Image Models (Modify Frame)
← Closed-source video models · Model catalog · Upscale models →Image models edit, stylize, or transform a single frame in the Modify Frame workspace. All current image models accept instruction-based prompts. The most important habit is including preservation directives.
💡 Preservation directives — For most edits, instruct the model to preserve what should not change: “Keep the original composition.” · “Keep the original outlines.” · “Keep the original character intact.” · “Do not change the lighting.” · “Keep the original framing and proportions.” These work with GPT Image 2, Nano Banana 2, and Nano Banana Pro.
GPT Image 2
OpenAI image model. Strong at precise editing; respects original content very well.- Output resolution: 2K.
- Prompting: instruction-based.
- Quality setting: standard or high. High improves fine detail like small text. Doesn’t change output size but increases generation time. Most useful paired with a video model that can preserve that detail.
- Image style reference: optional input guiding the visual style.
- Mask image input: a painted mask area limiting the edit to a specific region.
Use when — Precision editing of a specific element. Adding/removing objects. Text-sensitive edits (GPT Image 2 handles text better than most models).
Nano Banana Pro
Strong general-purpose image edit and style transfer model.- Output resolution: up to 2K.
- Prompting: instruction-based.
- Image style reference: optional.
⚠️ Known behavior — When adding an image reference, Nano Banana Pro may confuse which input is the source and which is the reference. If results are wrong, retry with explicit phrasing: “use the first image as the reference” or “use the second image as the source”; if it still fails, invert the images in the prompt.
Nano Banana 2
- Output resolution: 1K, 2K, or 4K.
- Prompting: instruction-based.
- Image style reference: optional.
- Best for: higher-resolution outputs, since it supports 4K natively.
Legacy & specialized image models
Still available but generally superseded by the models above.| Model | Notes |
|---|---|
| Nano Banana | Lighter, faster predecessor to Nano Banana 2 and Pro. |
| Kontext | Older edit model. CFG range 1.1–15. |
| MAGO (img2img) | Older Mago-internal style transfer model with ControlNet support. |
| Seedream | Original default. Still available for compatibility. |
Selection guide
| Goal | First choice | Alternative |
|---|---|---|
| Precise edit, one element | GPT Image 2 | Nano Banana Pro |
| Style transfer | GPT Image 2 | Seedream |
| 4K output | Nano Banana 2 | — |
| Mask-based edit | GPT Image 2 (mask input) | Mago Inpaint on video |
| Character preparation for video | GPT Image 2 | Nano Banana 2 |
| Text-sensitive edit | GPT Image 2 | Nano Banana 2 |
← Closed-source video models · Model catalog · Upscale models →