Model guide
Overview: model choice and shared settings
Meet the image and chat models available here and choose by task: web references, negative prompts, palettes, transparency, and sequential images.
Open featureView current models and settings in the workspaceOn this page
Start with the task, not a universal best model
Choose chat for developing ideas and image generation for producing pictures. Then distinguish text-only creation, reference fusion, edits, text layouts, and related series. Find required controls before choosing a model; Pro or Max is not automatically the right choice.
This chapter explains capabilities of models and controls integrated in BANANAI, not unintegrated features. Suggested uses are starting points, not local benchmark rankings or guarantees of speed or first-try success.
Model capability does not grant every option to every account. Membership, references, output count, dimensions, thinking, and ranges follow the current workspace. Check permissions and inputs when an upgrade message or disabled control appears.
Expand the actual examples in each model-family chapter for prompts, settings, original references, unretouched outputs and observed deviations. Generation Settings screenshots are captured in both real interface languages without paying for a second generation. Examples include realistic people, products, interior edits, transparency and sequences; they are not an exhaustive parameter benchmark.
Choose a starting point by task
| Task | Models to try | Check first |
|---|---|---|
| Explore composition and direction | nano banana 2 lite, FLUX.2 Klein 4B, Seedream 5.0 Flash | Starting points, not a speed or cost ranking. Define subject and purpose and compare equivalent requests. |
| Fuse references or retain character/product features | nano banana 2/2.1/pro, Wan 2.7, GPT Image | Give each reference a subject, style, or scene role and check count. |
| Posters, packaging, titles, and information layouts | Ideogram 4.0, Qwen Image 3.0, nano banana pro | No dedicated typesetting switch or guarantee of greater text accuracy; proofread results. |
| Fine-tune the generation process | FLUX.2 Flex | Steps and guidance are exposed here. Fix inputs and seed, changing one control at a time. |
| Related characters, stories, or image sets | Sequential-capable Seedream 4.0 and Wan 2.7 | Candidates differ from sequences. Describe shared character and each change; check combined reference limits. |
| State exclusions separately | Qwen Image 3.0/Pro | Use Negative prompt. In unsupported models, state requirements in the main prompt. |
| Specify a brand or series color direction | Wan 2.7/Pro | Enable Palette and set proportions. This is not exact pixel locking or direct recoloring. |
| Transparent materials or delivery formats | GPT Image 2.5 Flare/Sunburst | Choose transparency with PNG or WebP. JPEG cannot retain alpha; white is not transparent. |
| Real-world research or planning before creation | Gemini chat; nano banana 2/2.1/pro for image web references | Specify sources, scope, and date needs. Web use neither guarantees accuracy nor uploads website images automatically. |
Model-specific controls available here
Writing a parameter name in a prompt does not create an unsupported control. A seed written to a seedless model does not set its random seed; writing 'transparent' to GPT Image 2 does not enable transparent-file output.
Reference and dimension limits are not universal account allowances. Recheck after changing model, membership, references, or output mode.
| Control | Available models | When useful |
|---|---|---|
| Image web references | nano banana 2, nano banana 2.1, nano banana pro | For real-world information; existing references and clear instructions do not always need web search. |
| Image thinking level | nano banana 2 lite, nano banana 2, nano banana 2.1, nano banana pro | For more constrained instructions; use only currently available levels. |
| Separate negative prompt | Qwen Image 3.0, Qwen Image 3.0 Pro | To organize exclusions separately from positive requirements. |
| Palette | Wan 2.7, Wan 2.7 Pro | For color ranges and proportions. Verify strict brand colors afterward. |
| Steps and guidance | FLUX.2 Flex | Compare the effect of process settings with fixed input rather than just change a model name. |
| Transparency, WebP, compression | GPT Image 2.5 Flare, GPT Image 2.5 Sunburst | For transparent files or JPEG/WebP size. PNG offers no compression control. |
| Multiple candidates | GPT Image 2/2.5, Wan 2.7/Pro, Qwen Image 3.0/Pro | Alternative images for the same requirement, not an automatic story. |
| Sequential image sets | Seedream 4.0, Wan 2.7/Pro | Related images with a shared subject and specific changes per image. |
Gemini chat models
Chat develops goals, alternatives, consistent prompts, input analysis, and real-world information. Discussing an image does not submit generation. Transfer a finalized plan into the image workspace.
All four expose web search and low/medium/high thinking here. Equal controls do not mean equal answers. No speed, cost, or ability ranking is claimed; compare the same material against explicit goals. Account entitlements still limit available levels.
Use web search for current facts, product information, or source-backed work. Summaries, ordinary rewrites, and brainstorming do not always need it. More constrained analysis may benefit from higher thinking; simple wording can start lighter.
Check web sources and dates. Summarizing material does not grant permission to use its images. A chat-generated prompt is not automatically saved as material or inspiration.
Prompt optimization uses an available chat model, not an image model, and is a separate call charged under that model's actual-use rules. Availability does not make optimization unlimited or free.
| Model | Tasks to try | Settings and limits here |
|---|---|---|
| gemini 3.5 lite | Short questions, rewrites, summaries, and simple prompt organization. | Web search and low/medium/high thinking. Start light for clear small tasks. |
| gemini 3.5 | General questions, creative goals, alternatives, and prompt optimization. | Web and low/medium/high thinking. Existing-content rewrites need not use web search. |
| gemini 3.6 | Compare material, check multiple constraints, organize plans, and explain image/text content. | Web and low/medium/high thinking. These tasks are not exclusive to it; compare equivalent materials. |
| Gemini 3.8 | Creative analysis with several constraints, factual questions, and planning discussions. | Web and low/medium/high thinking. A higher version or thinking level is not a correctness guarantee. |
What the controls are for
| Control | Use | Avoid this assumption |
|---|---|---|
| Web search/references | Current or real-world information. | Not guaranteed factual or automatic reuse of website images; must be offered by the model. |
| Thinking | More constrained instructions, analysis, plans, or comparisons. | Not resolution, sharpness, or step count. Reasoning ability need not expose a control. |
| Negative prompt | Separate exclusions in a supported model. | Not an absolute prohibition or moderation exemption. |
| Seed | Keep a random starting point while comparing changes. | No identical-image guarantee across models, versions, or settings. |
| Steps/guidance | Fine-tune FLUX.2 Flex generation and constraints. | More is not always better; changing everything obscures the cause. |
| Palette | Specify series or brand-color direction with Wan. | Not pixel locking; verify precise colors. |
| Candidates/sequences | Compare alternatives versus create related images. | Not all models offer both; multiple images do not guarantee identity consistency. |
| Quality, dimensions, format | Output tier, pixel range, and delivery/transparency. | Different controls. Enlarging a preview adds no pixels; JPEG has no alpha. |
| Compression | JPEG/WebP file delivery in GPT Image 2.5. | Not thumbnail compression or another quality tier. Recheck after changing format. |
Recheck after switching models
Cards and Canvas share capability, entitlement, and submission rules. Historical settings describe the original task; switching does not turn an old image into a new model's output. Regeneration still rechecks current access and inputs.
- Check account access; being listed in Help does not grant authorization.
- Check reference count and each image's role. A single-reference model cannot accept a multi-source plan.
- Recheck dimensions, ratio, quality, count, sequences, web, and thinking. Values are matched to the new model and may differ.
- Recheck negative prompt, seed, palette, transparency, format, and compression. A disappearing control is not still applied invisibly.
- Check both text-field limits and the new full estimate. Higher quality, size, and count are separate choices with different costs.
Find a suitable model through comparable attempts
- Define purpose, subject, essential details, and allowed changes. Explain multi-reference roles.
- Choose an available model and check inputs and settings. Use deterministic tools for exact cropping, mirroring, splitting, or annotation.
- Keep main inputs and description and change one model or key parameter at a time. Compare adherence, details, text, and consistency, not only attractive thumbnails.
- After selecting direction, choose quality, dimensions, and format. Inspect actual files for text, alpha, detail, and deliverability. Collect inspiration to keep the method.