Tools & Resources Archive Details

LTX-2.3 Video Engine (Lightricks)

What it is

Lightricks’ LTX-2.3 is a diffusion-transformer video model that supports text-, image-, and audio-to-video, available as open-source weights for local runs or as an API billed per second of output.

Gabriel’s notes

LTX-2.3 is the latest release in the LTX-2 model family: a diffusion transformer (DiT) foundation model for generating high-fidelity video with synchronized audio, with support for text-to-video, image-to-video, and audio-to-video (including native portrait output up to 1080×1920). ([ltx.io](https://ltx.io/model/ltx-2-3))

Quick take: Video models have come a long way, and LTX-2.3 is a nice example of “serious tool” energy instead of “look what I can do, mom.” If you want a locally runnable video stack (or an API when you’re feeling lazy), this one’s worth a look. ([ltx.io](https://ltx.io/model/ltx-2-3))

I saved this under Video & audio because it’s one of the clearer “from local tinkering to production pipeline” bridges in the current video-gen zoo. ([ltx.io](https://ltx.io/model/ltx-2-3))

Good fit if you want to:

  • Generate video from text prompts and iterate quickly with “fast vs pro” style tradeoffs. ([ltx.io](https://ltx.io/model/ltx-2-3))
  • Do image-to-video that’s less “Ken Burns panic zoom” and more actual motion. ([ltx.io](https://ltx.io/model/ltx-2-3))
  • Drive motion/structure with audio-conditioned generation (voice/music/sfx as timing/pace). ([ltx.io](https://ltx.io/model/ltx-2-3))
  • Run locally with open weights (and keep your infrastructure boundaries intact). ([ltx.io](https://ltx.io/model/ltx-2-3))
  • Plug into ComfyUI workflows (official docs + community ecosystem). ([docs.comfy.org](https://docs.comfy.org/tutorials/video/ltx/ltx-2-3?utm_source=openai))

Pricing snapshot (auto-enriched):

The LTX API is billed per second of output video. As of the pricing doc, LTX-2.3 text/image-to-video starts at $0.06/sec (1080p, “fast”) and $0.08/sec (1080p, “pro”), scaling upward with resolution. ([docs.ltx.video](https://docs.ltx.video/pricing))

Work-use / compliance snapshot (auto-enriched):

Open-source use: Lightricks describes LTX as “open by default,” free to use (including commercial/production use) for companies under $10M in annual revenue, with a commercial license needed above that threshold. ([ltx.io](https://ltx.io/model/license))

API use: The public pricing page covers billing; deeper contractual terms, data retention, and privacy specifics are Unknown / not confirmed from the sources I reviewed here, so treat any production use like a normal vendor-risk exercise. ([docs.ltx.video](https://docs.ltx.video/pricing))

About that “22B + ComfyUI + GGUF + 12GB VRAM” stuff (cleaned-up from my raw note):

  • The official GitHub quick start references 22B checkpoints (e.g., ltx-2.3-22b-dev and ltx-2.3-22b-distilled-1.1), plus an upscaler and a distilled LoRA—so the “22B” and the existence of a “distilled 1.1” artifact are confirmed. ([github.com](https://github.com/Lightricks/LTX-2))
  • There are community Hugging Face repos packaging GGUF variants / workflows and referencing ComfyUI-GGUF tooling; this supports the general idea that GGUF/quantized workflows exist in the wild. ([huggingface.co](https://huggingface.co/Viral2AI/LTX-2.3-GGUF?utm_source=openai))
  • Claims like “runs on 12GB VRAM” show up in community discussions; I’d treat that as possible but highly configuration-dependent (resolution, batch, stages, upscalers, and how brave you are with quality). Unknown / not confirmed as an official requirement. ([reddit.com](https://www.reddit.com/r/comfyui/comments/1sd8j6r/im_looking_for_a_working_wf_for_ltx_23_that_runs/?utm_source=openai))
  • Specific workflow claims in my original note (e.g., “chunk feed forward,” “61KB all-in-one workflow download,” certain multi-image / IC-detailer integrations) are Unknown / not confirmed from the official sources I pulled in this pass.

Alternatives (auto-enriched):

  • OpenAI Sora: typically positioned as a proprietary alternative (terms/availability may vary). If you want open weights and local control, LTX-2.3 is the more “bring your own infra” vibe. Unknown / not confirmed beyond general positioning. ([en.wikipedia.org](https://en.wikipedia.org/wiki/LTX_%28AI_Model%29?utm_source=openai))
  • Google Veo (and other proprietary leaders like Kling): often referenced as competitors in benchmarks/discussions; if you need “easy mode” and don’t care about local runs, those ecosystems may fit better. Details are Unknown / not confirmed here. ([en.wikipedia.org](https://en.wikipedia.org/wiki/LTX_%28AI_Model%29?utm_source=openai))

Before you adopt it:

  • Decide up front: API convenience vs local control. You’ll optimize very different things. ([ltx.io](https://ltx.io/model/ltx-2-3))
  • Start with the official ComfyUI examples before downloading a 900-node community workflow you can’t debug. ([docs.comfy.org](https://docs.comfy.org/tutorials/video/ltx/ltx-2-3?utm_source=openai))
  • If you’re a company anywhere near the licensing threshold, read the license page and don’t “just assume” your use case is covered. ([ltx.io](https://ltx.io/model/license))

Sources

  • https://ltx.io/model/ltx-2-3
  • https://docs.ltx.video/pricing
  • https://ltx.io/model/license
  • https://github.com/Lightricks/LTX-2
  • https://docs.comfy.org/tutorials/video/ltx/ltx-2-3

Visit the resource