WaveSpeed LTX 2.3 Image-to-Video

WaveSpeed LTX 2.3 Image-to-Video is the slightly newer, slightly more controllable sibling of the pack's LTX 2 Pro node. Same idea - feed it an image URL and a prompt, get a video URL out - but LTX 2.3 stretches to 5–20 seconds, lets you pick the resolution, and, most importantly, treats its seed as a real generation seed that actually gets sent to the API. If you need reproducible clips, that's the difference that matters.

If you've read the LTX 2 Pro article, you already know the drill: WaveSpeed-hosted, billed per generation, output is a video_url STRING that plugs into the pack's Preview Anything node. The family context is the same too - Lightricks' speed-first video models, great for motion drafts and camera moves, a notch below the heavyweight models on faces and complex motion. The KB notes LTX-2.3 shipped in March 2026, so this is the current LTX release being served here.

The inputs

• image - source image URL to animate (a string)
• prompt - what the animation should do
• resolution - 480p / 720p / 1080p, default 720p; aspect ratio is inferred from the source image
• duration - 5 to 20 seconds, default 5
• seed - a genuine generation seed for reproducibility, -1 for random
• client - optional, falls through to ERPK Settings

The output is video_url (STRING). The aspect-ratio-from-source behavior is worth knowing: unlike the text-to-video nodes in this pack, there's no aspect ratio dropdown here - the model reads it off your image. That's usually the right call (you don't want the model guessing a frame shape that crops your subject), but it does mean your source image's shape is locked in, so pre-crop the image if you want a different frame.

LTX 2 Pro vs LTX 2.3

The pack deliberately ships these as separate nodes because their APIs differ. The practical differences: LTX 2 Pro offers 6/8/10 seconds with a generate_audio toggle, while LTX 2.3 offers 5–20 seconds, explicit resolution, and a real seed. The README spells out the seed difference explicitly: LTX 2 Pro's seed is cache control only (never sent), LTX 2.3's is a true generation seed. If reproducibility matters - you're iterating on a clip and want the same base motion between tweaks - 2.3 is the one. If you want audio, check whether 2.3's serving tier on WaveSpeed gives it to you; the node itself exposes no audio toggle, so that's the trade.

Install and setup

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Or install "ERPK Custom Nodes" via ComfyUI Manager, restart, and put a WaveSpeed API key in ERPK Settings (right-click canvas > ERPK Settings). Generation runs can take a couple of minutes; the pack polls every 10 seconds with a 15-minute timeout.

Troubleshooting

Same surface errors as the other WaveSpeed video nodes: "Image URL is required" and "Prompt is required" if you queue it empty. And remember the URL-string output - if you connect video_url to a plain Save Image node it'll do nothing useful; Preview Anything (or another URL-aware previewer) is the intended sink. One honest note: a 20-second clip is a long generation and a bigger bill - for iterating, start at 5 seconds and only stretch the duration when the motion direction is right.