Google Veo 2 Image-to-Video Generation: Capabilities, Limits, and How to Access It

You have a still image and need a short video clip from it. Google Veo 2 image-to-video generation accepts a JPEG, PNG, or WebP file up to 20 MB and 8192×8192 pixels, as documented by Google Cloud on 2026-09-26, and returns a 24 fps H.264 MP4 clip. The maximum native clip length is 8 seconds. Usable motion coherence drops after roughly 4 to 5 seconds. Access runs through two separate paths: VideoFX via waitlist and Vertex AI via API. Here are the exact capabilities, the documented limits, the specific failure modes you will hit, and how to get access now.

Veo 2 outputs silent video only. It does not generate audio for image-to-video clips. The native resolution reaches up to 4K (3840×2160 pixels), but the effective detail is closer to 1080p upscaled due to the latent-space compression in the generation pipeline. The claimed versus real resolution gap is about one tier: a 4K output file analysed frame-by-frame shows visible softening and hallucinated detail, especially in fine textures like hair or fabric patterns. This is documented in the Google Cloud Vertex AI documentation dated 2026-09-26.

Supported aspect ratios are 16:9, 9:16, 1:1, 4:3, and 3:4. Input images outside these ratios are resized during pre-processing, which causes aspect-ratio distortion. Subjects stretch or compress unnaturally if your source image is, for example, a 5:4 crop from a camera sensor. Crop your source image to one of the five supported ratios before uploading.

Clip Length: Claimed Versus Real

Google documents a maximum native clip length range of 4 to 8 seconds for image-to-video. The 8-second figure is the claimed maximum. Motion coherence degrades visibly after 4 to 5 seconds. The cause is the model's temporal attention losing the conditioning signal from the source image over longer sequences. Subject drift, limb mutation, and background sliding all accelerate after the 5-second mark. Plan your clips for 4 seconds if source fidelity is your priority. Use 6 to 8 seconds only when you accept that the final portion will show artefacts.

Veo 2 does not expose a discrete motion-strength slider or a separate end-frame parameter. The start frame is your input image. The output is generated from that single reference frame. You control camera movement by describing it in the text prompt: pan, zoom, dolly, tilt, crane, or orbit are all possible. The model then predicts pixel motion from its training distribution. There is no persistent 3D world model. The result is a statistical guess at what motion looks like, not a physical simulation.

Prompt adherence is strong for simple camera moves. It is weak for complex subject actions. If you prompt "a horse galloping left to right," the motion may collapse into a nearly static clip because the model biases toward low-motion outputs that produce fewer visible artefacts during training. This is a structural consequence of the trade-off between motion magnitude and source fidelity: stronger motion always increases the risk of subject drift and artefacting.

What Breaks and Why

The failure modes are not bugs. They are consequences of the single-reference-frame limitation. Subject drift transforms your subject into something else over the clip duration. Limb mutation warps, multiplies, or detaches arms, legs, or fingers because the model lacks any skeletal model. Background sliding moves the background independently of the foreground as if on a separate layer, caused by the model failing to infer a consistent 3D scene structure. Identity shift changes a face's ethnicity, age, or gender as the model interpolates toward its training data distribution rather than anchoring to the source. Flicker produces rapid frame-to-frame changes in brightness, colour, or texture from inconsistent latent-space decoding across frames.

Veo 2 motion fidelity is high compared to open-weights models like Stable Video Diffusion or LTX-Video, especially on camera-motion control. The camera moves described in your text prompt translate to smoother, more temporally consistent results. The model cannot overcome the fundamental architecture limitation: temporal attention across 8 seconds from a single image means the model loses track of the source image. The practical takeaway: Veo 2 gives you good control over camera motion but poor control over subject motion. Design your shots around camera movement, not character action.

Animating Real People

If your source image depicts a real, identifiable person, obtain explicit consent before generating and publishing a Veo 2 clip of them. The model can produce outputs that alter a person's appearance or place them in fabricated contexts. Publishing such clips without consent may violate privacy laws, platform policies, or both. This applies even if the clip is short or the motion is minimal.

Attribute	Documented Value	Real-World Note
Input format	JPEG, PNG, WebP	RGB colour space required
Input size limit	20 MB, up to 8192×8192 px	Larger files increase pre-processing time
Output clip length	4–8 seconds	Useable coherence ~4–5 seconds
Output resolution	Up to 3840×2160 (4K)	Effective detail closer to 1080p
Frame rate	24 fps	Standard for downstream editing
Aspect ratios	16:9, 9:16, 1:1, 4:3, 3:4	Crop input to avoid distortion
Audio	None	Silent output; add audio in post
Watermark	SynthID (invisible metadata)	Applied by default per Google DeepMind 2024
Video codec	H.264 in MP4 container	Standard delivery format
Motion control	Text prompt only	No motion-strength slider or end-frame parameter

Google Veo 2 access runs through three routes: VideoFX, Vertex AI, and YouTube Dream Screen. The distinction matters for your workflow and commercial terms.

VideoFX: The Web Interface

VideoFX is a Google Labs product accessed via waitlist. As of 2026-09-26, the waitlist is open for registration. VideoFX provides a web interface with no API. Upload your image, enter a text prompt describing the desired camera motion, and receive the output. The moderation strictness is high: the safety filters block outputs that violate the Google Generative AI Prohibited Use Policy (2025 version). Benign source images are sometimes rejected as false positives due to overly broad classifiers that flag skin tones, certain object shapes, or patterns resembling prohibited content. Budget for rejected images and prepare alternative sources.

Vertex AI: The Programmatic Path

Vertex AI gives you the same model through the Google Cloud Platform API. This is the commercial access path. Use is permitted under the Google Cloud Terms of Service and the Generative AI Acceptable Use Policy. The default rate limit is 2 requests per minute per project, but this varies by your quota tier. Use Vertex AI to integrate Veo 2 into automated pipelines: batch job processing, integration with editing software, or client-facing tools. The pricing structure is pay-per-request with no unlimited tiers. Google Cloud does not publish live retail prices in the documentation, so consult the Vertex AI pricing page before building workflows.

Commercial use is permitted via Vertex AI. There is no attribution requirement, but Google recommends disclosing AI use. All Veo 2 outputs receive a SynthID invisible metadata watermark by default per Google DeepMind documentation (2024). This metadata watermark is embedded in the file header. A human cannot see it. A tool can read it. Platform detection systems may not read that specific metadata format. YouTube's labelling policy for realistic AI-generated content (effective 2024-03-18) requires a label, but YouTube uses its own detection system. The SynthID watermark does not automatically satisfy platform compliance. Check each platform's current detection method before uploading.

YouTube Dream Screen: The Limited Path

YouTube Dream Screen is a limited-access feature for select YouTube creators. It is not a general-access tool. If you are not a YouTube creator in the beta program, ignore this path. VideoFX and Vertex AI are your options.

If you publish Veo 2 outputs in the European Union, the EU AI Act imposes transparency obligations. Article 50(2) requires labelling of AI-generated content from 2 August 2026. Article 50(4) requires disclosure of a summary of the training data used, also from 2 August 2026. Google DeepMind has not published a standalone model card for Veo 2 as of 2026-09-26. The training data summary required by Article 50(4) is not yet available. Monitor the Google Cloud documentation for the model card release.

In the United States, the US Copyright Office position (policy statement, 16 March 2023, reaffirmed in registration guidance, 2025) states that AI-generated works without human authorship are not copyrightable. A Veo 2 clip generated from your source image may qualify for copyright protection only if your creative choices in selecting, framing, and prompting the image constitute sufficient human authorship. The legal boundary is not settled. Ongoing litigation shapes the case law: Andersen v. Stability AI (US District Court, N.D. California, case no. 3:23-cv-00201) and New York Times v. OpenAI (US District Court, S.D. New York, case no. 1:23-cv-11195).