LTX-2 19B LoRA · Image to Video

The same LTX-2 shot, wearing a LoRA

Functionally this is the LTX-2 19B image-to-video node with one extra socket. Practically, that socket is the difference between "a generated clip" and "a generated clip in the look you spent a weekend training." You connect a PopNow LoRA node to loras, point it at an LTX-2 19B adapter, and the animation picks up that style - a camera/lens look, a grain, a character, whatever the file was trained for.

The cost of the privilege is modest: about $0.018/s at 480p, $0.024 at 720p, $0.036 at 1080p, versus $0.014/$0.019/$0.029 for the plain model. Roughly a quarter more for a look you can't get otherwise. If you're making a series where consistency matters, that's the cheapest consistency you'll find in this pack.

One thing to be clear about: LTX-2 video LoRAs are their own small ecosystem, and they're not interchangeable with image LoRAs or with the LTX-2.3 files. Family match or nothing.

What it does under the hood

Your opening IMAGE is encoded and uploaded (PNG, falling back to JPEG above 10 MiB), the LoRA reference is resolved by the upstream LoRA node - uploaded once from ComfyUI/models/loras or imported by PopNow from a Hugging Face or Civitai page - and the whole thing goes out as a single job. The returned clip keeps the aspect ratio of the source image, which is why this node has no aspect_ratio field while its text-to-video sibling does.

And the usual LTX caveat still applies: it's a speed-tier family. Scenery, architecture, camera moves and mood are its strengths; a face turning convincingly mid-shot is where it embarrasses itself. Audio, by contrast, is genuinely its own - LTX-2 puts sound in the same network as the picture.

Inputs you'll actually touch

• image_url - the opening frame, and the thing the clip drifts from over time.
• prompt - action, camera and sound, in prose. Put the LoRA trigger words first; the tooltip says so, and it's the difference between the adapter landing and it being decorative.
• loras - up to three stacked, each with its own scale. Over 2 GiB and the file is rejected.
• resolution - 480p / 720p / 1080p, price lever included.
• duration_seconds - 5 to 20.
• seed - fixed by default, so a re-queue reuses the previous result rather than charging for a second render.

Outputs: video and task id.

Install

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

Manager users can just search PopNow and Install; the CLI has comfy node install popnow-comfyui. No dependencies are declared and nothing is downloaded beyond the pack itself. Add a key from popnow.ai/api/console/keys in Settings > PopNow, or export POPNOW_API_KEY, then restart once so the catalog refreshes and the node appears under PopNow > Video > Lightricks.

Honest troubleshooting

Nothing changed. Nine times out of ten the LoRA family is wrong - an SDXL or image LoRA in a video node - or the trigger words aren't in the prompt. Check both before you blame the scale. A scale of 0.3 on a two-word prompt is a rounding error.

The clip doesn't match my still. LTX-2 image-to-video drifts from the source frame; it's a known characteristic of the family at launch, not a setting you got wrong. Shorter clips, with the last frame of each feeding the next, is the workaround people actually use.

Gated LoRA won't import. Civitai downloads need CIVITAI_API_KEY in ComfyUI's environment; Hugging Face gated repos need HF_TOKEN. Tokens are read from the environment only - never type them into a node, since widget values travel with every shared workflow.

Wallet errors. PopNow 40201 is "not enough credits" and 40004 is "this combination of settings isn't servable" - the message names the offending field. Both are cheap to diagnose if you read the text instead of re-queueing.