Generating images & video

Image models

• Seedream (default) — fast, strong general-purpose generations.
• Nano Banana / Nano Banana Pro — alternative looks; Pro adds resolution tiers up to 4K.
• GPT Image 2 — quality tiers (low / medium / high).

Images support 1K/2K/4K resolutions (per model), 9:16 / 1:1 / 16:9 aspect ratios, up to 4 variants per generation, and up to 10 reference images for image-to-image guidance.

Video models

• Kling (default) — 5–15 seconds, 720p/1080p/4K.
• Seedance / Seedance Fast — 5–60 seconds, 480p/720p/1080p (Fast is 720p); the only model family with audio reference support for voice and lip-sync.
• Veo 3.1 / Fast / Lite — 4, 6, or 8 seconds, 720p/1080p/4K (no 4K on Lite).

Video models are image-to-video: the render starts from a reference image — the first one attached, or the workspace's most recent image render. Optional audio generation is available per render.

Premium resolutions

The top-end SKUs — Seedance 1080p, Kling 4K, Veo 4K, Nano Banana Pro 4K, and GPT Image 2 high — are available on the Max plan and above.

Credits & refunds

Video renders are charged in render credits when the render starts, scaled by model, resolution, duration, and audio. The cost is always shown before you generate. Image renders cost credits on Free/Pro/Max and are plan-included (free) on Team and above.

Every failure refunds automatically. If a render fails outright you get the full charge back; if a long chain partially completes you keep the finished part and get the difference back. The refund amount is reported on the render itself.

Reference images & anchors

Renders can be steered with reference images: your own uploads, a pinned past render, a locked character, or cast characters. Janusly composes these in a fixed precedence order so identity anchors always win, and trims to each model's reference cap.

Programmatic access

The same pipeline is available over HTTP on any paid plan — see the API reference — or from agents like Claude Code via the MCP server.