🎬 Agnes Text-to-Video

Local video generation has a reputation problem that is earned: Wan and LTX are brilliant, and both will happily eat your VRAM, your disk, and several minutes of your life per clip. Agnes_Text2Video is the opposite side of that trade - it turns a prompt into an actual video clip using Agnes AI's cloud agnes-video-v2.0 model, then drops the result back into ComfyUI as frame tensors, audio, and a URL. No model download, no VRAM budget, just a prompt and an API key.

If that sounds like cheating to a die-hard local generation crowd, sure. But for iterating on ideas before committing to a heavy local render, a hosted video API is unbeatable: you get real results in your graph, and the only cost is whatever Agnes charges per clip.

How it works

Same architecture as the pack's image-to-video node, minus the image. It POSTs {prompt}, model: "agnes-video-v2.0", width, height, num_frames (parsed from the duration preset), frame_rate, and num_inference_steps to {base_url}/videos. The API returns a task ID immediately, and the node polls until the job is done - printing progress to the console every ten seconds, up to a 15-minute ceiling.

When the video lands, three things happen in sequence:

1. The original mp4 is downloaded into ComfyUI's output/ folder (look for agnes_t2v_*.mp4).
2. Every frame is decoded with OpenCV into an IMAGE tensor batch shaped [N, H, W, C].
3. The audio track is extracted into the AUDIO format - via VideoHelperSuite's lazy_get_audio if it's installed, otherwise through ffmpeg (from imageio-ffmpeg or your system).

You get all three outputs - images, audio, url - which is exactly what VHS_VideoCombine wants to build an in-app preview. Wire images and audio into VHS and you're watching your clip inside ComfyUI instead of digging through the output folder.

The inputs that matter

• config - from Agnes_Config. Non-negotiable.
• prompt - describe the shot. The default eagle-over-mountains is a demo prompt; treat it as a placeholder.
• duration - five presets from 3.4s (81 frames) to 13.4s (321 frames), default 5.0s (121 frames). The frame counts fit the 8n+1 constraint video models expect, and the node clamps to it.
• width / height - 512 to 1920, stepping by 64, default 1152x768. Bigger is slower and pricier; start at the default.
• frame_rate - 12 to 60, default 24.

Optional: negative_prompt (defaults to a generic quality list), seed (-1 = random), and num_inference_steps (10–100, default 30). With a hosted model, the inference steps slider is your main quality/price dial - the high end is genuinely slow, and your credits notice.

Install

Via ComfyUI Manager (search "AgnesAI") or:

cd ~/ComfyUI/custom_nodes
git clone https://github.com/Watchcats211/ComfyUI_AgnesAI_Nodes.git

Restart ComfyUI. Core deps are torch, numpy, and opencv-python - standard in any working install. For the in-app preview workflow (the good one), install ComfyUI-VideoHelperSuite too; without it you get the mp4 in output/ but no preview node to feed.

Common issues

• The graph "hangs" for minutes. That's the async poll doing its job. [Agnes] [120s] processing (45%) in the console means it's alive; only worry if the console goes quiet without a progress line.
• Empty frames out. Any failure - bad key, dead endpoint, task error - returns a blank 64×64 image batch and empty audio. The console has the real error; the node itself is polite to a fault.
• Silent clips. Some generations come back with no audio track; the node prints it detected no audio and hands you a placeholder. Not a bug, just how the model sometimes rolls.
• It's not free forever. The "free" framing you'll see in community posts about Agnes applies to a limited tier; heavy video use burns through credits fast, and a 321-frame render is the spendy end of the menu. Budget accordingly.

For the beginner who wants to see a video come out of their graph without first building a VRAM shrine to Wan, this is the friendliest on-ramp in the pack. The patience requirement (it is a queue) is the only real tax.