kie-gemini-omni
Kie Gemini Omni — image + video generation harness
You drive Kie AI's job queue (/api/v1/jobs/createTask + /api/v1/jobs/recordInfo) to produce images and short clips through Google's Gemini-family models. Veo3 is deprecated — use Nano Banana Pro for stills, gemini-omni-video for clips.
When to use which model
| Use case | Model ID | Aspect / resolution | Latency |
|---|---|---|---|
| DEFAULT — X-reply image, blog hero, brand visual, poster, diagram | gpt-image-2-text-to-image (GPT Image v2, Anton's pick) |
aspect_ratio auto/16:9/1:1/9:16/… , resolution 1K/2K/4K | 30-120 sec |
| Edit / restyle an existing image, or composite with reference images | gpt-image-2-image-to-image |
same; pass refs via input_urls (public URLs) |
30-120 sec |
| Fallback if GPT Image v2 errors or rejects the prompt | nano-banana-pro (Gemini 3 Pro Image) |
16:9, 1:1, 9:16 — 1080p/2K | 10-30 sec |
| Cheap batch images (~20+ at once) | nano-banana-2 (Gemini 3.1 Flash Image) |
same | 5-15 sec |
| Short-form vertical clip 6-10s (TikTok/Reels/Shorts B-roll) | gemini-omni-video |
9:16, 16:9, 1:1, up to 10s | 90-300 sec |
Default still model is gpt-image-2-text-to-image (per Anton, 2026-06-01). It renders on-image text far more reliably than Nano Banana and is the right tool for posters/cards with words. Nano Banana Pro is the fallback only. Note the param differences vs Nano Banana: GPT Image v2 uses resolution: 1K|2K|4K (NOT 1080p), takes aspect_ratio (default auto; 1:1 cannot go 4K), and does NOT take output_format (returns PNG). Result still comes back as .data.resultJson → .resultUrls[0], same poller, same Cloudflare-UA download gotcha.
Never use Veo3 — Kie still exposes the endpoint for backwards compat, but the model is outdated, the prompts you'd write for Veo3 don't translate, and Gemini Omni Video produces equivalent or better output at lower credit cost.