About
We build an AI image studio you can actually talk to
ImgPilot is built around one idea: most people get better results from describing edits in plain language than from writing the perfect prompt. The rest of this page is how we deliver on that — which models we expose, how we test prompts before publishing them, and what we will and won't do to chase rankings.
What we're optimizing for
Open the chat. Type what you want changed. The image responds. That's the entire interaction loop we're optimizing for.
Behind it sits a careful set of choices: which models to expose, how to handle failures, how to refund credits when a generation breaks, and how to teach beginners which model is right for which job. None of those should be decisions our users have to make.
Models we use
We expose 8 image models in one chat. We don't host them ourselves — we route to Kie API for the primary call, with Replicate as a fallback when Kie is unavailable. We pick which models to add based on real testing, not press releases.
| Model | Why it's in our lineup |
|---|---|
| GPT Image 2 | Default for new sessions. High text-in-image accuracy, supports up to 16 reference images, up to 4K (caveat: 'auto' aspect is 1K only, 1:1 doesn't upscale to 4K). |
| Nano Banana Pro | Best for multi-character consistency (5+ characters) and text-in-image accuracy (94% by Everypixel evaluation). Supports up to 4K. |
| Nano Banana | Cheapest entry point for general text-to-image and quick edits. No 4K. |
| Nano Banana 2 | Adds Google Search grounding to the Nano Banana family. Useful for fact-anchored visuals like product shots that need real-world specifics. |
| Flux Kontext Pro | Single-image edits with strong instruction-following. We use it most for clean background swaps and object removal. |
| Flux Kontext Max | Higher-fidelity Kontext variant. Worth the extra credits when Pro misses the mark on a hard edit. |
| GPT-4o Image | Up to 5 reference images. Good for compositional remixing where you want to stitch elements from multiple sources. |
| Grok Image | 720p output ceiling — included for stylistic variety and i2i moods, not for production-grade output. |
How we curate prompts
We publish prompts on /ai-image-prompts only after we've actually run them. The Before / After slider on those pages is real output, not stock photography.
Screen
Reject prompts that copy verbatim from another site without adaptation, or that promise effects we know the available models can't deliver.
Test at least 3 outputs
Generate the prompt at least 3 times across different seeds before deciding whether to publish. One lucky output isn't evidence.
Record actual params
Save the model, aspect ratio, resolution, and any observations into the post's metadata so the public page can show what we tested with.
Capture before / after
For edit-style prompts (style transfer, restore, background swap), we save both input and output and render them as a draggable comparison slider.
Publish or reject
Prompts that don't reliably produce a usable result don't get published, even if they're trending elsewhere. The library should be a shortcut to good results, not a graveyard of broken ideas.
Editorial standards
No fake bylines
We will not invent author personas to game E-E-A-T signals. When a real human writes a deep-dive, you'll see their real name and a real link.
Claims are anchored to sources
Model capability claims — 4K support, multi-character counts, text rendering accuracy, output ceilings — come from official model documentation (Kie API docs, model release notes) or third-party benchmarks. We don't repeat marketing copy verbatim.
Pricing is machine-readable
Our full pricing — tiers, credit allocations, monthly vs yearly, what each plan can do — is exposed at /pricing.md as plain markdown so AI agents and humans can read it without scraping the rendered page.
Updates are dated, never backdated
If a page hasn't materially changed, we don't refresh its publish date to fake freshness. Sitemap lastmod only moves on real content changes — that's a contract baked into our build.
Failures get refunded
When a generation breaks because of a model or pipeline failure, we refund the credits automatically. This isn't a goodwill gesture — it's how the system is designed.
Contact
Bug reports, prompt suggestions, partnership inquiries — email is the fastest channel.
