Generation

The panel opens on a model gallery headed "Select a model to get started." There's no separate "what do I want to make" switch — the model carries its modality, so choosing a model is choosing the output type. To narrow the list:

• Search — type in "Search models…" to filter by name or description.
• Type pills — filter by modality: All, Text to Image, Image to Image, Text to Video, Image to Video, Video to Video, Text to Audio, Lipsync, Image to 3D, and Upscale.
• Recently Used — your last-used models sit at the top (with a use-count badge) so repeat work is one click.

Each card shows the model's name and a short description; the provider is hidden (one card per logical model). A model that runs on your own key shows a "your key" badge ("Runs on your desktop using your local provider key"). Click a card to select it; use ← Back in the header to switch models.

A few cards are opt-in only — Flux Kontext Pro and Flux Kontext Pro Edit today. They are fully available here and you can pick them from this gallery like any other model, but the Assistant will never choose one for you: not as a default, and not as a stand-in when the model you named is switched off for this project. Kontext is flagged this way because its provider moderates the endpoint on its own side and returns all-black frames for a lot of ordinary production content. If you want it, name it ("use Kontext") or pick it here.

Models that take media (image-to-image, image-to-video, lipsync, upscale, and so on) show input slots above the prompt. Each slot is labeled by the role it plays:

• First Frame / End Frame — the start and end images for a video.
• Reference Images — one or more images the model draws style or subject from.
• Input Video — the clip a video-to-video model transforms.
• Reference Audio — audio to drive lipsync or scoring.

Click a slot to open the picker, which browses this asset's own files — navigate categories and folders, then Done. The Pinned pill switches the grid to whatever is pinned in the Assistant's context, from anywhere in the project, so a reference you were just discussing is one click away. It uses the full-resolution source, not the thumbnail.

Every tile shows its status badge, and a status dropdown next to the swatch slider controls what's offered. The default is Published files from anywhere, plus Selected files that belong to this asset — another asset's work-in-progress is never offered, only what it has published. Untick or tick statuses in the menu to widen or narrow that (Show All, or opt into New, Dropped or Untracked); Default restores the rule. When everything in a folder is hidden, the picker says how many files the filter is holding back.

Reference images get @-mention badges (@Image1, @Image2…) you can name in the prompt to say how each one should be used. To drop an input, click the × on its thumbnail in the form — no need to reopen the picker. Picking the same file again in the picker toggles it off rather than adding it twice.

The Prompt box is a resizable, multi-line field. Above it sits a small toolbar for reusing prompt language:

• Primers — inject reusable prompt blocks into the prompt.
• Save current prompt as primer — keep good language to reuse later.
• Browse previous prompts — pull from prompts used before on this asset.

Primers come in three scopes: System (always included, can't be turned off), Project, and your personal My Primers. The prompt actually sent is the composition of the active primers plus your typed text (see the Primers topic). Prompt-less models (image-to-3D, and upscalers, which are directed by their parameters instead) hide the prompt box; a negative-prompt field only appears if the chosen model defines one.

Parameters are generated from the model's own definition, so you only ever see controls the model actually supports (resolution, aspect ratio, duration, and the like). Rarely-touched fields — seed, steps, guidance — are tucked behind an Advanced (N) disclosure; defaults are sensible, so expand it only to tune.

A Number of Images / Number of Videos stepper sets how many outputs to make — default 1, up to 10. Ask for more only when you want variations to choose between.

Press Generate (or Cancel to back out). Loft checks that the prompt and any required fields are filled, then submits the job to the queue on the right, where it moves through Pending → Running → Complete (or Error), refreshing every few seconds.

1. Finished results appear as thumbnails on the queue card — click one to open the full-screen preview (videos play; siblings show as a filmstrip).
2. Expand a card for Copy prompt, Regenerate, the inputs grouped by slot (First frame, References, …), and a details table (model, size, seed, steps, and more).
3. Completed takes are delivered into this asset's files — the asset-name link on the card jumps you there. Remove a queued item with Delete.

A live cost estimate shows next to the prompt toolbar before you submit — in credits or dollars depending on your billing mode (it's hidden if the model has no pricing on file). Two ways a job can run:

• Loft keys — the Loft server runs the job and bills your credit balance. The default for Virtual projects.
• BYOK ("your key") — a model marked with the your-key badge runs on LoftDesktop using your own provider key, billed to your provider account, not Loft.

For a Local project with no desktop key, the server can still generate on Loft keys and deliver the output back into your local folder. Actual spend is recorded per job and rolls up in Costing and your wallet — see the Credits & Billing and Queue topics.

If your credit balance won't cover the estimate, the Generatebutton is disabled and the panel shows how many credits are needed versus available, with a Purchase credits link to billing.