LTX 2.5 · Image to Video

Two frames instead of one

Most image-to-video nodes take a starting picture and hope. This one takes a starting picture and, optionally, an ending one: connect last_image_url and the clip has to resolve on that frame. Bookending a shot is one of those small controls that changes what you can actually plan - you can now say "start here, end there, and figure out the middle" instead of "keep going and I'll pick the take that didn't wander."

And it's 4K-capable, which almost nothing else in this pack is. Resolution goes 720p, 1080p, 2k, 4k, and duration runs 5 to 20 seconds. That combination is the pitch: a long take, at delivery resolution, with control over both ends.

It's also the most expensive thing here, and not by a little. The per-resolution rates are roughly $0.09/s at 720p, $0.13 at 1080p, $0.19 at 2k and $0.30 at 4k - so twenty seconds at 4K lands near six dollars. That's still far below frontier cloud video, but it's an order of magnitude above the LTX-2 19B node, and you should know that before you queue a batch to "see how it looks." 2.5 also has a much smaller community footprint than 2.3 did - fewer field reports, fewer prompt recipes. You're early.

How it works

Your opening frame is encoded and uploaded (PNG, JPEG if it would exceed 10 MiB), the closing frame likewise if you connected one, and the job goes out with a quote taken first. The clip comes back as a VIDEO value. Because the output is usually too big to eyeball frame by frame in ComfyUI's preview, plan on Save Video plus a look in a real player before you decide anything.

The tooltip for the frame input says "Authorised opening frame", which is the API's way of reminding you that you're uploading someone's picture to a paid service - use what you have rights to. Same for the closing frame.

Inputs

• image_url - the opening frame.
• last_image_url (optional) - the closing frame. This is the reason to pick 2.5 over the 19B node if composition matters more than price.
• prompt - the action and camera between the two frames, in prose. The tooltip here is terser than its siblings' and doesn't mention audio, but the model page advertises a soundtrack - describe what you hear anyway.
• resolution - four tiers, and the big cost lever.
• duration_seconds - 5 to 20.
• seed - fixed out of the box, so an unchanged re-queue reuses the last result rather than billing again.

Outputs: video and task id.

Install

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

Manager > search PopNow > Install works too, as does comfy node install popnow-comfyui. There's nothing to download and no dependency step - no torch version roulette, no CUDA, nothing to break. Key from popnow.ai/api/console/keys into Settings > PopNow (or POPNOW_API_KEY in the environment), restart, and look under PopNow > Video > Lightricks.

Cost control, seriously

This is the node where I'd actually bother with guardrails. Settings > PopNow > Max credits per run sets a per-run ceiling: any job quoted above it stops before credits are held. Set it to something like your 1080p-10-second comfort number and let 4K be a deliberate decision rather than a mis-click.

Draft at 720p. The prompt, the timing and the camera move are all settled at 720p; resolution changes almost nothing about the composition and multiplies the bill by three. Only re-run the winner at 2k or 4k.

Also remember that seeds start fixed and cancelling doesn't refund: stopping your wait leaves the render running on PopNow, and it's charged if it succeeds. Failed or refused runs (generation_failed, content_rejected) are refunded, so a model-side failure costs you time, not credits.