LTX-2 Distilled (Replicate)

LTX is Lightricks' open video family, and its identity is speed. Where Wan takes minutes per clip, LTX takes seconds - the community built its whole workflow around "draft in LTX, refine in Wan." The distilled version takes that further: it's the few-step distillation, the one that runs on surprisingly small cards, and the one you'd reach for when iteration speed beats everything else.

So like Z-Image, the honest first question is why rent it? Because it's open, and the distilled tier runs locally on 4–6GB. If you have that GPU, you can run this yourself for free. If you don't - no local GPU, or your ComfyUI lives in the cloud - this node gives you the same model per call, no VRAM, no weights download. That's the legitimate use case, and it's a good one: LTX's whole value is drafts, and drafts are exactly what you don't want to meter forever. Don't iterate too long before noticing the bill.

How it works

Standard engine around lightricks/ltx-2-distilled. Prompt and optional start image go up, the MP4 comes back saved to output/video/, and the node returns video_path as a STRING (it's an output node, so the clip appears in UI history too). There's no audio handling in this schema - despite the "synchronized audio" marketing you may have seen on the wider LTX-2 family - so treat this as a silent-video producer.

The inputs that matter

• prompt (required) - the text prompt. One LTX quirk that carries over: it punishes short prompts. Give it a proper scene description or you'll get mushy results.
• image - optional image-to-video first frame. When provided, aspect_ratio is ignored.
• aspect_ratio - 16:9, 9:16, 4:3, 3:4, 1:1, 21:9; default 16:9. Only applies to text-to-video.
• num_frames - default 121, range 25–241, and the tooltip is explicit about the rule: it must follow 8*k + 1 (e.g. 81, 97, 113, 121). Set a number that breaks the formula and the model will complain.
• enhance_prompt - off by default. LTX's built-in prompt enhancement; worth a try if you're fighting the short-prompt weakness.
• image_strength - 0–1, default 1.0. For i2v: higher follows the image more closely, lower gives the model room to move. 1.0 is "stay faithful to this frame."
• seed - set for reproducibility.

Output: video_path, a STRING.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Dependency: replicate>=1.0.7. Token via environment variable:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• The frame formula. num_frames isn't free-form. 121 is the sweet spot; anything off the 8k+1 grid is a failed call or a silent fallback.
• Short prompts. LTX is the model where "a cat" genuinely underdelivers. Flesh out the scene.
• It's still a draft-tier model. LTX loses to Wan on faces, bodies, and complex motion. For final-quality people shots this hosted node is the wrong tool even before you factor the meter.
• Path, not tensor. video_path is a STRING; the MP4 is a file. Load it back into the graph with a video node if you need frames downstream.