SkyReels V4 makes the Veo 3.1 audio-video race less one-sided

Veo 3.1 made audio feel normal in AI video. That matters. For a while, the usual workflow was to generate a silent clip, add a voice track or sound effects later, and hope the timing did not feel pasted on. Google's model changed the baseline for a lot of creators: short clips with sound already inside them.

SkyReels V4 (by Skywork) enters that race from a slightly different angle. It is not only trying to make a pretty short clip. The public paper and current Runware listing both describe it as a unified video and audio model for generation, editing, inpainting, and reference-guided work. That is a broader brief than "make an eight-second video with sound."

The interesting part is not the claim that it has audio. Many models are adding audio now. The useful question is whether audio sits close enough to the visual generation process that it changes how you work.

What actually launched

The SkyReels V4 (by Skywork) paper appeared on February 25, 2026. The Hugging Face paper page summarizes the model as a dual-stream system: one stream synthesizes video, another generates time-aligned audio, and both share text understanding. The same page lists the headline limits: up to 1080p, 32 frames per second, and 15 seconds.

Runware listed SkyReels V4 (by Skywork) as live on April 24, 2026. Its model page presents the same basic shape: text, images, video clips, masks, and audio references can guide the result, depending on the workflow. The supported outputs are short clips at 480p, 720p, or 1080p, with a duration range from 3 to 15 seconds and five seconds as the default.

That timing matters because the public conversation around AI video has been moving fast. Artificial Analysis posted that SkyReels V4 (by Skywork) took the top spot in its Text to Video With Audio arena, ahead of Kling 3.0 and Veo 3.1 at that point. Skywork reposted that result. I would treat it as useful market signal, not a permanent ranking. Video leaderboards move, and the prompt set behind a ranking may not match your project.

Audio is the point, but not the whole point

The model's paper makes a stronger claim than "sound effects are available." It describes joint generation of video and audio, with audio references able to guide sound while visual inputs guide the image stream. In plain English: the system is designed to think about timing, motion, and sound together rather than producing a silent video and bolting audio on afterward.

That does not mean every result will sound good. AI audio can still be tacky. Footsteps may be close but not quite right. Ambient sound may match the scene while feeling too generic. Lip sync can pass at social feed speed and fall apart under close inspection.

Still, a model that accepts audio as part of the generation problem gives you a different starting point. You can test a rain scene where the drops hit the window at the right moment, a product spin with a clean mechanical click, a talking character that needs the mouth to land near the voice, or a music-driven cut where motion should follow the beat.

Veo 3.1 is not weak here. Google's model is polished, easy to explain, and widely understood as an audio-capable video model. Where SkyReels V4 (by Skywork) gets interesting is the combination: audio, video references, editing, masks, multiple aspect ratios, and a longer 15-second ceiling in one model family.

The practical spec sheet

The supported duration range is short but useful: 3 to 15 seconds. That covers most social ad beats, product reveals, background loops, motion tests, and single-shot concepts. It does not cover a full narrative scene unless you are stitching clips together.

Resolution support runs from 480p through 1080p. The 1080p ceiling is important for review, but I would not start there unless the prompt is already working. The 480p and 720p options are better for learning how the model behaves. Once the subject, motion, and audio timing are close, then 1080p makes sense.

Aspect ratio support is broad enough for normal production work: landscape, portrait, square, and the common 4:3 and 3:4 shapes. That sounds mundane, but it affects how quickly a clip can move from test to publishable asset. A model that only behaves well in one shape creates extra editing work.

The current Runware listing also separates plain text or image generation from video-guided work. Video-guided generation costs more, which is fair if the model has to preserve or transform existing motion. It also means you should not use video-to-video casually. Use it when the source clip contains motion, framing, or character behavior worth preserving.

Pricing in normal language

Runware currently lists SkyReels V4 (by Skywork) pricing per generated second. For text or image prompts, 480p is about eleven cents per second, 720p is about fourteen cents per second, and 1080p is about thirty-five cents per second.

Video-guided work costs more: about eighteen cents per second at 480p, twenty-five cents per second at 720p, and a little over sixty-two cents per second at 1080p.

That makes a five-second 720p test from text or image roughly seventy cents. A five-second 1080p test is about one dollar and seventy-five cents. A full 15-second 1080p video-guided run is much more expensive, a little under nine dollars and forty cents.

The pricing suggests a sane workflow: test short and low first. If the composition works, increase duration. If the timing and motion hold up, move to 1080p. Paying for a 15-second 1080p edit before the idea is proven is usually wasteful.

Artificial Analysis described SkyReels V4 (by Skywork) in a minute-rate frame, roughly seven to eight dollars per minute depending on whether audio is included in that comparison. That is helpful for market positioning, but for actual use I would look at the per-second price in the tool where you are generating.

Where it can beat expectations

The 15-second ceiling is the first real advantage. Eight seconds is enough for a clean visual beat. Fifteen seconds gives you space for a setup, a motion change, and a payoff. That extra room matters for product demos, short ad variations, and anything with a sound cue that needs a little breathing room.

The second advantage is reference control. SkyReels V4 (by Skywork) can use images, video, and audio references depending on the path exposed by the provider. That is useful when the prompt alone is too vague. A product photo can pin down the object. A video reference can suggest motion. An audio reference can push the result toward a particular rhythm or vocal feel.

The third advantage is editing. The paper treats generation, inpainting, and editing as part of one system. That is a better fit for real creative work than a model that only creates fresh clips. Most teams do not need one perfect first pass. They need a way to steer a rough pass into something usable.

AI 動画生成

テキストや画像から動画を作成、既存素材も変換可能

→