grok-imagine-video-1.5 on Sume: needs an image, seven fields refused

On Sume, grok-imagine-video-1.5 is an image-to-video model. A request without image_url, first_frame_url or one reference_image_urls entry is refused with "model grok-imagine-video-1.5 requires image_url, first_frame_url, or one reference_image_urls entry." Seven other fields are also refused by name, so a text-only or multi-reference prompt must go to a different model.

The envelope

The catalog lists 480p and 720p, durations from 4 to 15 seconds, no end frame, no reference videos or audios and no audio output. The validation then goes further than the capability flags: a second reference image is refused with a message that says to use image_url for the source frame, and the fields in the table below each return "X is not supported by model grok-imagine-video-1.5."

Shape the frame with the image

Because aspect_ratio is refused for this model, shape the frame with the image itself: crop or pad the picture to the ratio you want before you send it. video_frames extracts a still from a stored clip, which is a handy source for a continuation.

Minimal valid request

The body has a model, a prompt, an image and a resolution. It prints the request and only sends when you set SEND=1, because a valid call is billed.

import os
import requests

body = {
    "model": "grok-imagine-video-1.5",
    "prompt": "The camera drifts right as steam rises from the cup",
    "image_url": "https://example.com/cup.png",
    "resolution": "720p",
    "duration": 6,
}
print(body)
if os.environ.get("SEND") == "1":
    r = requests.post("https://api.sume.com/v1/video-router/generate",
                      headers={"Authorization": "Bearer " + os.environ["SUME_API_KEY"]},
                      json=body, timeout=60)
    print(r.status_code)

When to pick something else

If you need an ending frame, use kling-3, a Seedance model or Wan 3.0, which have end-frame support. If you need references or audio, use a model whose capabilities list them. If you want a text-only prompt, pick any text-to-video model. Grok's role in this catalog is narrow: animate one image, at 480p or 720p.

A few habits avoid the refusals. Keep one request builder per model family rather than one generic builder with every field. Read the model's capability flags from the catalog endpoint before you add an optional field. And log the 400 message verbatim: for this model each refusal names the field, so the fix is usually a single deleted line.

Cost matters as well. The model is a 4 to 15 second image-to-video option at two resolutions, so check the per-second price in the catalog entry against Seedance or Wan before you commit a batch to it; the right choice depends on the look you need, not on the cheapest row.

Finally, because there is no end frame and no reference list, continuity across shots has to come from the images you feed in. Extract the last still of the previous clip and use it as the next image_url, and keep the prompt focused on motion rather than re-describing what the image already shows.