Grok Imagine Video 1.5 Official Release: 6-Second 720P Video in 25 Seconds, Audio and Video Generated Together

xAI today (June 17) made a not-so-high-profile announcement: grok-imagine-video-1.5 has ended its preview and is officially available on the xAI API. Counting from the preview release on June 3, this model completed the GA process in just two weeks — unsurprising for xAI’s usual pace, but in the video generation race, this speed is enough to make competitors seem a bit slow.

Just a few months ago, OpenAI had removed Sora from its product line citing “resource constraints,” which most outsiders interpreted as the business not being cost-effective. Veo is still crawling slowly in Google’s own ecosystem, and ByteDance’s Seedance 2.0 topped the Artificial Analysis Video Arena’s image-to-video leaderboard last month. But once Grok Imagine Video 1.5 entered the field, it pushed Seedance out, scoring Elo 1404 ±6, currently at the top of this race.

The official release compared to the preview didn’t change the foundation — mainly engineering improvements to speed, stability, and quota to make it production-ready. For developers, the main concerns boil down to three things:
Can it be tuned? Is it expensive? Is the output usable? Let’s break it down.

One-sentence positioning: Image-to-video with native audio

Grok Imagine Video 1.5 is an image-to-video model — you provide a starting image plus a prompt describing motion, camera angles, and sound, and it outputs a clip up to 15 seconds long, at most 720P, with audio.

The keyword here is “native audio.”

This is not just a gimmick. Most video models on the market — including early Sora and pre-Veo 2 versions — either add audio afterward, omit it entirely, or just provide background ambience. Grok Imagine Video 1.5 does audio and visuals together in a single generation, with aligned ambient sound, effects, dialogue, and lip sync. One inference, one output — eliminating the hassle of post-production audio alignment.

This is possible through xAI’s Aurora autoregressive MoE architecture. Unlike Sora and Veo’s diffusion Transformer approach, Aurora unifies text, image, video, and audio into a single token stream, predicting next-token sequentially. Video frames and audio samples are jointly modeled in the same sequence, making alignment a natural part of the model’s internal state, without needing a separate “lip-sync” module.

This architectural choice is aggressive from an engineering standpoint. Autoregressive video generation has long been criticized as “computationally impractical” — after all, predicting each patch serially means higher latency compared to parallel denoising in diffusion methods. xAI’s solution was to train it on 110,000 GB200 units, using MoE to reduce activation parameters, and heavily optimize inference. The result is that in Fast mode it now takes 25 seconds to produce a 6-second 720P clip — nearly halving the preview version’s 40 seconds.

Three upgrade points, explained

1. Audio–video sync: lip movements now match

When testing the preview version, the most noticeable impression was “finally no need for post-dubbing.” This time xAI highlighted three things:

• Single-pass generation outputs ambient sound, effects, and dialogue together, no longer rendering visuals first and adding audio later;
• Speech clarity significantly improved — the preview sometimes had a muddy “AI voice,” now largely gone;
• Lip sync is now short-film quality — not “close enough,” but matching syllable by syllable.

For developers making character dialogue, talking avatars, or product voiceovers, this is key to lowering post-production costs. Previously you’d use ElevenLabs to generate voice, then SadTalker or Hedra to sync lips — now it’s one API call.

2. Physical realism: people no longer twist, objects no longer float

The easiest place for video models to fail is physics — extra fingers, legs clipping through, clothes floating like underwater — common jokes in the field. Version 1.5’s improvements here include:

• Motion continuity — long shots won’t have characters reset abruptly;
• Weight and momentum — falling objects’ acceleration curves feel real, stride and center of gravity match for walking;
• Secondary motion — clothing, hair, fabrics respond logically with passive movement.

You can best see the difference in demos. Seedance 2.0 excels at beat sync and multi-material composition, but for single-shot physical realism, Grok Imagine Video 1.5 now clearly wins.

3. Speed: Fast mode delivers in 25 seconds

This number is worth highlighting.

6 seconds, 720P, generated in 25 seconds. Meaning if you build a consumer-facing video generation tool, user wait times are close to the “bearable” limit. The preview’s 40 seconds was barely usable, but 25 seconds is on another tier of product experience. Of course, Fast mode has trade-offs — visual quality and detail stability are a bit lower than Standard mode, so for final delivery projects Standard is still better.

API calls: same paradigm as other xAI models

Simply switch the model name to grok-imagine-video-1.5, dropping the preview’s -preview suffix. xAI’s official SDK call looks like this:

import os
import xai_sdk

client = xai_sdk.Client(api_key=os.getenv("XAI_API_KEY"))

response = client.video.generate(
    prompt="Slow cinematic push-in as embers drift across the battlefield, "
           "the helmet's crest stirs in the wind, faint distant horn",
    model="grok-imagine-video-1.5",
    image_url="https://your-host.com/helmet.jpg",
    duration=10,
    resolution="720p",
)

print(response.url)

Details to note:

• duration max is 15 seconds — exceeding this will be rejected;
• resolution accepts 480p and 720p only — no 1080P mode, a slight disadvantage compared to Seedance 2.0;
• image_url is mandatory — this is an image-to-video model, pure text-to-video requires grok-imagine-video.

If you don’t want to integrate directly with xAI, OpenAI Hub already supports grok-imagine-video-1.5 with OpenAI-compatible format — one key can call GPT, Claude, Gemini, DeepSeek, and other mainstream models, with direct access from China without proxies. Code looks like:

from openai import OpenAI

client = OpenAI(
    api_key="your-openai-hub-key",
    base_url="https://api.openai-hub.com/v1",
)

resp = client.videos.generate(
    model="grok-imagine-video-1.5",
    prompt="a girl turns to the camera and says hi with a warm smile",
    image="https://example.com/portrait.jpg",
    duration=6,
    resolution="720p",
)
print(resp.data[0].url)

For domestic teams, skipping xAI account creation, overseas payments, and network instability — just swapping the base_url — shows the value of such aggregation platforms.

Video continuation: extending 15 seconds into a long shot

The 15-second hard limit is a bit awkward — TikTok videos often run 30 seconds and e-commerce voiceovers start at 60 seconds. xAI’s solution is video continuation — generating from the last frame of the previous clip to extend into the next clip, allowing multiple 6–15 second shots to form a complete sequence.

The engineering significance is that character consistency across shots is maintained internally by the model, without manually locking reference images. Combined with “reference-to-video,” you can carry the same character, product, or art style across multiple shots.

A practical workflow:

1. Use grok-imagine-video-1.5 to create the first segment, setting tone and characters;
2. Take the last frame as the starting image for the next clip;
3. Describe shot changes in the prompt (e.g. “cut to a wide shot…”);
4. Finally stitch all clips together — since audio is natively generated, transitions only need light crossfade.

Compared to Sora’s ambitious “two-minute continuous shot” promise, this short-shot continuation to build long clips approach is simpler but workable, cost-control-friendly, and suits modern short-video pacing.

Comparing with Seedance 2.0, Veo, Kling

Native-audio video models aren’t unique to Grok. A quick comparison:

| Model | Vendor | Max Duration | Max Resolution | Native Audio | Focus | | --- | --- | --- | --- | --- | --- | | Grok Imagine Video 1.5 | xAI | 15s | 720p | Yes | Image-to-video, lip-synced dialogue, continuation | | Seedance 2.0 | ByteDance | 4–15s | 1080p | Yes | Multi-material composition, music beat sync | | Veo 3 | Google | 8s | 1080p | Yes | Physical realism, stable quality | | Kling 3.0 | Kuaishou | 10s+ | 1080p | Yes | Multi-shot, capacity-first |

Grok’s shortcoming is resolution — no 1080P — for teams delivering social media-ready video, either upscaling or tolerating lower resolution. Its strength is dialogue capability in the image-to-video route — Seedance leans on beat sync for music-focused use, while Grok targets character voiceover, making their positioning distinct.

As for Sora — the name that once launched the video generation era — it’s now quietly removed from OpenAI’s product line. An era’s symbol gone without fanfare.

Who will use it and how

Some obvious application areas:

• Talking character shorts: upload a portrait, have the person naturally say a greeting, with natural lighting and ambient sound — can be used for customer service, virtual hosts, educational content;
• Product dynamic ads: those e-commerce close-ups like “slow push on bottle with condensed water drops” — previously needed filming and effects, now a product image plus one prompt;
• Music MV segments: native audio can include melody, with video continuation to build a full song’s visual sequence;
• Storyboard dynamic previews: game/film storyboard artists can animate static boards in bulk to show clients.

Previously these required Runway, Pika, or stitching together open-source chains. Grok Imagine Video 1.5 takes the audio step out, shortening the pipeline significantly.

One observation

xAI’s release timing is telling. OpenAI dropped Sora, Google locked Veo in its own ecosystem, and ByteDance’s Seedance’s main battleground leans toward B2B — leaving an open developer market for image-to-video + native audio. xAI pushed from preview to GA in two weeks, making it clear — this product line is for long-term business, not a demo show.

With its entirely different autoregressive path from diffusion competitors, if Aurora can keep pushing resolution, duration, and richer multimodal inputs, the Grok Imagine series will only move further forward in video generation rankings.

For developers, the action today is simple: switch to the new grok-imagine-video-1.5 model name, rerun your demos, and see what 25-second clip generation can do for your product experience.

References

• ITHome: Grok Imagine Video 1.5 Official Release, Generates 6-Second 720P Video in Just 25 Seconds — Chinese coverage of xAI’s official release, including speed comparisons and capability descriptions.