TensorScale LTX-2.5 Fast Video (Soze)

Lightricks' LTX line has always been about speed - the local models were the "generate faster than you can watch it" crowd. This node is the hosted take on that philosophy: TensorScale's LTX-2.5 Fast, billed at a flat $0.04 per generated video second, with audio generated and synced in the same pass. No local weights to download, no VRAM to budget, just a prompt and a key. If you've ever run LTX locally and wished it came with a soundtrack, this is that wish fulfilled.

The inputs

The prompt is the single most important input, and the tooltip makes the reason clear: it describes both the video and the audio. "A lighthouse beam sweeps the fog; waves crash against the rocks below" is an LTX-2.5 prompt - the sound is generated from the same description, so write for the ear as well as the eye.

• resolution - five options from 576x1024 up to 1920x1088 (native 1080p, more or less).
• num_frames - 121 (5 seconds) or 241 (10 seconds). Frame rate is fixed at 24 fps, so the frame count is the duration; there's no other length.
• seed - defaults to 10; set it for reproducible takes.

Then the image-to-video path: connect an image and it conditions frame 0. image_strength (0–1, default 0 = let the service decide) sets how strongly that frame anchors the clip, and crf (0–51, default 0 = omit) overrides the output compression. One important quirk: this endpoint only accepts the image inline - it rejects URLs. There's no use_fal_upload option here, so the 10 MiB body cap applies to your conditioning frame; a normal PNG is fine, an enormous one isn't.

How it works

TensorScale's /v2/ltx2-5/fast runs synchronously with a model-scoped key (TENSORSCALE_API_KEY_LTX2_5_FAST, falling back to TENSORSCALE_API_KEY). The default timeout is 7200s - the maximum the pack allows - because a 10-second clip with audio is a genuinely long single request. The node logs the estimated cost, so you can see the per-run bill in the status output before it even lands.

Outputs: video (the VIDEO object), video_path, request_id, config (a run summary including the cost estimate), and status.

Install and key

From the Soze pack ("Quality of Life Nodes for ComfyUI"), via ComfyUI Manager or:

cd ComfyUI/custom_nodes
git clone https://github.com/SozeInc/ComfyUI_Soze.git
pip install -r ComfyUI_Soze/requirements.txt

Restart, then set TENSORSCALE_API_KEY or the model-scoped TENSORSCALE_API_KEY_LTX2_5_FAST (environment variable or .env in the repo root). Keep it out of git.

Common issues

• Cost surprise. At $0.04/second, a 10-second clip is $0.40 - cheap per render, not free when batched. The status output shows the estimate, so check it before you run a queue of fifty.
• Image conditioning rejected. This endpoint refuses URLs, so only the inline socket works - and it must stay under the 10 MiB cap.
• Length is locked to 5 or 10 seconds. If you need 7, this node isn't the tool; the frame-count dropdown is fixed.