Video Generation — Sora, Veo, Wan

The 2024-2025 wave of frontier video-generation systems — OpenAI's Sora, Google's Veo, Veo 2, Veo 3, Alibaba's open-weights Wan, plus competitors like Runway Gen-3, Kling, Hailuo, Hunyuan Video — moved generative video from "research demo" to "shipping product". By mid-2025, sub-minute photorealistic video from text prompts was a paid consumer feature across multiple platforms. This page summarises the post-Sora generation; for the architectural baseline see Sora.

Veo line — Google

Google's video-generation answer to Sora:

• Veo 1 (May 2024) — announced shortly after Sora, with comparable demos. 1080p, up to 1 minute.
• Veo 2 (December 2024) — substantial quality and physics improvement. Available in Gemini and Vertex AI.
• Veo 3 (May 2025) — adds synchronised audio generation. The model produces video and audio (dialogue, music, ambient) jointly, eliminating the previous gap where video models needed a separate audio pipeline.

Veo's architectural lineage matches Sora's: latent-space DiT operating on spacetime patches. Google's training-data advantage (YouTube) gives Veo a recognisable strength on real-world physical scenes. Veo 3's joint audio-video generation was a practical headline — the first frontier video model with built-in soundtracks.

Wan — open-weights frontier

Wan 2.1 (February 2025) and Wan 2.2 (mid-2025), released by Alibaba's Wan-AI team, are the strongest open-weights video models. Architecture:

• 14B-parameter Diffusion Transformer.
• 3D causal VAE for spacetime-patch latent compression.
• Native support for image-to-video, video-to-video, video editing, plus pure text-to-video.
• Apache 2.0 licence — fully open including weights and code.

Wan 2.1's release was a Sora moment for the open-source community: 14B parameters runnable on a high-end consumer GPU (with quantisation), generating 480p-720p video at quality competitive with paid closed-source services. By mid-2025 the Wan ecosystem had spawned a community of LoRA fine-tunes, control adapters, and pipeline integrations comparable to Stable Diffusion's image-side ecosystem from 2022-2023.

Other 2024-2025 systems

The video-generation market is unusually crowded:

• Runway Gen-3 (June 2024) — strong cinematographic style, the favourite of professional creatives at launch.
• Kling (Kuaishou, June 2024) and Kling 1.5 / 2.0 — Chinese leader, frequently top-rated on user comparisons.
• Hailuo (MiniMax) — strong stylised content.
• Hunyuan Video (Tencent, Dec 2024) — open-weights competitor to Wan, ~13B parameters.
• Luma Dream Machine (June 2024) — competitive consumer-facing product.
• Pika — long-running competitor, repositioning around editing rather than pure generation.
• LTX Video (Lightricks) — fast, real-time-ish generation at lower quality.

By mid-2025 there were ~20 viable video-generation systems. Compare to the image-generation market in 2022 (4-5 viable systems) — the video market in 2025 is at a similar phase of broad accessibility but pre-consolidation.

Capabilities by mid-2025

What works reliably:

• 5-30 second clips at 720p-1080p quality from text prompts.
• Image-to-video — animate a still image into a short clip. Often more reliable than text-to-video.
• Style consistency within a single clip.
• Camera-motion control via prompt or explicit camera-path tooling.
• Audio synchronisation (Veo 3) — dialogue, sound effects, music.

What's still hard:

• Long-form coherence. Multi-minute outputs drift; characters change identity.
• Editing — making a specific change to an existing video without disrupting other content.
• Physics on hard cases — fluids, soft bodies, complex object interactions.
• Text rendering inside video — slowly improving but unreliable.
• Subject persistence across cuts — even short montages require handholding.

Architectures: DiT-on-spacetime is universal

Every frontier system uses some variant of:

• VAE compression to a spacetime latent.
• Diffusion Transformer (DiT) on patch tokens.
• T5 or CLIP text encoder for conditioning.
• Classifier-free guidance.
• Multi-stage training (image, image + low-res video, full-res video).

Architectural diversity has decreased compared to image generation — the DiT-on-spacetime template won decisively. What differs is data, scale, and post-training.

What this enables

• Stock-video replacement. Many B-roll, advertising, and social-media use cases that previously required filming or stock licensing now use generated video.
• Low-budget creative production. Indie filmmakers, music-video creators, content creators using these tools as part of their pipeline.
• Rapid prototyping of cinematography and shot framing before live shoots.
• Provenance and safety concerns — deepfakes, misinformation, non-consensual intimate imagery. Watermarking, detection APIs, and platform-side moderation are catching up.

What to read next

• Sora — the architectural ancestor.
• Latent Diffusion — the recipe these scaled.
• Image and Video Generation (CV) — broader generative-video context.