Prompt Guide

A reference for every piece of syntax a prompt can carry, with examples you can paste straight in.

A single run of Omninator works through a whole list of prompts against Google Flow — images or videos. Beyond plain text, a prompt can pull in reference images, point back at an earlier result with @tag, hand a fresh result a name via #name for later reuse, and, on the video side, pin down each clip's frames with @start / @end.

Images and Videos are separate tabs. Each keeps its own prompt text, so setting one up never touches the other. Sections 1–5 apply to both; section 6 is video-only.

Basics Reference images @tag — reuse an image #name — name an image Chaining (consistency) Video prompts Precedence

1. The basics

Write or upload one prompt per block, with a blank line between them — each block turns into one generation (or several, when outputs-per-prompt is set above 1).

a cozy cabin in a snowy forest, warm light in the windows

a neon city street at night, rain reflections, cinematic

a studio portrait of an astronaut, soft key light

Import a .txt file or type directly, set your model / aspect ratio / quality once, then hit Start. Results appear as they finish and save automatically into your chosen folder.

2. Reference images

The Reference Images panel holds three slots — Subject (who or what), Scene (where), and Style (look and mood) — and whatever you put there applies to every prompt in the run.

5 per slot, capped at 10 total across all three combined.
Reach for these when the whole batch should share one look or one set of characters.

3. `@tag` — pull in an existing image by name

Drop @name or @[two words] into a prompt and that image becomes a reference for just that one prompt:

@lina standing in a misty forest, golden hour, wide shot

@name works for single words; wrap multi-word names in brackets, e.g. @[two words].
A file you uploaded as lina.png resolves as @lina — the extension is optional, so @[lina.png] also matches.
If a name doesn't match anything, the tag is left untouched as ordinary text in the prompt.

4. `#name` — label the image a prompt produces

Attach #name to a prompt and its result gets that label. Nothing about #name shows up in the picture itself — it exists purely so a later prompt can pull the result back in with @name.

a young woman, red hair, freckles, studio portrait #lina

Same bracket rule as tags: #name or #[two words].
Only one name sticks per prompt — if it produces several images, the first one gets it.

5. Chaining — carrying a character through a batch

Put the two together: give a character a name once, then keep calling it back for every prompt that follows, all inside one run.

a young woman, red hair, freckles, studio portrait #lina

@lina sitting in a sunlit cafe, candid, 35mm

@lina hiking a mountain trail at sunrise, wide cinematic shot

@lina in a spacesuit on a red-rock planet, dramatic light

The opening prompt both creates and labels lina; each prompt afterward pulls in that exact result — the same person, recognizably, all the way through.

Both can sit in one line — #hero @city walking out of the station at dusk labels the new result hero while also referencing your existing city image.

6. Video prompts

The Videos tab batches clips the same way — one prompt per block, blank line between them. Set model, aspect ratio, outputs per prompt and duration once, pick a video mode, and press Start.

Frame tags — `@start`, `@end`, and references

The same idea as @tag for images, but for pinning down a clip's exact frames:

Start frame mode: @start:name sets where that clip opens.
Start + End mode: @start:name paired with @end:name sets the first and last frame.
Reference mode: plain @name tags (up to 4) work exactly as they do on the Images tab.

@start:sunrise the camera pushes slowly through the misty valley, cinematic

@start:door_closed @end:door_open the old wooden door swings open into daylight

Multi-word frame names use brackets too: @start:[my hero].
Frame tags never reach the model as text — they're stripped out and shown instead as chips on the row, so you can see which frames a clip is set to use.

Models & durations

Omni Flash — the fastest option; covers Text, Start-frame, and Reference modes.
Veo 3.1 Lite / Fast / Quality — higher quality, locked to 8s; Quality is restricted to Text and frame modes but is the only one that unlocks 4K upscale.
An incompatible combination gets auto-corrected (Veo forced back to 8s, for instance) rather than sent to Flow to fail.

Finished clips show up live and save into their own Videos folder, kept apart from image output.

7. What wins when references overlap

If a prompt could draw from more than one source, Omninator checks in this order:

1. Tags written in the prompt — @tag, @start:, @end: take priority over anything else.
2. The shared reference set — Subject / Scene / Style from the panel, used only when a prompt doesn't specify its own.

Something unclear or missing? Open an issue on GitHub.