taotaojt9988 / xzai-me-media

Name: taotaojt9988/xzai-me-media
Author: taotaojt9988

Use when the user asks to generate a video, make an AI video, create an image or poster from text, do text-to-speech, or give an agent media generation — e.g. "generate a video of a corgi running in snow", "make me a 15s promo video", "turn this text into speech". Covers Seedance 2.0/2.5 video (10s/15s/30s), Seedream 5.0 images, 7-voice TTS and OpenAI-compatible chat via the xzai.me API, with async polling and local download.

Version

SKILL.md

4.4 KB

|---|---|
| video | Seedance text-to-video / image-to-video | 4–20 min |
| image | Seedream 5.0 text-to-image / image reference | ~1 min |
| tts | Text to speech (WAV) | seconds |
| chat | OpenAI-compatible chat (glm-5.3-flash) | seconds |
| models | List available models | — |
| status <task_id> | Check a video/image task | — |

Video Models (the #1 mistake is here)

`--model`	duration	notes
(omit)	10 s	Seedance 2.0 Fast, the default, cheapest
`seedance-2.0-fast-15`	15 s	must be spelled exactly
`seedance-2.0-mini`	15 s	must pass `--duration 15` too
`seedance-2.5`	30 s	only 30 s; also supports 21:9

**The API silently falls back to the 10-second Fast model if the model name
is wrong or --duration is missing for mini.** The CLI validates the
combination and refuses to submit an inconsistent pair — never bypass it by
calling the raw endpoint with unvalidated values.

Ratios: 16:9 / 9:16 / 1:1 / 3:4 / 4:3 (plus 21:9 for 2.5 and fast-15).
Reference images: pass --images url1 url2 (max 10). One image = first
frame; multiple = multi-reference.

Procedure

Confirm the key works: python scripts/xiaozhi.py models. If this
fails with 401, the key is wrong — stop and check before submitting paid
jobs.
Choose parameters deliberately: video or image; ratio; duration
(video). State them in your reply so the user can correct before you
submit — generation consumes credits.
Submit and wait: the CLI polls automatically and prints progress.
Videos take 4–20 minutes; do not re-submit while a task is pending —
duplicate submissions burn credits. Use status <task_id> in a new call
instead.
Download immediately: the CLI saves output to --out. Result URLs on
the CDN expire; a local copy is the only durable deliverable. Completion
criterion: the file exists on disk and is non-empty.
Report the local path, not the remote URL, to the user.

Image Notes

Output size approximates but does not exactly match the requested ratio
(16:9 came out 2848×1600 in testing). If exact dimensions matter, check
the file after download.
Up to 10 reference image URLs per job via --images.
Styles: pass --style web_image_style_2D (also _3D, _portrait, …).

Pitfalls

Two auth schemes: api.php endpoints use the x-token header;
/v1/* endpoints use Authorization: Bearer. The CLI handles this; if
you ever call the API directly, don't mix them.
Failed tasks auto-refund (state 2), so a failed video doesn't cost
credits — but a duplicate submission while waiting does.
TTS limit: 1–500 characters per call; longer text must be split.
Audio URLs expire in ~24 h.
Context window of the chat model is 1,048,576 tokens — don't assume
small-context limits, but do not send more than ~1M tokens.
Errors are standard OpenAI-shaped JSON; 402 means insufficient credits.

Verification

After any generation, confirm the output file exists and open/inspect it
(e.g. ffprobe for video, image dimensions for images).
python scripts/xiaozhi.py status <task_id> returns authoritative task
state: 0=pending, 1=done, 2=failed (refunded), -1=not found.