Nano Banana 2.1 and Hailuo 3.0 paired up: video from a single photo

The task breaks down into two steps: if you don't already have a frame, create one in Nano Banana 2.1, then bring it to life in Hailuo 3.0. This refers to Google's release from October 6, 2026 (the gemini-nano-banana-2.1 model) and MiniMax's video service, officially launched July 31, 2026 under the name MiniMax H3 (listed as Hailuo Minimax 3.0 on aggregators and Hailuo-03 in the API). These are all names for the same model.

Nano Banana 2.1: generation built for motion

Nano Banana 2.1 is Google's update to Nano Banana 2, positioned as a more cost-effective alternative to Nano Banana Pro. It's designed for image generation and editing.

Capabilities:

• Three resolutions: 1K, 2K, 4K (1K by default)
• Conversational editing: you can ask it to change a single detail while leaving everything else untouched
• Up to 14 reference images for setting style
• Preserves appearance for up to 4 characters and up to 10 objects
• Three reasoning levels before generation: minimal, medium (default), high
• SynthID watermark on all outputs

A frame created in Nano Banana 2.1 should already contain a clear phase of action: a pose captured mid-step, hair or fabric caught by the wind, an object balanced on the edge of a table. A frame like this answers half the timing questions on its own.

Hailuo 3.0: video from a frame with parameters

The model understands context from text, images, video, and audio, and generates video with stereo sound. According to the official announcement, its strengths include precise instruction-following, accurate rendering of text and brands, and motion transfer from a reference video. The developer acknowledges that image detail still needs improvement in some scenarios.

Technical parameters:

• Video: 4–15 seconds, whole numbers only
• Resolution: 768P or 2K
• Stereo sound support
• Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
• Supported image formats: JPG, JPEG, PNG, WEBP, HEIC, HEIF
• File size: up to 30 MB
• Image dimensions: 256–5760 px
• Aspect ratio range: 0.4–2.5
• Prompt length: up to 7,000 characters
• In image-to-video mode, the video matches the source image's proportions

Motion physics: data from an independent study

The independent Physion-Eval study (Physion Labs, Stanford, MIT, Harvard; preprint arXiv 2603.19607) tested five video models: Sora 2, Veo 3.1 Fast, Kling 2.5, Hailuo 2.3, Wan 2.2. Experts with physics backgrounds flagged errors across 1,719 third-person clips and 423 first-person clips. The conditions closely mirror the photo-to-video task: a real video's first frame plus its text description.

Hailuo 2.3 results, third-person: 93.1% of clips with at least one glitch, 1.42 glitches per clip, severity on a 1–5 scale: 3.61.

Average across five models, third-person: 83.3%, 1.28, 3.10.

Best result (Kling 2.5): 73.8% of clips with glitches.

First-person, Hailuo 2.3: 92.0% of clips with glitches, 1.92 glitches per clip, severity 3.86.

First-person average: 93.5%, 1.56, 3.32.

Common errors: broken temporal consistency (details change color or shape without cause), mismatched material or state, and loss of object permanence. Calm processes (flames, sunsets) look convincing. Models break down on scenes involving contact, irreversible changes, and abrupt transitions: impacts, cutting, shattering, spilling.

Four working rules

Supply the missing time axis through visible verbs and chains of action: bumped — tilted — fell — hit — rolled — splashed. Keep a single dominant mechanism per clip. Choose frames with a clear motion phase. Use images with sharp edges, a single light source, and minimal small text.