Model choice should follow the job the generated asset needs to do in the project. A background, product hero, storyboard frame, and motion clip usually need different prompts and review standards.
Start from the output role
Before choosing a model, decide whether you need a still image, a video source, a texture, a reference direction, or a final visual. This helps you judge the result against the layout instead of only judging it in isolation.
| Output role | Prefer | Review for |
|---|---|---|
| Background | Clean composition and negative space | Text safety, color contrast, and crop flexibility |
| Product hero | Stable subject, clean edges, and consistent lighting | Whether product scale and perspective stay believable |
| Style direction | Expressive model or reference-heavy prompt | Mood, palette, and usefulness as a direction |
| Motion source | Clear subject/background separation | Whether the subject can animate without visual noise |
| Final visual | Higher fidelity settings after layout is stable | Detail quality, artifacts, and export size |
Generate into the project
Generated outputs should land near the canvas where they will be used. Place promising results next to the frame, compare them at the intended size, and keep rejected directions only when they are useful references.
If the output will become part of a final scene, save it with a meaningful name and keep provenance where possible: source prompt, reference assets, model, and intended use.
Compare models deliberately
When testing models, keep the prompt, aspect ratio, and reference material stable. Change one variable at a time so you can tell whether the model, prompt, or composition caused the difference.
Use a simple comparison grid:
- Same prompt.
- Same aspect ratio.
- Same reference inputs.
- Same output count.
- Same target frame.
Then judge inside the layout, not in the asset preview alone.
Review before refining
Check crop, subject position, text safety, and motion suitability before spending time on polish. If the result needs to animate, make sure the subject has enough clean space and separation from the background.
Do not refine an image that fails the frame. Fix the role, crop, or prompt first. Refinement is best after the generated result already supports the composition.
Cost awareness
Generation and rendering can consume the workspace Flares balance. The app quotes Flares before each generation starts, based on the selected model, input type, resolution, duration, and output count.
AI model calls are priced at provider cost with no Flare model markup. Use subscription Flares for routine tests, then use top-up Flares for larger batches or paid Recipe runs. For larger batches, run a small test first, then scale once the prompt and layout are stable.
When to change models
Change models when repeated prompt edits are not addressing the real issue:
| Symptom | Try |
|---|---|
| Good style, poor object fidelity | A model or setting better suited to subject accuracy |
| Good subject, weak composition | A simpler prompt with stronger layout direction |
| Good still image, poor motion potential | A cleaner source with more subject separation |
| Good result, too expensive for batch work | Smaller proof runs, fewer outputs, or a cheaper model for exploration |