LTX-2.3 LoRA · Image to Video

Why 2.3 instead of 2

LTX-2.3 is the March 2026 release that, by mid-2026, became the local community's default video generator - the year's highest-scoring LTX threads are adoption posts, not announcement posts. It's 22B and it fixes the two things people complained about in LTX-2: sound is stereo rather than mono, and it drifts less from an opening frame. If you're animating a still and the still has to survive, this node is the one you want.

It's also the node where the LoRA stack is most likely to have something to load, because 2.3's adapter library filled in while 2.2-era and LTX-2-era files became yesterday's news.

Costs more than plain LTX-2: about $0.024/s at 480p, $0.036 at 720p, $0.048 at 1080p, and there's a flat per-request component in the rate table too (60 credits), so very short clips don't scale down as gracefully as the per-second number suggests. The quote before each run is the truth here, not my arithmetic.

How the node works

You connect an IMAGE, describe the motion, and the node uploads the frame (PNG, or JPEG above the 10 MiB upload ceiling), resolves any LoRA references, quotes the job, submits it, and polls. The clip inherits the source image's aspect ratio - so, unlike the text-to-video version, there's no aspect_ratio widget to set.

Resolution options are 480p, 720p and 1080p - note there's no 2K tier here, unlike the 2.5 family. Duration is 5 to 20 seconds in whole seconds.

There's a real prompt hint in the tooltip: put LoRA trigger words first, describe what happens after the opening frame (action, camera, sound), and keep it in prose. LTX rewards specificity and punishes one-liners with either a static shot or a mess.

Inputs

• image_url - the opening frame, and the aspect-ratio authority for the whole clip.
• prompt - trigger words, then action, camera, sound.
• loras - up to three LTX-2.3 adapters with individual scales. Over 2 GiB, rejected.
• resolution - 480p / 720p / 1080p.
• duration_seconds - 5–20.
• seed - fixed by default; queue twice, get the cached first result rather than a second bill.

Outputs are video and task id.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

That's the whole install. No pip step, no model download, no ffmpeg fuss - the pack declares an empty dependency list and leans on what ComfyUI already ships. Manager users: search PopNow, Install, restart. CLI: comfy node install popnow-comfyui.

Then a key from popnow.ai/api/console/keys into Settings > PopNow, or POPNOW_API_KEY exported before launch. Restart once after the key so the catalog fetch populates your node list; the node lives under PopNow > Video > Lightricks, next to the LTX-2 19B and LTX 2.5 families.

Where people get burned

Wrong LoRA family. 2.3 wants 2.3 adapters. An LTX-2 19B LoRA will upload happily and then underperform or get rejected; the upload step doesn't check architecture for you. The tooltip on loras names the family, which is more than most nodes do.

Expecting 4K. Not here. The 2.5 node has the 2k and 4k tiers; 2.3 tops out at 1080p.

Long clips and upload size. The source IMAGE is capped at 10 MiB as PNG and gets re-encoded to JPEG to fit. A heavy 8K source will lose detail before the model ever sees it.

Free-looking retries that aren't. Changing the seed and re-queueing is a new paid render. That's fine - it's how you get variations - but say it out loud before running a batch overnight.

Errors follow the pack conventions: 40201 is an empty wallet, 40004 is an unservable combination and names the field, content_rejected means the model declined and you were refunded. Set Max credits per run in Settings > PopNow if you'd rather a mistake cost nothing at all.