second total. The server refines in small batches as many as

needed, and even if you close the dialog or reopen the tab, the same job picks up where it left off. Long dialogue and sentences are split automatically at boundaries that fit the model's single-clip length, keeping the speaker and the original text.

#### When generating split clips

• The estimated credits on the [Generate clip N] button are the cost of the one clip you've selected now.
• You can generate the clips in any order. It doesn't enforce an order or nag you about it.

For a continuous shot where you want to use the previous clip's final frame, click [Prepare clip N from the last frame] below. A cut doesn't carry over the previous frame, so there's no need.

• A submitted clip keeps generating on the server even if you close the screen, and the result stays in the archive. Clips

you haven't clicked don't start automatically.

• For a continuous shot in the same space and time, click [Prepare clip N from the last frame] above the result.

The last frame of the finished video goes into the next clip's start frame, and you move to the next clip's tab after the frame is ready. If preparation fails, a notification gives the reason and the tab doesn't move (clicking twice on the same video extracts only once). After checking the prompt and start frame, you click [Generate clip N] yourself.

• For a cut where the location or time changes, pick the clip tab yourself and generate a new shot with the start frame

left empty. It's fine even if you haven't generated the previous clip yet. There's no batch or simultaneous generation — you click one clip at a time.

• Clip tabs with a result get a check mark showing how far you've generated.
• Generated clips are kept separately in the results pane ([Show results]) and the archive (shown as "Clip 1/3").

This screen has no feature to join several videos into one — join them in your editing tool.

• For the single scene and video goals, the server checks again that the original @calls and Speaker: "line" lines

are kept as they are. For the sequence sheet goal, instead of the dialogue it checks that @calls are kept and that the sheet structure is intact. If these aren't kept, it's treated as a failure and your credits are refunded.

A sequence sheet prompt is for images. If you put a refined sheet prompt into video mode as is, the existing safeguard blocks generation. To make a video, pick one shot from the sheet and write a new video prompt describing what moves.

This is a conditional feature. It's available with a signed-in account in an environment where a writing AI key is configured, and costs 5 credits per use. If sign-in, the AI key, credits or permissions are missing, or the preservation check fails, the reason appears on screen and the current prompt isn't changed.

4. Jobs — one result you're trying to make

The unit of saving on this screen is the job. If you generate the same shot five times while editing the prompt, that's not five results but one job with five attempts.

• When you click Generate, the screen at that moment (prompt, attached assets, frames, settings) is saved as a job.

There's no separate save button. If you leave the name empty, it's named from the start of the prompt, and you can change it any time in the name field at the top of the screen.

• Finished attempts pile up under that job, and the results pane on the right, opened with [Show results], shows the

current job's attempts in a two-row grid. Click one to put it on the big screen.

• Reopen earlier jobs from [Job list] (next to the name field) — the screen and attempts from that time come back as a

whole. Delete is in each row's ⋯ menu (generated files stay in the archive).

• [Start new] (top right) starts a new job. The current job stays in the list exactly as it is at that moment, and the

screen and results pane are cleared. A one-line notification at the bottom of the screen tells you whether it started or was canceled.

• Jobs are saved to your account — sign in to continue on other devices.

When you reopen a job, even the attached assets come back. But if you deleted an asset (a character image and so on) in the meantime, that item opens missing.

Image prompts and video prompts are written separately. The Image tab in Production is where you describe how one moment looks, and the Video tab is where you describe what moves and how. Each field keeps its contents when you switch tabs, and the one actually used for generation is recorded in the result version.

5. Set the specs — one row below the prompt

Right below the prompt is a row of small controls. Only values the selected model actually accepts appear.

Start frame · End frame and attachment chips · Model · Aspect ratio · Quality · Length (video) or Quality (image)

Sound isn't in this row but at the bottom right inside the prompt field.

• When you change the model, the other lists change with it, and if a chosen value doesn't exist on the new model, it moves

automatically to the closest value.

• If you bring a scene into video mode without setting a length, a value calculated from the amount of dialogue and number

of actions is filled in as the default. After that, the screen doesn't nag you about length — if it's too long the model trims it on its own, so it's better to decide how much to fit in one clip by generating a few times.

• The default aspect ratio for video is 9:16 (vertical). A value you picked yourself is kept, and on models that don't

accept 9:16, that model's first value is used.

• Direction isn't in this row — you set it once in Default project direction under Project look & rules

(natural contemporary drama, observational home video, vintage 16mm and so on). Choose only one: the shooting medium, camera movement and things to avoid are bundled as one set, so combining two gives conflicting instructions. The default, Project default direction, adds no shooting instructions.

• In projects with an anime look, you see animation presets instead of shooting presets (TV anime limited,

theatrical anime, action animation). Live-action shooting presets use camera vocabulary such as skin texture and camera shake, which pulls the drawing toward live action. A preset you already chose stays in the list even if you change the look.

• On some models, the quality fixes the length — in that case, Fixed at 8 sec at this quality appears at the end of the

row.

• The sound switch is in the strip at the bottom inside the prompt field (With sound / Without sound).

On models where it can't be turned off, it's locked as Sound always on. Even when on, no background music is added — only dialogue, location sound and sound effects are generated.

Check the estimated credits before generating

Once you've set the specs, the estimated credits calculated by the server appear inside the [Generate] button. The value reflects not only the model, quality, length, resolution and sound, but also the number of references attached, start/end frames and reference video length.

• Server estimate — calculated from the operating price table.
• Base price estimate — the model's actual cost hasn't been registered yet, so the server's default feature price is

applied. The credits shown on the button become the spending cap as is.

• While Checking estimated credits or an estimate error is shown, [Generate] is locked. This prevents clicking at the

previous specs' price right after changing the specs.

• For video in an anime project, if the start or end pose (key frame) is still empty, the frame row also shows the

total including key frames, such as 2 key frames 20 + video 100 = 120 credits in total. Key frames are made separately in image mode, so the number inside the button stays the video amount you're charged when you click. Attaching a key frame removes it from the total, and the row disappears once both are attached.

• The end pose is made from the start key frame. Once a start pose is set, the frame row shows [Make end pose].

It switches to image mode with the start key frame attached as a "composition anchor" reference and fills the prompt with a template: "exactly the same camera, framing, background as the start key frame; only one thing changes: (…)". Write the one change in the brackets and generate; under the result, [Use as end pose] drops the image straight into the end-pose slot of video mode. Generating an end pose from a prompt alone changes the camera and background, and the clip then jumps at the end. The anchor stays ahead of character sheets at the reference cap — it is not dropped even with three characters.

• If you have not chosen a video model in an anime project, clips are made with Seedance (when the operator

policy exposes it). Veo sometimes rejects key frames of anime characters who look underage, and Veo Lite/Fast do not accept an end pose. A model you chose yourself stays as chosen. If you have not chosen a resolution either, it is 720p (line art smears at 480p). A resolution you chose yourself stays as chosen.

• In an anime project, importing a script scene as an image drafts only the first shot (the first action

paragraph and its lines) as one moment, with a "one single frame" line appended. Putting the whole scene's action into one image makes the image model draw a strip of panels, one per beat, which cannot be used as a key frame. To make a key frame of another moment, rewrite the draft to that sentence. The video draft still carries every shot.

The moment you click the button, the server checks once more that the amount is the same. If the operating price or generation conditions changed in the meantime and the amount differs, it doesn't hold credits or start generating, and asks you to check the new estimate.

Reference image limit: each model accepts a different number of images. If what you've attached exceeds the limit, it keeps characters → locations → props in that order, and the rest isn't included in this generation. How many will be left out is shown on screen before you generate.

Reference videos — make a video by looking at a video

On models that accept reference videos, you can make a new video that references another video's motion and direction. Pick an mp4 file with [Upload image] and it becomes a reference video.

• 2–15 seconds each, up to 3, and up to 15 seconds in total (per the model documentation). Files outside the spec are

returned with the reason — the length is measured before uploading, so you won't upload a large file and wait.

• Files whose length can't be read can't be used (the spec is in seconds, so it can't be met without knowing the length).
• An uploaded video appears as a Reference video 3.4 sec chip next to [Start frame] · [End frame]; remove it with ✕.
• The rights to an uploaded video belong to the person who uploaded it. When using someone else's video, check the rights.

Some models can't use reference videos together with start and end frames. On those models, frames take priority and the reference video isn't included in this generation — the screen tells you in advance.

5-1. Import from another project (archive)

Use [Import from another project] above the archive to open the archive of another of your projects and copy selected items into this project. It's for reusing images and videos you've already made between projects split into seasons or set in the same world.

• It's a copy — nothing changes in the original project, and what you imported stays even if you delete the original project.
• Where it came from (project and asset) is kept as a record.
• It works only between projects in your own account. It costs no credits.

6. Continue from the previous scene

Click [Continue from this ending] on a generated video card and that video's last frame goes into the start frame slot. You can make a video that continues from the shot where the previous scene ended. Videos in the archive have the same item in the ⋯ menu.

• It costs no credits. Extracting a frame isn't AI generation.
• Nothing is generated right away. Only the start frame is filled in; editing the prompt and clicking [Generate] is still up to you.
• Clicking several times on the same video creates the frame only once.

A start frame is different from a reference. A reference means "keep this character's face," so the background and composition are deliberately ignored. A start frame, on the contrary, means continue from that shot. That's why they have separate slots.

Some models can't use both together. On those models, the moment you add a start frame, the screen tells you in advance how many attached references will be dropped.

>

Even if an image is dropped, that character doesn't disappear from the prompt. The appearance description saved in the asset (age, hair, fixed items) is included instead, along with a note that "the person seen in the first frame is this character." Still, it's not as good as an image, so for shots where the face matters, it's better to leave the start frame empty and generate with references.

If the aspect ratios differ, you're told in advance. If you put a 1:1 image in as the start frame of a 9:16 video, the model crops or stretches it. This isn't blocked — sometimes it's done on purpose, so you're only told the fact and the writer decides.

You can add an end frame too — give a start and an end and it fills in between them (first/last frame interpolation). The slot appears only on models that accept it, and some models only accept it together with a start frame, in which case adding only an end frame blocks generation (the reason is shown next to the button).

End frames are accepted by Veo 3.1 (full) and the Seedance family. On Veo Fast and Veo Lite, the [End frame] button doesn't appear — Google rejects Fast when a start and an end are sent together (confirmed 2026-09-08), and Lite hasn't been confirmed, so it's closed as well. If you need first/last interpolation, switch the model to Veo 3.1 or Seedance. Start frames work on all three models. Gemini Omni also accepts end frames, but only together with a start frame — adding only an end frame blocks generation, and the reason is shown next to the button.

You can also add start and end frames directly with the [Start frame] · [End frame] buttons in the settings row, by picking an archive image or uploading a file (→ section 2).

Extend by 7 seconds — Veo 3.1 and Veo 3.1 Fast only

Below a video made with Veo, next to [Continue from this ending], there's one more button: [Extend by 7 sec] Veo only. It's different from the last-frame approach — instead of a single still image, it sends the video file itself to Veo and gets back one file with 7 seconds added to the end. Motion and sound continue without a break.

• It works only on 720p videos made with Veo 3.1 or Veo 3.1 Fast (Google's official contract). For videos made with

Lite, Seedance videos or 1080p videos, clicking the button shows the reason in a notification and the original isn't added — use [Continue from this ending] in those cases.

• Clicking it adds an "Extend ○○ · Veo only" chip to the settings row and locks quality and length to 720p · +7 sec.

The start and end frame buttons are hidden (extending and frames can't be used together). Click the chip to remove it.

• In the prompt, write only the action for the 7 seconds that follow. There's no need to repeat what's in the previous video.
• Character and location reference images aren't included in this generation (the saved detail text remains).

If you switch the model to Lite or another, generation is blocked and the reason is shown next to the button.

• The cost is the new 7-second segment × that model's per-second price. The result goes into the archive as a new video,

and you can click [Extend by 7 sec] again to make it longer (up to 148 seconds per video).

• If there's no voice in the last 1 second of the original, the voice may not continue (per Google).

Preview with Draft, then make the 1080p final — Seedance 2.5 Draft

Seedance 2.5 Draft (Draft lane) in the model list makes Seedance 2.5 quickly at 480p so you can check framing, blocking and performance first. It goes up to 30 seconds like 2.5, and 480p is the only quality.

A video made with Draft gets a [1080p final] button above it. Clicking it makes the same scene, motion and sound again in 1080p — the provider reuses the draft's settings (including the seed), so the result is consistent with the draft. An identical video isn't guaranteed.

• Clicking it first shows the estimated credits; nothing starts until you click [Make].
• The prompt, characters, references, length and aspect ratio are reused from the draft. They can't be changed — make a

new draft if you want to change them.

• A draft older than 7 days can't be turned into a final (provider rule). The confirmation shows the deadline.
• It's charged twice, separately: the draft (480p, about 33 credits for 5 sec) and the final (Seedance 2.5 1080p, about

365 credits for 5 sec). The final is in the Final lane, so it uses credits you have topped up.

• The result goes into the results list and the archive as a new video. The draft video stays as it is.

When you switch to generating images, frames take a break. Start and end frames exist only for video, so in image mode the slots disappear and only a notice appears that they won't be included in this generation. What you picked isn't deleted — it's still there when you return to video. The same applies when you switch to a video model that doesn't accept frames (generation isn't blocked).

7. Generate

Click [Generate] to start. You can go to other screens while it's generating.

Video models have a limit on how many can be generated at once per account, so when things are busy it shows "Queue position N" and starts automatically when its turn comes.

If the button can't be clicked, the reason is shown next to it — the prompt is empty, or the selected model requires a start image.

If there's at least one reference or start frame, you can generate with an empty prompt. But without a sentence describing what moves and how, the result may look like a still image.

If you've attached character assets to a Seedance-family model, the first [Generate] waits about 15 seconds for that character's asset registration to finish, then continues generating. Only if it still isn't ready does it tell you "Please generate again in a moment" and stop (no credits are charged). Registration starts the moment you attach a character asset, so it usually starts without waiting.

Start and last frames go through the same registration. Seedance doesn't accept photo-like images containing people as is, so images with a large face in them, like the last frame of a previous shot, used to be rejected. Now frames are also registered as assets automatically before being sent — there's nothing for the writer to do. If registration is slow or rejected in review, the original is sent as is, and in that case it may fail by being judged a real person.

However, Seedance can't use frames and character assets together (a rule on the model's side). If you attach a start frame, character assets are dropped and the person in the frame takes their place. If you need a character or location that isn't in the frame, remove the start frame.

This row is pinned to the bottom of the screen, so it's never hidden no matter how large the result is.

Start a new job — [Start new] at the top right

[Start new] at the top right of the screen starts a new job. The current job stays in the job list with the prompt and assets exactly as they were at that moment, and the prompt and attached assets and frames on screen are cleared. It asks once before clearing, then tells you "Started a new job" if you went ahead or "Didn't start" if you canceled.

• A new job is still an empty draft, so it doesn't appear in the list — it appears in the list once you write a prompt or

attach an asset. Renaming a job doesn't create a new job; it renames the current one.

• Images and videos share one job (the mode buttons don't split jobs). The results pane shows only the current mode's

attempts, while the job list collects attempts from both modes.

• Generated results aren't deleted. Reopen an earlier job from the list and its attempts come back, and the files are

always in the archive.

• If there's nothing to clear, the button can't be clicked.
• Its role is different from [Generate] at the bottom — the bottom one means "generate with what's here now," while the top

one means "set this job aside and start a new one."

8. View results — to the right of the prompt

The work area has two panes — the prompt and specs on the left, and the whole right side is where you view results. The results area fills the screen except for the window's top and bottom margins, and is always fully visible without scrolling. The row above the result holds what you can do with the current result ([Edit result] · [Make a video from this image] · [Continue from this ending]) and [Show results], and which result it is, whether it's an image or video, and which model made it are noted in one line at the bottom left of the result.

• While generating — it spins with "Generating video." You can go to other screens, and when you come back the in-progress

indicator is still there. In the results pane opened with [Show results], in-progress attempts appear first as "Generating" placeholders.

• If you generate several in a row — the result that finishes first goes on stage and can be played right away.

While the rest are running, a line "Generating video · What you see now is the result that finished first" appears at the top of the stage, and it switches to the new results as they finish.

• When finished — it goes up right there. Videos have a playback bar so you can watch them in place.

Click an image to view it at full size. Tall results aren't cropped either; margins appear on both sides and the whole image is visible.

• If it fails — it says what the problem is and what to do next (the provider's original message remains in small text only

when the cause couldn't be determined). Click [Close], start a new generation, or pick a result from the results list to clear it. Charged credits are refunded automatically.

• For a temporary outage on the model provider's side (for example, "internal server issue"), the server **resubmits once on

your behalf after about a minute**. Meanwhile, the status line reads "Temporary error on the model's side — resubmitting shortly," and if the second attempt also fails, it reports the failure then. If failures from the same provider become frequent, a line "○○ is unstable right now" appears next to the model selector in the settings row — generate with another model or try again later. It never secretly switches to another model.

• If Veo reports that "it couldn't complete during the sound generation step" — this is an intermittent error on Google's

side (listed in the official documentation's limitations as "safety filters or other processing issues with the audio"). It isn't a content policy rejection, so trying again with the same conditions often works, and credits are refunded automatically. If it repeats, cut down the speech and sound descriptions in the action text or pick another model.

• When you attach start or end images, Veo and the Seedance family don't accept character or location reference images at

the same time (Gemini Omni does). A notice appears below the specs, and those references are also removed from the prompt (the saved character and location detail text is still included).

• Use [Continue from this ending] on a video to move on to the next scene (→ section 6).
• Click [Make a video from this image] on an image to switch to video mode with that image in the start frame.
• If you haven't generated anything yet, these buttons stay in place but locked — they turn on once a result is up.
• In the archive (⋯ menu), you choose where to put it — video reference / start frame / end frame. The same image serves

different purposes: a reference refers only to the character's appearance and deliberately ignores background and composition, while a frame, on the contrary, starts or ends on that shot. You can choose the same three in the Archive tab of the left drawer too.

• Either way, nothing is generated right away — check the prompt and click [Generate] yourself.

Edit result — fix what fell short and generate again

Click [Fix result] below the stage and the "What didn't work in the result?" dialog opens. It works for both images and videos.

• Choose what fell short. Videos have seven options (character, key action, camera, no cuts, motion, dialogue and sound,

broken shapes), and images have four (character, scene, composition, broken shapes). Images have no motion or dialogue options — those aren't things you can ask of a still picture.

• [It's flat — no cuts] asks for the shots to be split so it doesn't end as a single shot. You can choose it together with

[Too much camera] — one reduces camera movement and the other adds cuts, so they don't conflict. The result is two shots with a static camera.

• You can also write it yourself. Write requests that aren't on the list in the [Write your own] field below

(up to 300 characters). For example, "less blinking from the hazard lights" or "remove the sign text." What you write is placed before the chosen items — what a person actually asked for comes first.

• Click [Generate again like this] and it adds the chosen items and what you wrote to the prompt that made that result

and generates right away. Even after making the same clip several times, fixing an older result starts from that result's text, not what's in the input box now (results made from 2026-10-21 on; for older results it's added to the text in the input box now, and the dialog says so). It's a new generation, so credits are charged — the amount is on the button.

• The fix request goes only into this generation. After it's made, the prompt field keeps **only what you originally

wrote**. You can see the request you wrote for that result under [Prompt] above the result.

• The model, length, aspect ratio, quality and attached assets use the current settings as is. No need to choose them again.
• When [Generate] is blocked (checking the estimate, a model settings problem, the clips need re-splitting and so on), this

button is blocked for the same reason, and the reason is shown to the left of the button.

• For a video split into several clips, opening the dialog moves the screen to that result's clip, and that clip is

generated again.

• When you reopen it, the items you chose and the text you wrote last time are still there. Each generation replaces the

previous note, so instructions don't pile up.

• If you don't change anything, the button can't be clicked. This is so no credits are spent generating the exact same thing

again. If you want to generate once more with the same conditions, close the dialog and click [Generate].

• If you want to edit the prompt by hand, close the dialog and edit it in the prompt field.

See a result's prompt — reuse the text that made it

Click [Prompt] in the row above the result to see the prompt used to make the result on stage. The input box only remembers the last text for each clip, so use this when you've made a clip several times and want an older result's text back.

• Click [Bring into input box] to put that text into the result's clip input box. Edit it, then click [Generate].

It only brings the text in and doesn't generate, so no credits are charged.

• Results made before 2026-10-21 don't have the full text saved, so only the scene line is shown. The dialog says so.
• Results with no prompt left say "No prompt is left for this result." The button can't be clicked for results that

aren't saved to the archive yet.

The pane on the right collects the current job's attempts as small thumbnails. Click one to bring it on stage. Everything you make also collects in the archive — it stays even if you leave the screen or switch jobs.

You can also see which model made it. The line below the stage, the captions in the results list and the archive cards carry the model name, like "Veo 3.1 Fast." Hover over the name to see the provider and creation time and, if saved, the length, resolution, aspect ratio and seed. Results made without choosing a model directly, and older files, may have no name — only what's saved is shown, and nothing unknown is filled in.

The model name is a record of what it was made with. Generating again with the same model, prompt and seed can still give a different result (providers update their models). That's why the creation time is shown too.

The archive is a dense grid (6 per row on wide screens). It's a screen for choosing, so thumbnails are shown small and cropped. Click one and a popup opens — images at full size, videos playing in place.

When you hide something, you're asked why once. Click archive ⋯ → Hide and a dialog asks what fell short (face, outfit, dialogue, lighting, camera, space, other). To skip, press [Hide without a reason]. The answer is used to decide which part to improve first. The file isn't deleted.

The stage and the right pane show only the current mode's items. Switch from image to video and the results area empties; switch back and they reappear. The prompt and attached assets carry over as they are.

The [Generate] row is pinned to the bottom of the screen, so you can click it any time regardless of scrolling.

Start from the script — Send to production on an episode

Instead of building a prompt on a blank screen, you can start from the script.

Where it is: [Send to production] at the top right of the episode (script) screen — right next to "AI help." Click it to open the "Take scenes to production" dialog.

The old step tabs and per-episode video screen have been removed (2026-08-21). Download finished clips from the archive and join them in the editing tool you use (CapCut, Premiere and so on).

Send scenes to production

When the dialog opens, the episode's scene list appears. Each row has the scene number and scene name, plus one [Send to production] button. Sending itself costs no credits.

The image field and the video field are filled together. Whether you'll make a still or a video gets decided as you try things, so you don't have to pick one side when importing. What you choose in [Screen to open after sending] at the top of the dialog is only which screen to open; switch the mode tab on the production screen and the other side's draft is already there. No need to import again.

1. In image mode, you can choose single scene or sequence sheet in [Refine with AI], and scenes with dialogue have sequence

sheet selected first. AI refining tells you it costs 5 credits before it runs.

1. In video mode, check the action and dialogue, then choose specs such as the video model, length and sound.
2. The draft differs by mode. Action lines are the same in both modes, but in the video field dialogue goes in as

Speaker: "line" lines, while in the image field only "who is speaking" goes in, instead of the original dialogue — if you put quoted dialogue in a still, the model draws speech bubbles or subtitles. The characters and locations in the scene and the prop assets attached to that scene are shared by both modes, so remove assets you don't need with the chip's ✕.

1. After checking the prompt, model, specs and estimated credits, click [Generate].

If you're already on the production screen, you can bring in the same draft from the Scenes section on the left. That path adds text at the current cursor position and keeps the prompt you've already written and the assets you've picked.

If you want to join several clips, generate one, then click [Continue from this ending] on the result to use it as the next clip's start frame. Check each clip's prompt and assets and generate it yourself.