Grok Imagine multi-reference: up to 14 images
Grok Imagine Video 1.5 references: 7 at the 31 July 2026 launch, 14 in xAI's docs now, up to 720p, preset voices only. Versely's reference row takes 7.
Grok Imagine Video 1.5 can build a clip from several reference images at once. Each one holds a face, a product or a place. xAI launched references on 31 July 2026 with "up to seven references per generation". On 16 October 2026 its docs list up to 14 images per request on 1.5, at up to 720p. On Versely, the Reference to Video row takes up to 7.
Grok Imagine built its early reputation on being fast and a little unhinged: great for quick, chaotic image riffs, less trusted for anything that needed to hold a character together across more than one shot. The 31 July update changed that. The headline feature isn't the resolution bump. It's the reference system, which answers the question every AI ad and story creator eventually hits: how do I keep this specific thing the same while everything else in the scene changes?
What shipped on 31 July 2026
xAI's announcement and its release notes for 31 July list three additions to Grok Imagine Video 1.5:
- Text-to-video. "Describe the shot — no starting image needed."
- Native 1080p for text-to-video and image-to-video. Reference-to-video tops out at 720p, per the docs' Reference-to-Video guide.
- Reference-to-video, with image references and optional preset voices. The announcement said "up to seven references per generation".
The docs, read on 16 October 2026, now list reference images as "Up to 14 per request on grok-imagine-video-1.5; up to 7 on grok-imagine-video". The release notes don't say when the limit rose. Either way it is a large reference budget for a video model, and it turns Grok Imagine from a single-shot generator into something closer to a scene-continuity tool.
| Item | Detail |
|---|---|
| Image references | Up to 14 on 1.5 (7 at launch), up to 7 on the original model |
| Voice references | Up to 3; preset voices in the public API |
| Resolution | Up to 720p with references; native 1080p for text and image to video |
| On Versely | Reference to Video row: up to 7 images, 480p or 720p, 1 to 15 seconds |
Sources: xAI's announcement (31 July 2026), release notes and Reference-to-Video guide, read on 16 October 2026. The Versely row is from its model page on the same date.
Each reference locks one thing, not the whole image
The mechanism is the interesting part. Rather than treating a reference image as "make it look like this," each image in the set can be assigned to lock a specific element - a face, a product, or a location - while everything else in the prompt is free to change. TestingCatalog's coverage of the update puts it plainly: one reference image can hold a face in place, another can preserve a product, and another can define a location, independently of each other.
In practice that means you can:
- Keep a character consistent while swapping the scene around them.
- Keep a location fixed while swapping which character appears in it.
- Keep both character and location locked while changing only the action.
This is character consistency handled at the reference layer instead of through prompt engineering alone, and it's a direct response to the single biggest complaint about AI video across every model in this category: the same "person" looking like three different people across three generated clips.
Voice references: preset voices, not your own sample
xAI's announcement pairs the image references with voices: "Pass in a character image and a voice reference, and both hold — the same face and the same voice in every scene." You send the image and the voice with each request; xAI doesn't describe a saved pairing.
What counts as a voice depends on where you use it:
- xAI's public API takes up to three voice references per request, and they are preset voices picked by
voice_id. The Reference-to-Video guide says "voice references with your own audio files are available to trusted partners on request." So a voice sample of your own is not a public feature. - Grok's apps. On 31 July xAI said image and voice references "start today in the US for SuperGrok Heavy and SuperGrok Plus" on grok.com/imagine and iOS, rolling out to all tiers.
- On Versely, the Grok Imagine Video 1.5 rows don't offer voice references. Clips come with native audio from the prompt.
TechTimes (2 August 2026) described the feature as voice cloning from "a voice sample". That is its wording, not xAI's, and it noted the API version was gated behind a sales call.
For a recurring host or narrator across a multi-scene ad, a held voice removes the need to lay a separate voiceover on top. Today that means one of xAI's preset voices, not a recording of your own.
A reference-discipline walkthrough on Versely
On Versely, references run through Grok Imagine Video 1.5's Reference to Video row, listed with its sibling rows on the Grok Imagine Video 1.5 page. It takes up to 7 images, addressed in the prompt as <IMAGE_0>, <IMAGE_1> and so on, at 480p or 720p for 1 to 15 seconds. It costs 7 credits a second at SD and 12 at HD. Here's a concrete way to use reference discipline for a three-scene product spot inside the AI video generator:
- Reference 1 - the presenter. A clean portrait of your on-camera talent (or an AI-generated consistent face you've used before).
- Reference 2 - the product. A studio shot of the item, isolated, showing the label and shape clearly.
- Reference 3 - the location. A single still of the environment: a kitchen counter, a desk, a storefront.
With all three references loaded, write scene-specific prompts that vary only the action: "The presenter picks up the product from the counter and turns it toward camera" for scene one, "The presenter sets the product down and gestures toward it while speaking" for scene two, "Close-up push-in on the product as the presenter's hand rests beside it" for scene three. The face, the product, and the counter should hold steady across all three generations because each is pinned to its own reference rather than re-described from scratch every time - the actual discipline is one reference per element, not one reference per scene.
This is a meaningfully different workflow than the old approach of writing an increasingly long, increasingly fragile single prompt and hoping the model remembers what you said three sentences ago about the label design.
Long takes: the original Grok Imagine Video, without references
The references above are a Grok Imagine Video 1.5 feature, and its rows run 1 to 15 seconds. For a longer single take, Versely also carries the original Grok Imagine Video, an image-to-video row with 6, 10, 15, 20 and 30 second options. It doesn't take reference images on Versely. It bills 4 credits a second at SD and 6 at HD, from 24 credits for six seconds at SD; its model page lists the full range.
To go past one generation, Grok Imagine Extend continues an existing Grok Imagine clip from a chosen timestamp, 2 to 10 seconds at a time, at 5 credits a second at SD and 7 at HD. It continues the footage you give it; it doesn't take new references or a voice. For the general craft of stitching generations into a longer sequence, see our guide on extending video length, and for everything else in 1.5 (Lite, pricing, keyframes), read Grok Imagine Video 1.5.
Ship the next cut on Versely
Open the matching tool, keep exports clean (no third-party watermarks), and burn captions before you upload. Start with AI video generator, AI music generator, AI text to speech, or AI caption generator. Credit prices stay on the live model and tool pages.
FAQ
How many reference images can I use in one Grok Imagine Video generation? xAI's docs list up to 14 per request on Grok Imagine Video 1.5 and up to 7 on the original model, as of 16 October 2026. The 31 July launch said seven. Versely's Reference to Video row takes up to 7.
Can I use my own voice sample as a voice reference? Not in xAI's public API. It takes up to three preset voices per request, and own-audio voice references are for trusted partners on request. Versely's Grok rows don't offer voice references.
What resolution does Grok Imagine Video 1.5 output? Native 1080p for text-to-video and image-to-video since 31 July 2026, and up to 720p when you use references. On Versely the 1.5 text-to-video and image-to-video rows offer 480p, 720p and 1080p, and the Reference to Video row 480p or 720p.
Is Grok Imagine Video good for long single takes? The original Grok Imagine Video row on Versely runs up to 30 seconds per generation, without references. The 1.5 rows run 1 to 15 seconds.
How is this different from just writing a longer prompt? A long text prompt still asks the model to remember and re-derive every visual detail from scratch each generation. Reference-locking assigns specific images to specific elements up front, which is a more reliable mechanism for consistency than prompt length alone.
Closing takeaway
The multi-reference update is xAI treating Grok Imagine less like a novelty generator and more like a production tool: a face, a product and a location can each be pinned down independently while the rest of the scene stays flexible. Start with three tightly cropped references, one per element you actually need to hold constant, before you touch the prompt itself.
Versely AI Weekly
New AI models, every Thursday.
The image, video, audio and LLM models released that week, what each is good for, and what's new on Versely.