Most of the time you'll let the Agent handle generation — describe what you want and it picks the model and settings. But when you want direct control — an exact model, specific parameters, a fixed seed, or a precise edit to an asset you already have — Layer gives you a generation form you can drive yourself. It's the direct replacement for the old Layer's generation form, built right into the session alongside your board.

💡 This is the expert lane. If you know exactly what you want — a specific model, an exact prompt, a reproducible seed — the form lets you set it directly, no Agent in the loop.

Open the generation form

From inside any session:

The bottom toolbar — click Generate image, Generate video, Generate 3D, Generate audio, or Generate playable. The form opens above the toolbar, separate from the chat.
You can also switch the composer out of Agent mode into a direct generation mode.

Generate directly

In the form, every choice is yours:

Model — pick the exact base model (the live default is shown on the picker, e.g. Seedream 4.0)
Prompt — the exact text the model receives. Chatting with the Agent, it expands and refines your wording for you; here it goes straight to the model, so the phrasing is entirely yours
Attachments — add a Reference Image or Reference sets to keep results on-style
Cost & count — see the CU cost up front (e.g. ~1.2 CU) and set how many to generate (the batch)
Advanced settings — pin a Seed to reproduce a result exactly, and set the Prompt Language
Forge — for power users who want the most fine-grained control over a generation

Then hit generate. The result lands on the board, just like an Agent generation.

💡 Writing the prompt yourself is the expert move. Day to day, the Agent does the heavy lifting of prompt-craft — turning a short request into a full, model-ready prompt. In the direct form there's no Agent in between, so you write the precise prompt the model runs. It takes more prompt know-how, but you get exact, repeatable control. New to writing prompts directly? See [Writing better prompts].

💡 Pin the seed to reproduce a result. With the same model, prompt, and a pinned seed, you get the same image every time — invaluable when you're iterating one variable at a time or recreating an exact setup.

Edit an asset with an exact prompt

This is the fastest way to tweak something you've already made — a prompt edit, with no Agent interpreting your request:

On the board, click an asset to select it. A "Selection · 1 item" chip appears, marking it as the form's source.
Open the image form — your selection is used as the source image.
Type a prompt describing only the change — "make the jacket red," "add a soft rim light," "remove the background props" — then generate.

The model keeps your asset and applies just the change you asked for. Because it's a fresh generation, your original stays on the board untouched — so you can fire off several exact edits side by side and compare. Pair this with a pinned seed for tightly controlled, one-variable-at-a-time iteration.

Generate video — choose your frames

Switch the form to Generate video and pick a video model. Depending on the model, you can directly set:

First Frame and Last Frame — anchor the clip to specific images, so you control exactly where the motion starts and ends
Reference sets — guide the overall look
Duration and aspect ratio

This is how you produce a specific chunk of video with precise control, rather than describing it and letting the Agent decide.

💡 The same form handles image, video, 3D, audio, and playables — switch modality from the bottom toolbar, then pick a model that fits.

Direct form vs. the Agent

Agent (chat): fastest for most work — describe the outcome and it handles model choice, prompt craft, and settings.
Generation form (direct): when you want exact control over the model, parameters, seed, frames, or are reproducing a specific setup. It's the home for everything the old Layer's inference form did — now built into the session alongside your board.