Short video tools can already turn a sentence into a clip. The harder problem is control.
An uploaded still can do two different jobs, and they are easy to mix up because both start the same way: users attach a picture. In one job, the picture is the first frame. Composition, lighting, and subject placement are already decided, and the model only has to invent the motion that follows. In the other, the picture is a visual constraint. A face, a product, a garment, or a place is supposed to stay recognizable while the camera and staging are free to change.
Those two uses pull the model in opposite directions.
Most systems that animate photographs treat the upload as frame one. That is predictable when the still is already well composed. It is less useful when the same character or object needs to appear in a new shot that does not begin in the same framing.
Reference-style generation exists for that second case.
xAI's own documentation draws the line in those terms: reference images "incorporate specific people, objects, clothing, or other visual elements without locking the first frame (unlike image-to-video)."
Grok Imagine has supported both behaviors in its video models. The distinction was clearer in developer documentation than in the consumer product. The recent interface change is that, once an image is attached to a video generation, the product asks which job the image should perform.
On August 25, 2026, Kara (@karaebel), who lists herself as working on video content design for Imagine, posted a screen recording of the new menu:
When making a video in @imagine, you can now choose whether your uploaded image is a first frame or just guides the video.
The labels in that clip are "First frame/Image starts the video" and "Reference/Image guides the video."
The recording keeps the same uploaded portrait on screen while prompt suggestions cycle through storyboard, character, landscape, and product shot.
The still does not change. The instruction to the model does.
Kara also described a related path she called an "omni-reference video model," which she said allows multiple pictures or characters to be uploaded and animated into one scene.
She also noted practical limits in the same conversation: video upload is not available in that flow, a frame can be extracted from a clip already in Imagine’s canvas, and separate clips can be stitched together in canvas.
Those comments describe the current product surface; they are not a second set of API specs.
First-frame mode, which maps to image-to-video in the API, keeps the uploaded still as the opening composition. Later frames are generated forward from that locked starting point. xAI’s files documentation even names the input that way: "Image-to-video from a stored first frame."
On Grok Imagine Video 1.5, text-to-video and image-to-video are the paths documented with native 1080p. Text-to-video is described as generating a first frame internally, then animating it, without returning the intermediate image. Clip length on Video 1.5 is documented up to 15 seconds. Audio is generated in the same pass as the picture.
Reference mode does not lock the first frame.
Official docs allow up to 7 reference images per request, plus optional preset voices on Video 1.5. Each reference can be a public URL, a base64 data URI, or a stored file_id.
The July 31 note describes the intended split of labor as "each reference image locks one thing in place — a face, a product, a location," so a character can be kept while the scene changes, or the reverse.
The same documentation places a lower resolution ceiling on this path: reference-to-video is capped at 720p, and it cannot be combined with image-to-video or video editing in a single request.
Only one mode is active at a time. That constraint matters because a user who wants both an exact opening frame and additional style references has to choose which job the image will perform. The official Imagine account has said the same thing in public.
The toggle does not invent a new model.
It exposes a split that already existed between image-to-video and reference-to-video. Image and voice references, text-to-video, and native 1080p for the supported modes were announced on July 31, 2026, after Video 1.5 itself shipped in June.
Early replies to Kara’s first-frame/reference demo were limited.
One user generated the same source twice and preferred the guide setting because it left the model more room to change staging. Other comments simply noted that the video tool had continued to add controls. Those reactions do not amount to a systematic comparison.
The same tension appears across other video systems released through 2026.
Some models accept a first frame and a last frame and invent the transition between them. Some accept multiple subject references but still treat a single still as the opening picture. A smaller set offers a dedicated reference path that is not tied to frame one. Grok Imagine now presents the two options as an explicit choice inside one upload flow rather than as separate product names.
Whether that produces better clips depends on the source image, the prompt, and the usual limits of short generative video: brief duration, variable consistency across cuts, and audio that is synthesized rather than recorded. The update is a change in how the existing modes are offered, not a claim that either mode solves those limits.
