Sora2 视频生成同步 (Grsai)

Sora2 is OpenAI's closed video line, and this node is the "just give me a video" version of it inside the ComfyUI-nkxx pack: you type a prompt, optionally attach an image, hit Queue, and ComfyUI sits there polling the Grsai reseller API until a finished MP4 lands in your output folder. It's synchronous, it shows a real progress bar, and when it's done you've got three outputs instead of a task ID and a prayer.

The inputs that matter:

• prompt - multiline, default "A cute cat playing on the grass". Sora responds well to full cinematic sentences, not tag soup.
• aspect_ratio - 16:9 or 9:16. Two choices, that's it.
• duration - 10 or 15 seconds.
• size - small or large. Cost and resolution knob; there's no visible 4K toggle here.
• image (optional) - attach one and it becomes image-to-video (the node uploads it to Grsai and passes it as the start frame).
• remixTargetId - optional video-remix target ID (s_xxxx). This is "make a sequel to that other video I generated" - you'd grab the PID from a previous run's response. Leave "默认无" to ignore it.
• api_key - blank uses the pack-wide Grsai key (env var GRSAI_KEY or the global key manager).

The outputs are the nice part: video is a proper ComfyUI VIDEO (VideoAdapter-wrapped local file), video_url is the remote URL so you can re-download or share it, and response is a status blob with the task ID, the PID (which you'll want for remixTargetId later), and your remaining credit balance.

Mechanically it's honest about being a poller: it posts to /v1/video/sora-video, then loops /v1/draw/result every ~3 seconds for up to 900 seconds, driving the progress bar the whole way. Because IS_CHANGED returns a fresh timestamp, it re-runs on every queue execution - so don't leave it in a graph you're continuously queueing unless you actually want to burn credits each pass.

The whole pack installs the same way:

cd ComfyUI/custom_nodes
git clone https://github.com/jieg9341-lab/ComfyUI-nkxx

Restart; first boot auto-installs requests, pandas, yt-dlp, opencv-python and friends. You need a Grsai key from nkxx.grsai.ai - no key, no request, and the node tells you exactly that in the response string.

Grounded gotchas:

• This node blocks your graph. That's the point of "同步". A 15-second Sora2 clip can take minutes; if you want to keep the queue moving, use the async trio (Sora2SubmitAndRecordTask_Grsai → Sora2QueryTasks_Grsai → Sora2GetNextVideo_Grsai) instead.
• The download is yt-dlp first, curl fallback, three retries. If both fail you still get the URL in video_url - grab it manually rather than re-queueing and paying twice.
• Every run costs API credits. The node helpfully shows the remaining balance, which is the pack's way of making you feel the pain.
• removeWatermark is baked into the request payload, so outputs come back clean - a real advantage over some reseller channels that stamp their logo on everything.

For a one-off clip this is the node you reach for. For anything batch or background, the async set is strictly better, and that's what the author's own numbered workflows use.