Replicate has joined Cloudflare
Collections

Generate videos

These models can generate and edit videos from text prompts and images. They use advanced AI techniques like diffusion models and latent space interpolation to create high-quality, controllable video content.

Key capabilities:

  • Text-to-video generation- Convert text prompts into video clips and animations. Useful for quickly prototyping video concepts.
  • Image-to-video generation- Animate still images into video.
  • Inpainting for infinite zoom Use image inpainting to extrapolate video frames and create infinite zoom effects.
  • Stylization- Apply artistic filters like cartoonification to give videos a unique look and feel.

Models we recommend

For cinematic realism and physical accuracy

Runway Gen-4.5 is the top-rated video generation model, ranked #1 on the Artificial Analysis text-to-video benchmark. It produces videos with realistic physics — objects have weight, liquids flow naturally, and fine details like hair and fabric stay coherent across frames. Great for polished, cinematic clips where visual fidelity matters most.

Google Veo 3.1 and Veo 3.1 Fast are strong alternatives with native audio generation. Veo 3.1 Fast is a good pick when you want high quality with quicker turnaround. Note: Veo can sometimes produce slightly overbaked results.

For multi-shot storytelling with audio

Kling Video 3.0 generates cinematic videos up to 15 seconds with native audio — including lip-synced dialogue, sound effects, and ambient sound. Its multi-shot mode lets you define up to 6 connected scenes in a single generation, making it ideal for short narratives, product demos, and ads. Use standard mode for 720p or pro for 1080p.

Kling Video 3.0 Omni adds reference-based generation and video editing on top. Upload reference images to keep character appearance consistent across scenes, or feed in a reference video for style and camera movement transfer.

For fast, audio-rich social content

Grok Imagine Video from xAI generates short video clips with synchronized audio in around 30 seconds. It supports text-to-video and image-to-video with multiple aspect ratios (16:9, 9:16, 1:1), making it a natural fit for TikTok, Reels, and Shorts. It's best for rapid iteration and social content where speed matters more than maximum fidelity.

For realistic, home-video aesthetics

Sora 2 from OpenAI has a distinctive diffusion quality that gives outputs a natural, home-video feel. It's great for realistic scenes where you want footage that looks captured rather than generated. Sora 2 Pro offers higher fidelity for more demanding use cases.

For balanced cost and quality

Hailuo 2.3 from Minimax supports both text-to-video and image-to-video with standard and pro quality tiers. It's a solid middle-ground option when you need good results without paying top-tier prices. Hailuo 2.3 Fast trades some quality for speed.

PixVerse v5.6 is another cost-effective choice with a unit-based pricing system that keeps shorter, lower-resolution videos affordable.

For open source

The Wan video models are excellent open-source options, competitive with many proprietary models. Wan 2.5 is the latest generation, and the fast variants (Wan 2.5 T2V Fast, Wan 2.5 I2V Fast) are among the quickest text-to-video options on Replicate. Try adjusting the number of steps to trade off between speed and detail.

Other rankings

Generative video is a rapidly advancing field. Check out the arena and leaderboard at Artificial Analysis to see what's popular today.

Recommended Models

Frequently asked questions

Which models are the fastest?

The open-source Wan suite (like wan-video/wan-2.1-t2v-480p) is among the faster text-to-video options on Replicate, especially at lower resolutions and shorter durations. Many models also have “fast” variants, like google/veo-3-fast, designed for quicker turnaround.
Note: Faster runs usually mean lower resolution or simpler motion.

Which models give the best balance of cost and quality?

PixVerse v4 offers a strong balance for many use cases. It uses a unit-based system at $0.01 per unit — for example, a 5-second, 360p video costs about $0.30. Hailuo 02 is another good middle-ground option, with both standard and pro modes for different quality levels. Your ideal choice depends on how much resolution and runtime you need and how much you want to spend.

Which models are best for specific use-cases within this collection?

  • For cinematic realism with high resolution and optional audio, try Veo 3.
  • For fast prototyping (short clips, lower res), Wan models or PixVerse v4 work well.
  • For both text-to-video and image-to-video, Hailuo 02 supports both.
  • If you’re on a budget, stick with 480p or 360p outputs to keep costs low.

What’s the difference between key sub-types or approaches in this collection?

  • Text-to-video (T2V): You write a prompt and get a video.
  • Image-to-video (I2V): You provide a still image (or first frame) and animate it. Not all models support this.
  • Quality / resolution tiers: Some models focus on speed and lower res (e.g., Wan Fast), while others aim for higher resolution and richer motion (e.g., Hailuo 02, Veo 3).
  • Open-source vs proprietary: Open models like Wan are cheaper and often faster. Licensed models like Veo 3 offer higher fidelity but can be more expensive.

Which models are good for one common style or output type?

For short, stylized clips (5–10 seconds at lower resolution), PixVerse v4 and Wan models are great picks. They’re fast and relatively inexpensive, making them ideal for concept work, storyboarding, or rapid iteration.

Which models are good for another common style or output type?

If you want high-fidelity motion, longer clips, or more realistic physics, Veo 3 or Hailuo 02 Pro are better options. Hailuo 02 supports 768p in standard mode and 1080p in Pro mode, which makes it a solid choice for more polished results.

What types of outputs do these models produce?

Most text-to-video models generate short video clips (5–10 seconds) at 24 or 30 fps. Supported resolutions range from 360p to 1080p, depending on the model. Some, like Veo 3, can include audio as part of the output.

How much do runs typically cost?

Costs vary by model and resolution:

  • PixVerse v4: about $0.30 for a 5-second, 360p video.
  • Wan models: generally very inexpensive for short, low-res clips.
  • Veo 3 and Hailuo 02: prices vary and aren’t always listed publicly, so check the model page for up-to-date details.
    Generally, you’ll pay more for longer durations and higher resolutions.

How can I self-host or push a model to Replicate?

You can push your own model by packaging it with Cog and deploying it. If you’re working with open-source video models, you can also fine-tune them and publish your version for others to use.

Can I use these models for commercial work?

Yes, but always check the model’s license. Most text-to-video models on Replicate are available for commercial use, but some authors include additional restrictions.

How do I use or run these models?

You can use the Replicate playground or run them programmatically.

  1. Pick a model from the text-to-video collection.
  2. Add your prompt (and optionally an image for I2V).
  3. Run the model and wait for the video to generate.
  4. Download or embed your output.

Any other collection-specific tips or considerations?

  • Start with short durations and lower resolutions to experiment without overspending.
  • If animating a still image, choose a clean, well-framed starting image for better results.
  • Be specific in your prompts — details like camera motion or scene type improve output quality.
  • Not all models handle character consistency or motion equally well; higher-tier models tend to do better here.
  • Compare resolution and duration to match your budget and needs.
  • Check for updates, as text-to-video models evolve quickly and new versions can improve speed and quality.