Choosing a model
A model is the engine that turns your prompt into an image or video. Different models are good at different things. You don’t have to learn them all — by default the studio picks one for you. This page covers the default, then a plain-language guide for when you want to choose yourself.
Start with Auto
Section titled “Start with Auto”When you open the composer, the model is set to Auto. Auto reads your task and picks the right model:
- A plain prompt with no reference image — a strong all-rounder for new images.
- A prompt plus a reference image — a model that works from your reference.
- An edit to an existing image — a model built to change one thing and leave the rest alone.
Auto is marked Suggested and shows “Picks the best model for your prompt.” For most work, leave it on Auto and write a clear prompt. See Writing better prompts.
Choose a model yourself
Section titled “Choose a model yourself”To pick a specific model, click the model name in the composer to open Browse image models (or Browse video models for video). You get a scrollable list with filters across the top:
- Provider — filter by maker (Google, Black Forest Labs, Ideogram, Recraft, xAI, and more).
- Features — Reference support, imported-file support, or Fast.
- Best for — a goal, such as Photoreal, Illustration & art, Text & logos, Portraits, Editing & references, Brand & marketing, Fast drafts, or Video & motion.
- Search — type a model name.
Each row shows the model, its typical render time, and a few tags (References, Fast, Imported files). Pick a model to pin it; close the browser to stay on Auto. Pinning a model with more than one version shows a small set of version chips so you can switch between, for example, the standard, faster, and highest-quality version.
To go back to Auto after pinning, click Auto at the top of the model area.
Which image model should I use?
Section titled “Which image model should I use?”You rarely need to choose, but if you want a specific look, here’s a starting point. The version names below match what you see in the browser.
| Your goal | Try | Why |
|---|---|---|
| A reliable all-rounder | Flux 2 Pro | Handles plain prompts, reference images, and edits in one model |
| Up to 10 references, or a choice of output size | Flux 3 | Text-to-image and edits from up to 10 references, at 768px up to 2K |
| Print-ready or large-format work | Gemini Pro | Studio-grade 4K output and the strongest reasoning of the Gemini tiers |
| Readable text in the image | Ideogram | Renders clean, legible text — posters with copy, logos, social cards |
| Text plus editing in one model | Gemini Flash | Strong text rendering, conversational edits, and multi-reference consistency |
| Stylized illustration or design | Recraft V3 / V4 | Vector and brand-safe styles; V4 takes longer style instructions |
| Bold, creative looks | Grok Imagine | Loose, expressive styles for editorial and mood work |
| Precise instruction-following | Filter Best for to Editing & references | Sharp at multi-step instructions and high-fidelity image inputs |
| A fast draft | Flux Klein, Gemini Lite, or any model tagged Fast | Quick, low-cost passes for iterating on an idea |
| Keep a subject consistent across edits | Flux Kontext | Built for character and subject consistency across a series |
| Use several reference images | Flux 2 Pro, or Gemini Lite for the highest count | Composes from multiple references in one shot |
| A transparent background | Filter Best for to Editing & references | Output an isolated subject on a transparent background |
Need transparent or print-ready output, or want to control resolution and aspect ratio? See Generation settings.
Which video model should I use?
Section titled “Which video model should I use?”Video models take longer than image models and are priced by the second. Auto starts you on a strong general video model. To choose yourself:
| Your goal | Try | Why |
|---|---|---|
| Animate an existing still | Gen-4 Turbo | Built for image-to-video — bring a hero still to life |
| A cinematic short | Gen-4.5 | Flagship quality from text and image |
| Realistic human or character motion | Kling 2.1 | Strong motion and physics for movement-heavy clips |
| Longer realistic motion, up to 20 s | Kling 4 Flash | Kling 4 at 720p, the same price from text or a start image |
| Best value | Hailuo 2.3 | Strong motion at a low per-second cost (add audio separately) |
| Camera moves like a push-in or orbit | Ray 2 | Named camera controls |
| A quick draft | Hailuo 2.3 Fast or Ray Flash 2 | Faster, cheaper passes for testing a shot idea |
| A clip with sound | Seedance 2.5 | Native sound, holds a product’s shape and logo, up to 30 s, 9:16 social clips |
| A cheaper clip with sound | Seedance 2.0 | The previous Seedance, on the same card, up to 15 s |
| The cheapest Seedance | Seedance 2.5 Flash | On the same card, 480p or 720p, up to 30 s with sound, half the price of 2.5 or less |
| Cheap drafts up to a 4K final | Gemini Omni Flash | Native sound, 360p drafts at a third of the 720p price, up to 4K |
| The cheapest clip with sound | Wan 3 | Sound on every clip, up to 30 s, from text or a start image |
Seedance 2.5 makes clips with sound, up to 30 s. Attach one image to animate it from the first frame, or several, up to 30, as references. Pick 480p, 720p or 1080p and any of six aspect ratios, 9:16 included, and turn sound off when you don’t need it. Its card also offers Seedance 2.0, the cheaper option: up to 15 s and 9 references, and Seedance 2.5 Flash, the cheapest: 480p or 720p only, 4 to 30 s, up to 9 references, at half the price of 2.5 or less and a minute and a half a clip. Sound, a start image and references don’t change Flash’s price. 2.5 and 2.0 take about two minutes a clip, 1080p costs more than twice 720p, and a clip the model refuses still costs credits. Seedance’s content filter is strict: it refuses mature prompts and images on any brand.
Gemini Omni Flash makes clips with sound from text or images, 3 to 15 s long, in 16:9 or 9:16. Attach one image to start the clip from it, or up to six as references. Pick 360p for cheap drafts, 720p or 1080p for most work, or 4K for finals: 360p costs a third of 720p and 4K costs three times 720p. Sound, a start image and references don’t change the price. Clips take about a minute and a half. Gemini Omni Flash refuses mature content per run, and a refused run isn’t charged. Runs saved on the older Veo models now open on Gemini Omni Flash.
Wan 3 makes clips with sound from text or one start image, up to 30 s. Every clip has sound and there is no switch to turn it off. Pick 480p, 720p or 1080p and 16:9, 9:16, 1:1, 4:3 or 3:4; with a start image the clip takes the image’s shape. 720p costs twice 480p and 1080p twice 720p, and a start image doesn’t change the price. Clips take about two minutes. Set a seed to repeat a take.
Cauldron models
Section titled “Cauldron models”Cauldron models are models Cauldron runs for your studio. They come with the plans that include them; if yours doesn’t and you’d like them, ask us. In a studio that has them, the Image, Edit and Video pickers list them in their own Cauldron section beside every other model, and each row says how many of the brand’s trained styles and imported files fit it (Takes your styles before there are any). Owners and admins add or remove Cauldron models for the whole studio from Manage Cauldron models, at the end of the picker or under Settings › Studio › Cauldron models; an added model shows in Create for everyone in the studio. Bookmarks to the old pages for these models open the right model in Create.
A note on cost and speed
Section titled “A note on cost and speed”Each model row shows its typical render time, and pinned image and video models show their cost in credits before you generate. Faster, lower-cost models are good for drafts; higher-quality models are worth the wait for final work. Auto already balances this for the task at hand.
If your brand has mature content on, you’ll see an optional toggle.
For how to write prompts that get the most from any model, see Writing better prompts. To start generating, see Creating images.