Prompt Studio is the Classic workspace profile for repeated image iteration. It keeps the image viewer in the center, recent generations on the side, and the composer plus generation controls close to the current image.
Use this page when you are not sure what a Prompt Studio control does or when you want a reproducible way to report a workspace issue.
| Area | What it does | Notes |
|---|---|---|
| Composer | Holds the prompt for the next request. | Selecting an existing image is view-only. It should not overwrite the composer. |
| Storyboard | Maintains character and scene continuity across sequential frames. | Toggle in the composer. Works for image and video generation; image keyframes are composed for video production. |
| Multimode | Starts several separate image requests from the current prompt. | Each slot is a candidate output, not a collage panel or a guaranteed scene sequence. |
| 1:1 Direct | Sends the prompt through with less rewriting by the app. | Use it for exact wording, strict prompt experiments, or provider-side prompt syntax. |
| Model quick menu | Changes the image model and reasoning effort from the sidebar header. | The full Settings workspace remains the detailed configuration page. |
| Recent generations | Shows the visible Prompt Studio history domain. | Arrow keys move inside the same visible recent domain instead of hidden older rows. Video items render as video thumbnails. Drag any thumbnail to the composer to add it as a reference image. Video results expose First, Mid, and Last frame buttons to copy keyframes. |
| Video settings | Controls Grok video duration, resolution, aspect ratio, and planner model. | Default planner model is grok-4.3; grok-4.6 and grok-4.5 are selectable. |
| Gallery | Browses saved local images, All/Favorites tabs, and folders. | Favorite toggles should preserve the gallery viewport you were browsing. |
| Prompt library | Imports saved prompt text into the composer intentionally. | Library insert/continue actions are explicit prompt imports; passive image selection is not. |
Multimode repeats one generation request shape across several candidate slots. It is useful when you want alternatives, not when you need one combined multi-panel image.
For related candidates, put the shared subject first and the variation rule second:
Same character design in every image: a silver-haired courier in a red raincoat.
Make 4 alternatives that vary only camera angle, lighting, and background street.
Keep face, outfit, age, and color palette consistent.
For unrelated candidates, say so directly:
Create 4 unrelated sticker ideas for a local image generation app.
Each image should use a different mascot, color palette, and composition.
If you need a true two-panel comic, contact sheet, before/after comparison, or collage, ask for that in a single normal image request instead of relying on multimode slots:
Create one 2-panel comparison image. Left panel: rough sketch UI. Right panel:
polished Prompt Studio UI. Add small labels inside each panel.
Use 1:1 Direct when the exact prompt text matters. It is helpful for:
- comparing prompt wording changes,
- preserving a structured prompt template,
- using provider-specific instruction style,
- avoiding app-side phrasing changes during troubleshooting.
Direct mode can be used together with Multimode. In that case, each multimode slot receives the same direct prompt request shape.
Reasoning effort controls how much planning the selected model may spend before or during generation. Start with the default for everyday work. Raise it when the prompt has many constraints, references, or composition requirements.
The sidebar model label opens quick settings for both model and reasoning. The Settings workspace still shows the full model configuration and explanatory copy.
Prompt Studio separates browsing from composing:
- Passive image selection is view-only.
- Clicking a history or gallery image focuses the image for viewing.
- Favorite on/off changes the saved image metadata and should not jump your gallery browsing position.
- The All and Favorites tabs are browsing filters. Switching between them does not intentionally import a prompt.
- Prompt Library insert, "continue from this image", and explicit reuse actions are the actions that intentionally change the composer.
Before generating, glance at the composer if you were using explicit prompt import actions. Passive image selection should leave your draft alone.
The v1.1.13 Prompt Studio fixes tightened these user-visible contracts:
- concurrent multimode completions keep their slot/request identity,
- keyboard navigation follows the visible recent history domain,
- the gallery button remains reachable beside recent history,
- long prompts no longer starve the default image viewer,
- Direct and Multimode states can be seen at the same time,
- gallery favorite toggles and tab changes preserve the browsing viewport,
- passive image selection does not refill the composer.
When opening an issue, include:
ima2-genversion and operating system,- browser and viewport size if layout is involved,
- workspace profile, mode toggles, model, reasoning effort, and image count,
- numbered steps from a fresh
ima2 servesession, - whether the problem happens in All, Favorites, or recent history,
- safe screenshots or screen recordings if they do not reveal private prompts.
Do not share ChatGPT cookies, OAuth token files, API keys, raw upstream responses, private prompt history, or generated base64 data.
When you draw memos (notes), arrows, or boxes on the canvas and save, the notes reach the model as text instructions, not pixels.
- Saving a canvas version keeps the burned-in annotated image for UI display and history only; model references always use the clean source. Adding the current image as a reference also swaps a canvas version for its original.
- Note contents attach to the composer as a removable "Canvas notes" chip, so the next generation carries the instructions (position + content + a markup-removal rule).
- Drawing boxes and running the mask edit calls the edit API with the clean source, a box mask, and the note instructions.
If markup still shows up in a result:
- Phrase notes as what should be visible in that spot, not only what to remove. For example, "show the continuous wooden desk surface where the sticky note was" instead of "remove the sticky note".
- Generate several candidates from the same request and pick the cleanest.
- Box only the residual trace and run one more mask edit.
Canvas versions that contain baked notes offer Revert baked notes. When no later pixel edit was added, the clean source is restored and the saved notes return as an editable draft. If background cleanup or another pixel edit was added afterward, Prompt Studio asks for confirmation because reverting also discards those pixel edits.