🧬 Sequencer Generate

The name is half the pitch: 🧬 Sequencer Generate is a universal generation node, and it really does let you pick any of ~178 models from one dropdown - GPT Image, Flux 1.1/2 Pro, Ideogram, Nano Banana, Veo 3.1, Kling, Wan, ElevenLabs TTS, even Topaz and SeedVR2 upscalers. Here's the part the marketing blurb is quieter about: none of that runs on your machine. This node is a thin client for sequencer.media, a paid cloud API gateway that resells all those closed models. You supply a key and credits, it supplies the GPUs, and your local ComfyUI just watches the polling.

If that sounds like the opposite of what you usually run ComfyUI for, you're not wrong. The local-vs-cloud gap is a real wall in this ecosystem, and this node is squarely on the cloud side: zero local VRAM load, zero model downloads, but also zero privacy - every prompt and image you connect leaves your machine and hits their backend. The honest use case is narrow but real: you want a frontier closed model inside a ComfyUI graph without maintaining half a dozen per-vendor API accounts. Video is where that genuinely earns its keep - the community still has no open-source equivalent to Veo 3.1's native audio (the KB's closed-source-models essay calls it a real capability gap), and nothing local does ElevenLabs-tier TTS. For still images, open weights have basically closed the gap, so the image models in this catalog are "nice to have," not "need."

How it works

The mechanism is straightforward once you know the shape. On startup the node pulls the model list from the Sequencer Firestore registry (it caches for ~5 minutes and refreshes when ComfyUI restarts), so the dropdown is the service's catalog, not yours. When you queue a job it resolves your API key, resolves your workspace ID, uploads any connected images, creates a "media document" on the backend, and then polls for up to ~4 minutes (120 attempts at ~2s each), printing status to the console.

That polling and the outputs are where beginners get burned. For image models the node downloads the result and hands you a real IMAGE tensor you can wire into Preview/Save. For video and audio models, the IMAGE output is a dummy 1×64×64 black tensor (or a thumbnail if the backend happens to provide one) - the actual deliverable is the VIDEO_URL or AUDIO_URL string. If your preview shows a black square after a 90-second video job, that's not a failure; grab the URL string and save it with a Save Text node. One more grounding gotcha: the code base64-encodes input images inline into the request (its own comment admits "this isn't ideal for large images" - proper upload is a TODO), so keep connected source images reasonable.

Inputs that matter

The required trio is model, prompt, and api_key - leave the key blank if it's in your config file and the node picks it up. The ones you'll actually fiddle with:

• model - the whole point. Pick from the live catalog; each entry is tagged like ⭐ [IMAGE] GPT Image 2 (High).
• aspect_ratio - 16:9, 9:16, 1:1, 4:3, 3:4, 21:9. Not every model honors it.
• input_image - connects for img2img, image-to-video, upscale, style transfer.
• duration - video length in seconds (1–30), only used by video models.
• strength - influence of the input image; lower = more faithful.
• reference_image_1/2/3 - subject/style refs for multi-ref or element-to-video models.

Also present: negative_prompt, seed, resolution (auto/1K/2K/4K, for upscalers), and workspace_id_override. One stale-README warning: the README's table calls the key input api_key_override, but the actual node input is api_key - trust the node.

Installing it

ComfyUI Manager → Install Custom Nodes → search "Sequencer", or:

cd ComfyUI/custom_nodes
git clone https://github.com/sequencer-media/sequencer-comfy-nodes

Restart ComfyUI. No pip dependencies - the requirements file is intentionally empty; it uses stdlib plus the torch/Pillow ComfyUI already ships. Then configure:

mkdir -p ~/.sequencer
echo '{"api_key": "sk_YOUR_KEY", "workspace_id": "YOUR_WORKSPACE_ID"}' > ~/.sequencer/config.json
chmod 600 ~/.sequencer/config.json

Get the key at sequencer.media/settings (Settings → API Keys), and find your workspace ID in the dashboard URL after /workspace/. Environment variables (SEQUENCER_API_KEY, SEQUENCER_WORKSPACE_ID) work too.

Common issues

• "No Sequencer API key configured" - you skipped config, or the config JSON is malformed (it'll print a warning). Fix the file or paste the key into the node's api_key input.
• Dropdown shows "(error loading models - restart ComfyUI)" - the registry fetch failed at startup; check your internet, then restart to retry.
• "Generation timed out after 120 polling attempts" - video can take 1–3 minutes; the poll cap is ~4. If it still dies, check the Sequencer dashboard.
• Black IMAGE output for video/audio - not a bug; read the URL output (above).

Bottom line: it's a tidy convenience wrapper, genuinely zero-maintenance once configured, and video/audio access is its real superpower. But it's a paid, lightly-trafficked service (essentially no community discussion of sequencer.media anywhere on the usual boards), so go in knowing your credits and your data both run on a meter. For images alone you probably don't need it - your local box already does that part for free.