Chat Bar: Generating Images

The chat bar is the floating prompt composer for generating images (and video / 3D) on the canvas. This page covers image generation: writing prompts, picking models, and feeding in references, LoRAs, and aspect ratios.

Generate an image (text-to-image)

To make an image from just a prompt:

Make sure the category pill (top-left of the bar) shows Image. With nothing selected on the canvas its icon is a T (text-to-image).
Type your prompt in the input.
Press the Send button (the round arrow-up button, bottom-right) or hit Ctrl/Cmd+Enter. Plain Enter inserts a newline — it does not send.

The result appears as a node on the canvas.

Image-to-image

To generate from an existing image:

Select one image node on the canvas. The category pill icon changes to a single image (one selection) or stacked images (2+); the mode is now image-to-image.
Selected images show as thumbnails in the selection strip above the prompt.
Type your prompt and Send.

With 2 or more images selected, the @ key in the prompt switches to image references — type @image1, @image2, etc. to refer to a selected image by index.

Category selector (Image / Video / 3D)

The leftmost pill picks what you generate:

Image — stills from your prompt.
Video — animate from a prompt (and optional input image).
3D — generate a mesh.

A fourth option, Agent, hands the task to the conversational AI agent and opens the full agent surface. It is for Pro and Max plans, so a free account sees three choices here.

Enhance is not one of the pill's choices. It is a locked mode rather than a switchable category, and it lives only in the 3D editor's Scene tab, where the chat bar opens already pinned to it — see aspect ratio to reframe before generating.

In the 2D editor the category is locked to Image.

Generation mode indicator

The Image pill's icon doubles as a live indicator of what will be generated:

T — text-to-image (no canvas selection)
single image — image-to-image with one reference
stacked images — multi-image input
dotted box — inpainting (a selection has been drawn in the Editor2D editor)

Model picker (override per prompt)

Next to the category pill is the active model pill — the model the chat will actually call, auto-suggested from your category, references, LoRAs, and image inputs.

Click the model name to override the model for this prompt (a popover lists compatible candidates, cheapest first, with credit cost on hover).
Auto at the top lets the suggestion engine pick.
Click the pin on any row to set it as your default for the current mode (text-to-image, image-to-image, multi-image, etc.). An override lasts until the generation mode changes — selecting or deselecting images, or switching category, resets it to Auto for the new mode; pins persist.

Prompt favorites & history

Docked inside the prompt field, at its bottom right:

Star (Sparkle icon) — toggles saving the current prompt to favorites; click the favorites popover to refill the input from a saved prompt. Favorites are scoped to the current mode.
Clock — prompt history; click any entry to refill the input.

Presets

The chat bar has one-click presets to fill your prompt with curated starting points. Presets exist per mode — text-to-image (T2I), image-to-image (I2I), enhance, and inpainting each have their own preset library.

To apply a preset:

Click the BookHeart icon (next to the model selector, top-left cluster).
Pick a preset from the popover. Cards with images flip on click to reveal the full prompt text.
Click Select preset to fill the input with the preset's prompt.

The preset bubble then shows the preset's name. Once you edit the prompt, the bubble auto-clears — a visual cue that you've drifted from the saved preset.

Presets are also available on node toolbars in the graph. Click the BookHeart icon on any generation node to browse presets for that mode.

Share a saved prompt with your org

From the Prompt Favorites popover (Sparkle icon, top-right cluster):

Right-click any saved prompt card or click the organization button on hover.
Toggle the Building2 icon (org logo or building icon) to share / unshare with your organization.
Shared prompts appear in the org filter tab at the top of the favorites popover, visible to all members of your workspace.

Refine with AI (prompt enhancement)

Click the wand (Wand2) icon at the bottom right of the prompt field to open a chat-style prompt-refinement surface. You converse with the AI to improve your prompt, then accept a result to drop it back into the input. This is a Pro feature — free users are prompted to upgrade. See Sketch pipelines & prompt tools for the full enhancer.

Auto-enhancement (Enhance ON/OFF)

The Enhance toggle pill in the bottom row (scroll on it or click to flip) turns on a vision pre-pass: for image-to-image, the input image is analyzed and your prompt is reconciled before generation. Off by default. (Distinct from the 3D editor's Enhance bar, which upgrades the scene render.)

The "+" menu: Design reference, LoRA, Look, Matrix

The + button (bottom row, left) adds optional capability panels:

Design reference

An image the model uses alongside your prompt to adopt color, composition, or structure (model-dependent).

Add via + → Design reference, then click the panel to pick an image, drag a node in, or use the eyedropper to pick from the canvas.
Shortcut: hold Alt and click any image/video/model node on the canvas to set it as the reference.
Strength (header pill, ~0.20 loose to ~1.00 close): scroll to adjust or click for a slider.

LoRA

Add via + → LoRA (up to 4). Each LoRA bubble lets you pick a trained model, set its scale, and open its details.
Picking your first LoRA on an empty prompt drops in the trigger-word phrasing as a starting point.
When a LoRA is configured, type @ in the prompt to mention its trigger word.

See Training & using LoRAs for training and trigger words.

Look (camera / lens / film)

Stacks cinematography modifiers onto your prompt. Add via + → Look.

A Combo dropdown applies curated presets (Cinematic, Epic, Portrait, Street, Dreamy, Anamorphic).
Three slots — Camera, Lens, Film — set independently. Click a slot for its picker, or wheel-scroll over it to cycle.
The dice randomizes all three.
Shares state with the AI sidebar's modifier bar. Warns if your prompt already mentions camera/lens/film info.

Matrix (variations grid) — image only

Generates a grid of variations in one call. Add via + → Matrix.

Composition: 2×2, side-by-side, stacked, triple side-by-side, triple stacked.
Creativity: Subtle, Balanced, Wild.

Aspect ratio

The ratio pill (bottom row) sets output width-to-height. Click for a grid of ratios, or scroll on the pill to cycle. The picker only shows ratios the resolved model supports, and may snap to the model's default when you switch models. Auto (when offered) matches the input image's dimensions. Hidden in the Editor2D editor (output is sized from the canvas).

Number of images & parameter variation

The N image(s) pill (bottom row, Pro/admin only) sets how many images per prompt (1–4). Scroll to adjust or pick in the popover. Cost scales linearly.

When set to 2 or more, a Vary across batch section appears: click the eyedropper, then click any varyable setting pill in the bottom row to vary it; set a ± range and the values are spread evenly across the batch (one image per step).

Model parameters (quality / guidance / steps / seed, etc.)

Each model exposes its own inline setting pills in the bottom row (e.g. steps, guidance, resolution, seed). Click a pill for its control (slider, segmented buttons, or a Random/Fixed seed toggle), or scroll on the pill to adjust. A pill's value turns the category color when it differs from the model default; each has a per-pill reset.

Estimated cost

A coins count sits left of the Send button showing the estimated credit cost for the current configuration (model, image count, parameters).

Reset

Reset props (circular arrow, bottom row) — snaps model parameters, image count, and aspect ratio back to defaults. Only shows when something has changed.
Reset all (circular arrow, top-right) — also clears the prompt and removes all LoRAs. Canvas selection and the Design reference / Look / Matrix panels are left untouched.

Help mode

Open the … menu at the end of the bar and pick Explain the chat bar. While on, hovering any pill shows a description of what it does instead of opening its editor. Press Esc to exit. (This used to be a ? icon in the left icon rail; it moved into the menu with the rest of the bar's settings.)

Voice input

Dictation is the prompt field's left bookend — a microphone button inside the field, available whether or not the field has text in it. Recording transcribes on-device and fills the prompt. Click the stop square to finish. (Send is a separate control and always shows an arrow.)

Collapse / expand & docking

Press Tab (when not typing in the prompt) to collapse or expand the bar, clearing it off the canvas.
The Position switch in the … menu moves the bar between bottom and left orientation. In the 2D editor the bar is force-docked left and the switch is shown disabled rather than hidden.

Editor2D editor: source switch ("Current View")

When editing an image in the 2D editor, the selection strip is replaced by a source switch:

Current View — generate from the visible canvas composite (image-to-image), or inpaint within a drawn selection.
New image — ignore the canvas and generate text-to-image.

A caption under the switch states what Send will do. Drawing a selection always takes priority and inpaints within it.