The Previz tab turns rough previsualization into a finished look, from a single frame to a whole clip. Pick one of its two input modes at the top of the settings panel.
Still Frames and Previz are built to work together. The source stays loaded when you switch between them, so you can stylize a frame of a clip in Still Frames, then use it to restyle the whole clip in Previz.
To turn a sketch into a finished image, use Sketch in the Image tab.
Still Frames
Load the source
Drop a previz image or video on Previz Source. With a video, the Input card in the working area shows its frames: select up to nine and click Done. Each selected frame becomes its own generation, with the same settings. Select Frames reopens the picker.
Set the look
Load a Style Reference image, or choose a style preset under Style. Without a style, write the look in the prompt.
Add references
Under References, turn on Identity, Product or Environment to add a Character, Product or Environment slot. Each slot you turn on must be filled.
Generate
The prompt is optional once a style or a reference is loaded.
Preprocessor is on by default in Still Frames. It reads the depth of the source (Depth Map (Depth Anything V2)) to keep pose and composition closer to the previz. The depth map is sent as one more image, so it counts toward the model’s input limit.
Not every model takes references. With Identity or Environment on, the list keeps Nano Banana, Seedream 5.0 Pro, ChatGPT Image and Muse Image. With Product on, it keeps Nano Banana Pro, Seedream 5.0 Pro, ChatGPT Image and Muse Image.
Previz
The Previz input mode restyles a whole clip: the model keeps the motion of the source video and takes the look from a stylized frame.
Load the clip
Drop a video on Previz Source. Each model has a length limit: up to 10 seconds on Kling and Gemini, 18 seconds on Luma. The source card says when a clip is out of range.
Load the stylized start frame
Stylized Start Frame is required: the first frame of the clip, already in the finished look. Make it in Still Frames from the same clip, then drop the result on the slot.
Add references
Add images, and on some models videos, in Reference 1 and the slots after it. The prompt is optional once the start frame is loaded.
| Model | Extra references |
|---|---|
| Kling | Up to 3 images |
| Seedance and MiniMax | Up to 8 images and 2 videos |
| Gemini | Up to 5, of which at most 3 videos |
| Luma | Guide frames instead (below) |
With Luma, the first slot is Stylized Guide Frame 1, and you can add up to 63 more guide frames. Each one is a stylized frame pinned to a moment of the clip, shown as its Frame Timecode. On an empty guide card, Pick Frame lets you choose that moment in the source video. Guide frames are images, and each moment can hold only one.
Style presets and the Preprocessor are not used in this input mode: the stylized frame carries the look.
Models
Each input mode offers its own models. Still Frames starts on ChatGPT Image 2.5 Flare.
Still Frames
OpenAI- ChatGPT Image 2.5 FlareDefault
- ChatGPT Image 2.5 Sunburst
- ChatGPT Image 2.0
GeminiGoogle- Nano Banana 2.1
- Nano Banana 2.0
- Nano Banana 2.0 Lite
- Nano Banana Pro
SeedreamBytePlus- Seedream 5.0 Pro
- Seedream 5.0 Flash
- Seedream 4.5
FluxBlack Forest Labs- Flux 3.0
- Flux 2.0 Pro
- Flux Klein 2.0 9B
WanAlibaba- Wan Image 2.7
- Wan Image 2.7 Pro
- Wan Image 2.6
KlingKlingAI- Kling 3.0 Omni
GrokxAI- Grok Imagine 2.0
- Grok Imagine
Meta- Muse Image
Ideogram- Ideogram 4.5
Previz
Hylope- Hylope V2V 0.1
KlingKlingAI- Kling 3.0 Omni
SeedanceBytePlus- Seedance 2.5
- Seedance 2.0
- Seedance 2.0 Fast
Luma- Ray 3.2
MiniMax- H3
GeminiGoogle- Omni 1.1 Flash