Grok video: infinite canvas guide

This page covers Grok video features currently connected to this site's infinite canvas and video workspace, not the full xAI API. Model visibility and permissions depend on your own API key and its authorized groups. The table does not claim that every upstream combination has been tested with paid generation.

Before you start: configuration and permissions

1. Create your own API key on the main site and authorize a group that includes the video model you need. See Keys and groups. Never share the key or include it in screenshots.
2. Enter that key in the canvas's NekoCloud API configuration, load available presets, and enable only video models actually returned for that key. Advanced custom channels must use the xAI JSON video protocol under an OpenAI channel; ordinary chat examples are not video upload instructions.
3. Select the video model and mode before adding images and choosing duration and resolution. If the new entry or help text is missing, switch to another menu and return, then check your configuration against the matrix below. Do not clear site data: local canvases and settings may be lost.

A successful GET check or a model appearing in the list does not guarantee upstream generation will succeed, or that the key has every video capability. Check the full model ID, group authorization and balance before generating.

Current support matrix

Durations are whole seconds. 480p / 720p / 1080p are output resolution options, not image counts or input dimensions.

Model	Text / single first frame	First and last frames	Reference mode
classic: grok-imagine-video	1–15 seconds; 480p / 720p	Not supported	Up to 7 reference images; 1–10 seconds; 480p / 720p; no explicit first/last frame roles
1.5: grok-imagine-video-1.5	1–15 seconds; 480p / 720p / native 1080p	1–15 seconds; 480p / 720p	Up to 14 reference images plus explicit first/last frames, up to 16 distinct input images in total; 1–15 seconds; 480p / 720p
Lite: grok-imagine-video-1.5-lite	1–15 seconds; 480p / 720p / 1080p upscaled from 720p	Not supported	Not supported; no explicit first/last frame roles either

For 15 seconds with multiple references and pinned endpoints, choose 1.5 + Reference mode + 720p (or 480p), not classic or Lite. Reference and two-frame workflows cannot use 1080p.

First/last frame mode: 0 / 1 / 2 images

• 0 images: text-to-video; describe the shot, movement, lighting and sound.
• 1 image: the image is the first frame; describe what happens next. Single-first-frame output follows the source image's aspect ratio; do not rely on the ratio selector to force a different crop.
• 2 images: 1.5 only, in first-then-last order; sent as image / last_frame, at up to 720p.

This mode is not a multi-reference input. If you exceed the selected model's image limit, change modes or remove unwanted images yourself; the app should not silently take just the first two.

Reference mode: reference images and endpoint roles

References guide people, clothing, products and locations; they do not automatically lock the first frame. This entry requires at least 1 input image. Lite is not supported.

In 1.5 Reference mode, first/last frame image indices refer to the current input image list, starting at 1. Leave a field empty for no pin. The first pin is sent as image, the last as last_frame. Only assigned images are removed from the reference list; every unassigned image is sent to reference_images in its original order. Pinning endpoints does not discard the other references.

• Assign just a first frame, just a last frame, or both.
• You may assign the same image to both roles to try a loop. This uses 1 distinct input image and does not free 2 reference slots.
• At most 14 images may remain unassigned as references. With 15 inputs, assign at least 1 distinct image to an endpoint; with 16, assign 2 different endpoint images. Sixteen inputs with no pins, or with both roles assigned to the same image, are rejected before HTTP submission.
• Upload, drag/drop, paste and asset selection all follow the active mode's capacity. Oversized Grok batches are rejected as a whole with a message, not silently truncated. Switching modes does not delete existing images, but you must resolve incompatible settings before generation.

Example: 5 images, pinned endpoints, all other references kept

Suppose inputs are images 1 through 5, with image 1 as the opening, image 5 as the ending, and the other 3 as character / clothing / location references:

1. Choose grok-imagine-video-1.5 and Reference mode.
2. Add all 5 images and check the current list order. Set the first-frame index to 1 and the last-frame index to 5.
3. Set duration to 15 and resolution to 720p; describe the action and camera movement between the endpoints.
4. Image 1 → image; image 5 → last_frame; images 2, 3, 4 → reference_images. All 5 images are used, not only the two endpoints.

If you need 5 separate character references plus 2 additional endpoint images, add all 7 images. For example, add the 5 references first and the endpoints last, then use indices 6 / 7. The first 5 remain intact as references. Do not delete character references to select endpoints.

Prompt image tags are not role indices

Use the exact official prompt tokens <IMAGE_0>, <IMAGE_1>, etc., including angle brackets and underscores. These start at 0, unlike the UI's 1-based endpoint indices.

Without a first-frame pin, the first reference is <IMAGE_0>. With an image first-frame pin, that pin occupies <IMAGE_0> and the remaining references start at <IMAGE_1> in their retained order. In the 5-image example above, image 1 is the pinned opening; images 2, 3 and 4 correspond to <IMAGE_1>, <IMAGE_2> and <IMAGE_3>. Set the ending through its last-frame role; do not guess its prompt token from the original input index.

For example: “Keep the appearance of the person in <IMAGE_1>, wearing the clothing in <IMAGE_2>, in the setting from <IMAGE_3>, moving smoothly from the first frame to the specified ending.” Tags express constraints; they do not guarantee pixel-identical output.

Mentioning “use this reference video” in a prompt or pasting a video description does not upload a valid video reference. This Grok entry does not accept reference video files.

Validation and switching modes

Before creating an HTTP request, the app rejects unknown models, invalid modes, excessive image counts, out-of-range role indices, invalid / empty durations or resolutions, and unsupported aspect ratios. Valid parameters can still be rejected upstream. Supported ratios are 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 5:2; see the single-first-frame note above for that workflow.

Switching a model or mode preserves existing inputs and settings. It does not silently shorten a billable duration, change resolution or discard references to make a request fit. For example, after switching 15-second references from 1.5 to classic, explicitly reduce duration to 10 seconds or less. After switching 1080p single-frame generation to references / two frames, explicitly choose 720p or 480p. Emptying a field does not restore a valid default automatically; inspect the actual values before resubmitting.

Features not connected, and the Agent boundary

• Audio uploads, video uploads and watermark options are not supported by this Grok generation entry. Generated audio is not the same as uploaded reference audio.
• xAI's keyframes (keyframes), preset voices (reference_audios), video editing and video extension are not connected to this UI yet. An audio upload is not a preset voice; a video upload is not video editing. Generic controls for other models do not establish Grok support.
• The current server-side Agent tool schema does not expose the new endpoint-role parameters. Manual-panel or client-tool support does not imply full Agent support for these roles. Use the adapted manual panel when you need precise endpoint assignments.

Cost, timeouts and troubleshooting

• Before generation, check the Model Plaza for the model, resolution and group's per-second price and multiplier. Check actual charges per request in the main site's Usage records. This page does not hard-code prices, and selecting a model does not make it free. See Billing.
• Failed generation is not guaranteed to be free or automatically refunded. A timeout does not mean no task was created. Check Usage and any existing task status first; with a task ID, query that task instead of submitting creation again. Records may be delayed, and an absent record does not prove no execution occurred. Avoid unlimited retries. See Troubleshooting.
• For authentication / group errors, check your own key's authorization first. For image count, duration, resolution or role errors, fix the parameters first. For support, provide the model ID, mode, input count and role indices, settings, task ID and a redacted error—not your key or private media.

Official references and scope

• xAI Video Overview (raw Markdown)
• xAI Reference-to-Video (raw Markdown)

Reference counts were checked against the official model-specific raw Markdown: 14 references for 1.5 and 7 for classic. An older page's “7 images” must not be applied to every model. Upstream capabilities may be broader than this site's entry; this guide describes the current canvas integration, not an end-to-end paid test of all upstream features, output quality or billing.

简体中文