AI video generation

Connection and authentication

Base URL: https://caheoai.com/v1

Authorization: Bearer $STALI_API_KEY

Set STALI_API_KEY in your environment. GET /v1/models checks authentication without creating a video.

read -r -s -p "STALI_API_KEY: " STALI_API_KEY
export STALI_API_KEY
printf '\n'
curl -fsS https://caheoai.com/v1/models -H "Authorization: Bearer ${STALI_API_KEY:?STALI_API_KEY}"

cURL commands use Bash; on Windows, use the PowerShell setup. Replace sample media URLs with your own provider-accessible files.

Creating a video is billable. Save the id from a successful response to track the same job.

If a job fails after an id is returned, refunds depend on provider results and reconciliation. A failed status does not confirm an immediate wallet refund.

Polling, downloading and cancellation do not create another generation. Successful cancellation does not automatically reverse a recorded charge.

Models and capabilities

Five main Seedance models (ByteDance, st1/ line): 2.0-1080p, 2.0-720p, 2.0-fast-720p, 2.5-1080p, 2.5-720p. Billed by the exact seconds you send. Veo / Omni / Grok keep fixed duration price tables.

One image via input_reference/images is the first frame (aspect auto). Use reference_image_urls or omni_reference_task_type=reference to guide content with a fixed aspect ratio.

veo-3.1 accepts only 4, 6 or 8. google-omni-flash accepts only 5 or 10. grok-imagine-video-1.5 accepts only 6, 10 or 15. Check seconds in the response.

Endpoints and parameters

POST /v1/videos — create a job

Send application/json. This endpoint reads JSON, not multipart/form-data; use URLs or image JSON supported by the model.

GET /v1/videos/{id} — track progress

Include ?model= with the original model. Start with a 10-second polling interval and back off on 429. The JSON below is illustrative, not a live result.

• queued: the job was accepted and is waiting.
• in_progress: the provider is processing the job.
• completed: the result can be downloaded.
• failed: inspect error and keep the id for support.

{
  "id": "video_abc123",
  "object": "video",
  "model": "st1/seedance-2.0-720p",
  "status": "completed",
  "progress": 100,
  "seconds": "10",
  "size": "1280x720",
  "error": null
}

GET /v1/videos/{id}/content — download MP4

Download only after completed. Send the same API key and ?model=; save the binary response as .mp4.

DELETE /v1/videos/{id} — request cancellation

Cancellation depends on the route and job state; some routes currently do not support DELETE. Cancellation does not guarantee a refund.

Pricing by duration

Prices follow this website’s current configuration, including applicable adjustments. Check them before creating a video.

Seedance duration pricing. Unit: VND per successful video. 4–15s or 4–30s is the supported duration range, not a flat package price. Choose seconds within the supported range; exact prices are listed below. Any fixed generation fee is already included.

Other video models

Seedance bills by the seconds you send. Veo / Omni / Grok keep fixed duration price tables.

Integration examples

Each POST creates a billable video. Choose one example; example.com media URLs are placeholders and must be replaced.

import os
import time
from getpass import getpass
from urllib.parse import quote
import requests
from requests.adapters import HTTPAdapter

api_key = os.getenv("STALI_API_KEY") or getpass("STALI_API_KEY: ")
headers = {"Authorization": f"Bearer {api_key}"}
model = "st1/seedance-2.0-720p"
http = requests.Session()
http.mount("https://", HTTPAdapter(max_retries=0))
created = http.post("https://caheoai.com/v1/videos", headers=headers, json={
    "model": model, "prompt": "Cinematic sunset over ocean waves",
    "seconds": 5, "aspect_ratio": "16:9",
}, timeout=(10, 120))
created.raise_for_status()
video = created.json()
video_id = video["id"]
print(video_id, video.get("status"), flush=True)
job_url = "https://caheoai.com/v1/videos/" + quote(video_id, safe="")
for _ in range(120):
    if video.get("status") in ("completed", "failed"):
        break
    time.sleep(10)
    progress = http.get(job_url, headers=headers, params={"model": model}, timeout=(10, 60))
    progress.raise_for_status()
    video = progress.json()
    print(video_id, video.get("status"), flush=True)
if video.get("status") != "completed":
    raise RuntimeError(str(video))
content = http.get(job_url + "/content", headers=headers, params={"model": model}, stream=True, timeout=(10, 120))
content.raise_for_status()
with open("video.mp4", "wb") as output:
    for chunk in content.iter_content(chunk_size=65536):
        if chunk:
            output.write(chunk)

Error handling

Do not automatically repeat POST after a timeout or network error: the first request may have been accepted. If you have an id, poll it; otherwise keep the request ID and check Usage or contact support before recreating.

Prompt design

These are creative guidelines, not a guarantee that a provider will reproduce every detail. Each model’s content policy still applies.

Write a clear prompt

1. Define subject, setting and main action before effects.
2. Set duration with seconds, not only in the prompt.
3. Keep left/right positions and outfits consistent across shots.
4. Prefer one clear camera movement per shot.
5. Describe light sources, colors and materials.
6. Avoid small text, logos or exact numbers in frame.
7. Start with one short scene; change one factor at a time after reviewing.

Arrange reference images

• Number Image 1, Image 2… in submission order.
• State which image controls character, outfit or setting.
• Use authorized, clear images with a consistent visual direction.
• Match characters and composition in first/last frames for natural motion.
• Stay within the model’s image limit; references do not guarantee exact reproduction.

Prompt template

确认生成。参数已确定：16:9 横屏，三个连续镜头组成一个完整视频，立即生成，无需再次确认或提问。
The reference images are colored-pencil character design sketches: use them only for design and outfits. Render the video as a photoreal 3D CG game cinematic with dramatic lighting.
Image 1 = PROTAGONIST: [character design].
Image 2 = ANTAGONIST: [opponent design].
Image 3 = ENVIRONMENT: [setting].
FIXED LAYOUT: altar in the background CENTER; protagonist on the LEFT; antagonist on the RIGHT. Nobody disappears, teleports or swaps sides.
LIGHTING: torchlight from the left wall, cool moonlight from the ceiling.
Action: well-choreographed, orbiting camera. No blood, no wounds, no gore.
Sound: heavy impacts, wind, no dialogue.
镜头1: camera pushes in, protagonist draws a weapon, antagonist raises claws.
镜头2: fast exchange, camera orbits.
镜头3: climax, both stay on their sides, protagonist lands ready.
Keep faces and outfits consistent. No text, no numbers, no subtitles, no watermark.

The template guides the scene. Shorten it if actions are missed; do not use prose instead of API parameters.

Describe non-graphic action

Build a sequence

1. Split the story into scenes with clear starts and ends.
2. Reuse character descriptions, lighting and references.
3. Download and review each scene before generating the next.
4. Join scenes in an editor; each POST remains a separate billable job.

When results miss the mark

• Missing actions: reduce events in one scene.
• Swapped positions: specify left/right and stable composition.
• Inconsistency: reuse references and identifying details.
• Wrong duration: check seconds and fixed-duration models.
• Rejected: review error and content policy; do not repeatedly resend the same request.

Review motion with a frame grid

ffmpeg -i clip.mp4 -vf "fps=1,scale=320:-1,tile=6x5" -frames:v 1 frame_grid.png

This ffmpeg command reads local clip.mp4 and creates a frame grid without calling the API. Install ffmpeg first.

Preflight checklist

• Choose the correct model ID and supported duration.
• Check pricing and balance before creating.
• Keep keys server-side, out of public browser code.
• Use JSON and provider-accessible media URLs.
• Save id, model and request ID for tracking.
• Poll the same id with ?model= and back off when limited.
• Download after completed and verify the MP4 file.
• Do not recreate automatically when acceptance is uncertain.
• Review errors and costs in Usage; send the id to support if needed.