By default, Air automatically selects a compatible model based on your requested operation and supplied source type, including reference-based video workflows, when you prompt the Canvas AI agent. If you prefer more control over output style, speed, cost, generation duration, or supported input types, toggle off the Auto setting and select a specific model from our industry-leading options. If you select a model before sending the first message in a new Canvas, Canvas uses that model for the first run and saves the selection with the Canvas once it exists. The badges beside each model show its approximate credit cost and generation time. See Explore Credits for consumption estimates.
Image models
Air offers a variety of image generation and editing models to suit your creative needs:
Nano Banana 2: Google's Gemini 3.1 Flash image generation and editing model. It creates vibrant, high-fidelity visuals quickly, with strong text rendering, natural-language understanding, and character consistency for up to five people.
Nano Banana 2 Lite: Google's Gemini 3.1 Flash Lite efficiency option offers sub-two-second latency, cost-effective generation and editing, fast multi-turn local edits, and 14 supported aspect ratios.
Nano Banana Pro: Google's Gemini 3 Pro production-quality image generation and editing model offers strong text rendering, character consistency for up to five people, and support for 1K–4K resolution.
Flux 2 Pro: Black Forest Labs' production-optimized image generation and editing model delivers consistent, studio-grade results for high-volume creative work.
Flux 2 Max: Black Forest Labs' image generation and advanced editing model focuses on exceptional realism, precision, and consistency.
GPT Image 2: OpenAI's previous-generation image model creates high-fidelity images with strong prompt adherence while preserving composition, lighting, and fine-grained detail.
GPT Image 2.5 Flare: OpenAI's fast, general-purpose model supports precise, repeatable edits and high-volume generation for social, creator, and product imagery.
GPT Image 1.5: OpenAI's model that generates high-fidelity images with strong prompt adherence, preserving composition, lighting, and fine-grained detail.
GPT Image 2.5 Sunburst: OpenAI's premium model produces polished campaign and product imagery with tighter editing control and greater fine-detail fidelity, with longer generation times.
GPT Image: OpenAI's multimodal model accepts text and image inputs and produces higher-quality image output at a slower speed.
Seedream 4.5: ByteDance's unified image generation and editing model supports stylized transformations and complex edits.
Seedream 4: ByteDance's previous-generation unified image generation and editing model supports stylized and transformation workflows.
Video models
Model-supported references and video editing are generative workflows. They do not guarantee unchanged source footage or source quality. AI creates brand-new video frames from scratch rather than just cutting and pasting your original clips together.
For dynamic motion and cinematic clips, you can choose from the following video generation models:
Flux 3: Black Forest Labs' model creates 5–20 second videos at up to 1080p with native audio from text or image. It supports start- and end-frame interpolation, clip extension, and instruction-based editing that preserves motion, timing, and framing.
Seedance Pro 1.5: ByteDance's model creates 4–12 second videos at up to 1080p from text or image, interpolates start and end frames, and generates synchronized dialogue, sound effects, ambient audio, and multilingual lip-sync.
MiniMax H3: MiniMax's Hailuo-03 model creates 5–15 second videos at 2K and 24 FPS with native stereo audio from text, an image with optional first and last frames, or references of up to nine images, three videos, and three audio clips.
Seedance 2.0: ByteDance's model creates 4–15 second videos at up to 720p from text, image, or up to nine image, three video, and three audio references. It interpolates start and end frames and generates native audio.
Seedance 2.0 Mini: This faster, lower-cost Seedance option creates videos from text, image, or reference inputs and supports start- and end-frame interpolation.
Kling 3 Pro: Kuaishou's model creates 3–15 second videos from text or image with cinematic visuals, fluid motion, and native audio. It supports start- and end-frame interpolation, multi-shot generation, and video-to-video editing.
Seedance 2.5: ByteDance's model creates 4–30 second videos at up to 720p from text, image, or up to 50 multimodal references. It supports start- and end-frame interpolation, native audio, and single-pass long-form coherence.
Gemini Omni Flash 1.1: Google's model creates 720p video with synchronized native audio from text, image, or reference inputs. It supports natural-language video editing, character and voice consistency, cinematic camera control, and on-screen text.
MiniMax H3 Max: This Hailuo-03 variant emphasizes prompt adherence and aesthetics, creating 5–15 second videos up to 768p with native audio. It supports an optional last frame, up to nine image and three video references, clip extension, and five-second renders in under three seconds.
Wan 2.7: Alibaba's model creates smooth 2–15 second videos at up to 1080p from text or image. It supports start- and end-frame interpolation, reference-to-video, and instruction-based editing.
Veo 3.1: Google DeepMind's model creates 1080p video with dialogue, sound effects, and ambient audio, plus accurate lip-sync and realistic motion. It supports start- and end-frame interpolation and clip extension.
Wan 3.0: Alibaba's model creates smooth 2–30 second videos at up to 1080p from text or image. It supports start- and end-frame interpolation, image, video, and audio references, and native audio.
Now that you know which models power Air's AI features, check out our guide on writing effective prompts to get the best possible results from your generations.