Search "LTX 2.3 GitHub" and you get four different repositories, none of which is named LTX-2.3. The weights are not on GitHub at all. The ComfyUI integration lives somewhere else again. And the repo that looks most official — Lightricks/LTX-Video — is the previous generation.
This is a map. What each official repository contains, which one answers which question, and where the community forks fit. Compiled from the LTX-2 repository, LTX-Desktop, the Hugging Face model repository, and the official ComfyUI tutorial, checked July 2026.
The four official homes
| Repository | Type | What it is for |
|---|---|---|
Lightricks/LTX-2 | GitHub | Inference code, pipelines, LoRA trainer, architecture docs |
Lightricks/LTX-2.3 | Hugging Face | The actual weights — checkpoints, upscalers, distilled LoRAs |
Lightricks/LTX-Desktop | GitHub | Open-source desktop app for local generation |
Lightricks/ComfyUI-LTXVideo | GitHub | Lightricks' ComfyUI integration |
Rule of thumb: code on GitHub, weights on Hugging Face. Every "where do I download LTX 2.3" confusion comes from expecting weights in a git repo. They are 46.1 GB each — they were never going to be.
Inside the LTX-2 monorepo
Lightricks/LTX-2 is organized as a monorepo with three packages, and knowing the split saves a lot of aimless directory browsing:
packages/ltx-core — the foundational library: raw model definitions, component implementations, and loading logic used by the other two packages. It holds:
model/— PyTorch implementations of the transformer, Video VAE, Audio VAE, vocoder, and upscalertext_encoders/gemma— the Gemma text encoder with tokenizers and feature extractorsquantization/— the FP8 backends (FP8 scaled MM, FP8 cast)components/— schedulers, guiders, noisers, patchifiersconditioning/— image, video, and keyframe conditioning tools
Its README is also the best architecture document Lightricks publishes: the asymmetric dual-stream design, the 14B video / 5B audio split across 48 blocks, and the exact VAE tensor shapes.
packages/ltx-pipelines — the high-level generation interfaces. This is where the eleven runnable pipelines live (ti2vid_two_stages, distilled, ic_lora, keyframe_interpolation, a2vid_two_stage, retake, hdr_ic_lora, lipdub, t2a_one_stage, and the one-stage/HQ variants). Its docs/ folder holds the practical references most people actually need:
installation.md— install steps and the shared CLI flagspipelines.md— full reference for every pipeline and its constraintspipeline-selection.md— which one to useoptimization.md— FP8,torch.compile, gradient estimationmultimodal-guidance.md— CFG, STG, and modality guidance parametersconditioning.md— replacing vs. guiding latentsmultigpu/— sequence parallel, distributed decoder, tiled data parallel
packages/ltx-trainer — LoRA training, full fine-tuning, and IC-LoRA, with ready-made configs (t2v_lora.yaml, i2v_lora.yaml, t2v_lora_low_vram.yaml, audio and IC-LoRA variants) plus Accelerate configs for DDP and FSDP.
Only want to generate, not read source? Cloning a monorepo to make one video is a lot of ceremony. ltx23.app runs LTX 2.3 in the browser — worth trying first, and the repos will still be there when you need real control.
What is on Hugging Face
The Lightricks/LTX-2.3 repository holds the weights — three 46.1 GB checkpoints, distilled LoRAs at 7.61 GB each, spatial upscalers around 1 GB, and a 262 MB temporal upscaler. About 157 GB in total, which is why targeted hf download calls beat cloning.
The task-specific LoRAs live in separate repositories, not in the main one:
Lightricks/LTX-2.3-22b-IC-LoRA-Union-ControlLightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-ControlLightricks/LTX-2.3-22b-IC-LoRA-HDRLightricks/LTX-2.3-22b-IC-LoRA-LipDubLightricks/LTX-2-19b-IC-LoRA-Detailer,-Pose-ControlLightricks/LTX-2-19b-LoRA-Camera-Control-*(one per camera move)
Access is gated: accept the terms and use a Read token, with the "read gated repos" scope enabled for fine-grained tokens.
LTX-Video vs LTX-2: the repo people open by mistake
Lightricks/LTX-Video is the original LTX Video repository — the 2B and 13B generation. It is still there, still has stars, and still ranks well in search.
LTX 2.3 does not live there. If you are following instructions that reference ltx-video-2b-v0.9.5.safetensors or a T5 text encoder, you are reading documentation for the older family. LTX 2.3 uses Gemma 3, not T5, and its checkpoints are named ltx-2.3-22b-*.
That mismatch is the single most common source of broken LTX 2.3 setups, and it is entirely a search-results artifact.
Community repositories, and how to treat them
Two community hubs come up constantly and are genuinely useful — with a caveat.
Quantized builds. Community repositories publish GGUF and INT8 conversions the official repo does not ship. For example, Kijai/LTX2.3_comfy hosts ComfyUI-oriented files including an INT8 transformer build, a text projection file, and a VAE file. These explain search terms like ltx 2.3 text projection and ltx 2.3 vae that have no match in the official repo.
Workflow and LoRA sharing. Civitai and similar sites host community LoRAs and workflow JSON.
The caveat, stated plainly: these are not signed by Lightricks. For base weights, prefer the official repository. For quantized experiments, community builds are the only game in town — just know which is which, and expect community files to lag official point releases.
Which repo answers which question
| Your question | Where to look |
|---|---|
| "Which file do I download?" | Hugging Face Lightricks/LTX-2.3 |
| "How do I run it from Python?" | LTX-2 → packages/ltx-pipelines/docs/installation.md |
| "Which pipeline do I use?" | pipeline-selection.md, then pipelines.md |
| "How do I fit it in my VRAM?" | optimization.md |
| "What do CFG/STG do here?" | multimodal-guidance.md |
| "How is the model actually built?" | packages/ltx-core/README.md |
| "How do I train a LoRA?" | packages/ltx-trainer |
| "How do I use it in ComfyUI?" | ComfyUI's own LTX-2.3 tutorial + ComfyUI-LTXVideo |
| "Can I just use an app?" | LTX-Desktop |
FAQ
What is the official LTX 2.3 GitHub repo?
github.com/Lightricks/LTX-2 for code. Weights are on Hugging Face at Lightricks/LTX-2.3.
Is LTX-Video the same project? Same family, earlier generation. LTX 2.3 is in the LTX-2 repo and uses different checkpoints and a different text encoder.
Where are the ComfyUI nodes?
ComfyUI has native LTX-2.3 support with built-in workflows; Lightricks also maintains ComfyUI-LTXVideo.
Where do I report a bug?
The repository matching the layer you are using — pipelines/trainer issues in LTX-2, app issues in LTX-Desktop.
Is there an official Colab? The documented paths are the Python package, ComfyUI, and LTX Desktop. Colab notebooks you find are community work.
Bottom line
Code on GitHub, weights on Hugging Face, app in its own repo, LoRAs in one repo each. Once that layout is clear, the docs are genuinely good — ltx-core for architecture, ltx-pipelines/docs for everything operational, ltx-trainer for training.
And if you would rather generate than clone, ltx23.app is the zero-repo path.
Sources
Verified July 2026 against primary documentation:
- Lightricks/LTX-2 — official GitHub repository — monorepo structure, packages, model list
- ltx-core package documentation — module layout and architecture reference
- ltx-pipelines documentation index — pipeline reference
- LTX-2 installation & CLI flags — runnable modules and flags
- ltx-trainer package — training configs and requirements
- Lightricks/LTX-2.3 — Hugging Face repository — weights, sizes, gated access
- Lightricks/LTX-Desktop — desktop application repository
- Lightricks/ComfyUI-LTXVideo — official ComfyUI integration
- Lightricks/LTX-Video — previous-generation repository, for disambiguation
- ComfyUI official LTX-2.3 tutorial — native workflow documentation
- LTX-2 technical paper (arXiv 2601.03233)
Community repositories are named for orientation only; they are not maintained or signed by Lightricks.
