LTX-2.3 Prompt Generator

Verdict

Feature snapshot

What's new in LTX-2.3

Lightricks frames 2.3 as an engine upgrade, not a cosmetic one. CEO Zeev Farbman calls out three changes that came straight from community feedback: a new latent space, a bigger connector, and a better treatment of audio. The latent space matters most for the eye, a very compressive latent is efficient but loses detail, so the rebuilt one means that at the same resolution you get more detail and less smear, with stronger motion at the same speed even on consumer hardware.

The connector, the piece that links the language model to the diffusion model, grew to around 20 billion parameters, which translates into closer prompt adherence: the video tracks nearer to what you actually wrote. On top of that, 2.3 makes portrait video native rather than a crop-and-pray workaround, and supports native 4K at industry-standard frame rates. The practical upshot for prompting: detail pays off. A precise, structured brief gets rendered more faithfully than it would have in earlier versions.

How LTX compares to other AI video models

Where LTX-2.3 sits against the rest of the field on value and output quality, and how it scores capability by capability. Hover or tap any model for the detail.

The open-source advantage

LTX-2.3 is the only model in this lineup you can actually download and own. Lightricks ships open weights on Hugging Face (millions of downloads and counting) under an open license, and the whole pitch is that 'closed APIs are not enough', they want people building on top of an open multimodal engine. That's not just ideology; it's a different workflow. You can run it locally with no per-clip API bill, fine-tune it on your own footage, and tap a community LoRA ecosystem that adds capabilities the base model doesn't ship with.

If you don't want to touch ComfyUI, you don't have to. LTX Studio is Lightricks' own web app running 2.3, and major AI video platforms bundle LTX 2.3 alongside other models, so the same prompt you generate here works whether you're a tinkerer running open weights or a creator who just wants to paste and go. The deeper control (dialogue, camera transfer) rewards the ComfyUI route; the hosted surfaces are the fast, no-setup path.

Dual-character dialogue & custom audio

One of the most useful community workflows is the IC-LoRA Dual Character V2 setup in ComfyUI, built for the 'two people talking, podcast-style' scene that most models fumble. It runs on the distilled 2.3 checkpoint with the dual-character IC-LoRA doing the heavy lifting, and you can stack an enhanced-motion LoRA and a detailer LoRA to lift overall quality. The workflow flips between text-to-video and image-to-video, and lets you set duration directly in seconds.

The key prompting decision is audio. You can let LTX generate dialogue from your prompt, or feed your own custom track, for example lines you made in ElevenLabs or another TTS model. When you supply custom audio, the advice from the workflow author is to stop writing dialogue in the prompt and instead describe motion and shot type per beat (shot 1, shot 2, shot 3) and, crucially, who speaks first and who answers. When you don't supply audio, you write the dialogue lines in the prompt and LTX voices them. Either way, the prompt this tool produces is laid out as a clear shot list so it drops cleanly into that workflow.

Camera control with IC-LoRA

The other standout open-source trick is camera-motion transfer via the IC-LoRA Camera Man V2 (also from Hugging Face). The idea: build your scene as a keyframe image first, then hand LTX a reference video purely for its camera movement, a pan, a tracking shot, a rise-and-fall, and it transfers that exact camera path onto your new footage. It's not really image-to-video; it's image-plus-borrowed-camera-move. In the demos a dancing clip's tracking shot got mapped onto a motorbike on a stormy highway, and a sports-car clip's left-right cuts got mapped onto a spaceship over a city.

Two settings make or break it, both from the workflow author: keep resolution at 960×512 or above for clean output, and set the conditioning image strength between 0.5 and 0.7 to balance keyframe consistency against motion. The big win is that you stop writing paragraphs trying to describe a camera move in words, you just show LTX the move you want. You can even record your own phone footage as the camera reference.

LTX vs the closed models

In the 2026 4K head-to-head, LTX-2.3 was tested side by side with Veo 3.1 and Kling 3.0 on the same starting frame and prompt, at 4K quality. It impressed on detail, the reviewer singled out how defined the boxer's muscles looked and how solidly the hits landed, and it could open at high intensity before dropping into slow motion. The honest caveat: on that action shot it delivered one punch plus a long slow-motion beat, which felt 'a bit less dynamic' than the alternatives, and like Kling it lacked Veo's multi-image character-reference feature.

So the trade is clear. The closed models can edge it on raw dynamism and a few convenience features. LTX-2.3 gives you comparable 4K quality plus the things they can't: open weights, local execution with no metered API, fine-tuning, and an open LoRA ecosystem for dialogue and camera control. If you value ownership and control, write your prompt for LTX; if you only care about the single most dynamic action take, the gap closes by chaining shots and using first/last-frame guidance.

Other model prompt generators