LTX 2.3 LoRA: IC-LoRA, ID-LoRA, and Training Your Own

"LoRA" means something broader in LTX 2.3 than it does in image generation. Alongside the familiar style-and-likeness adapters, Lightricks ships LoRAs that act as control mechanisms — camera moves, pose transfer, HDR output, lip dubbing — and one that is a mandatory pipeline component rather than an aesthetic choice.

That last one catches people out constantly: ltx-2.3-22b-distilled-lora-384-1.1.safetensors is required by most two-stage pipelines for the stage-2 refinement pass. It is not optional flavor. If your two-stage render errors out, this file is the first thing to check.

This guide covers the full official LoRA catalog, what each category does, how to apply them, and what it really takes to train your own. Sourced from the LTX-2 repository, the ltx-trainer package, the LTX-2.3 model card, and the ComfyUI tutorial, checked July 2026.

The three kinds of LTX 2.3 LoRA

Most guides list LoRA files. The useful move is separating them by what they change, because that determines whether you can stack them.

1. Pipeline LoRAs — infrastructure. The distilled LoRA (7.61 GB) that two-stage pipelines require. Documented exceptions where it is not needed: DistilledPipeline, ICLoraPipeline, and LipDubPipeline.

2. IC-LoRA — in-context control. These condition generation on a reference input — a clip, a pose sequence, a depth map. They change what the model is looking at, not what it looks like. Critically, ICLoraPipeline can only be used with a distilled model checkpoint.

3. Subject and style LoRAs — likeness. ID-LoRA for a consistent person (with voice cloning), plus anything you train yourself for a character, product, or visual style.

Category matters because IC-LoRAs come with pipeline constraints that identity LoRAs do not. Loading an IC-LoRA in a workflow that has no reference input does nothing useful.

The official LoRA catalog

Published by Lightricks on Hugging Face:

Read the version prefix. Union Control, Motion Track, HDR, and LipDub are 22b LTX-2.3 era. Detailer, Pose Control, and every Camera Control adapter are 19b LTX-2 era. Mixing generations is a well-earned source of confusing output.

In ComfyUI the naming differs again — the ID-LoRA template uses ltx-2.3-id-lora-talkvid-3k.safetensors, and the IC-LoRA template pairs ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors with a geometry estimator, moge_2_vitl_normal_fp16.safetensors, in models/geometry_estimation/. That estimator is what converts your reference video into the depth and normal signals the LoRA consumes; without it the control workflow has nothing to condition on.

Applying LoRAs

On the CLI, LoRAs load through a repeatable flag with an optional strength:

--lora path/to/lora.safetensors 0.8

Default strength is 1.0, and the flag can be repeated to stack adapters. The distilled LoRA has its own dedicated argument in the two-stage pipelines:

python -m ltx_pipelines.ti2vid_two_stages \
    --checkpoint-path path/to/checkpoint.safetensors \
    --distilled-lora path/to/distilled_lora.safetensors 0.8 \
    --spatial-upsampler-path path/to/upsampler.safetensors \
    --gemma-root path/to/gemma \
    --prompt "..." \
    --output-path output.mp4

Note the 0.8 on the distilled LoRA in Lightricks' own example — even the official refinement adapter is not applied at full strength by default in their documented command. That is a useful calibration for your own stacking: start below 1.0 and raise it.

LipDubPipeline is stricter than the general case — it requires the --lora flag exactly once, with a lip-dub IC-LoRA, on a distilled checkpoint.

Trying LoRA-style results without the setup? If what you actually want is consistent characters or controlled motion rather than a local training rig, ltx23.app runs LTX 2.3 in the browser — a faster way to find out whether the base model already does what you need before you build a dataset.

Training your own: the honest hardware bill

This is where expectations need adjusting. The ltx-trainer documentation states that training requires an NVIDIA GPU with 80 GB+ VRAM for standard configurations, with a low-VRAM config for 32 GB GPUs using INT8 quantization.

That is a very different world from training an image LoRA on a gaming card. Realistically:

• 80 GB class (A100/H100) — the documented standard path
• 32 GB — possible via the INT8 low-VRAM config; there is a dedicated t2v_lora_low_vram.yaml config in the repo for exactly this
• Below 32 GB — rent, don't fight it

The counterweight to that cost is speed. The LTX-2.3 model card states that training for motion, style, or likeness — sound and appearance together — can take less than an hour in many settings. A short rental on a big card is a genuinely reasonable plan.

Train against the dev checkpoint: it is the variant the model card describes as trainable in bf16.

What the trainer can actually train

The repository ships ready-made configs, and the list tells you the real scope of what LoRA means here:

Read that list as a capability map: you can train an adapter for audio behavior alone — an accent, a voice style — and validate it with the audio-only T2AOneStagePipeline before spending a single frame of video compute. That workflow is barely discussed anywhere and it is the cheapest way to iterate on a voice.

Multi-GPU is supported through Accelerate configs for DDP and FSDP (with compiled variants of each), so scaling out is a config change rather than a rewrite. The trainer also supports full fine-tuning, not just LoRA, if you have the hardware for it.

Practical stacking rules

• One identity LoRA at a time. Two likeness adapters fight; the result is neither person.
• Identity + control usually composes. An ID-LoRA plus a camera-control LoRA is asking for two different things, which is exactly when stacking works.
• Lower strength first. Start around 0.7–0.8 and raise until the effect appears without the output going plastic. Lightricks' own example uses 0.8.
• Match the era. 22b LoRAs with the 22b checkpoint; 19b adapters were trained against a different base.
• Check the pipeline constraint before blaming the LoRA. IC-LoRA needs a distilled checkpoint. Wrong base, wrong result — the adapter is fine.

FAQ

Where do I download LTX 2.3 LoRAs? The official ones live in Lightricks' Hugging Face repositories, linked from the LTX-2 README. Community LoRAs appear on Civitai — useful, but unverified.

What is IC-LoRA? In-context LoRA: it conditions generation on a reference input (video, pose, depth, audio) rather than restyling output. It requires a distilled checkpoint in the official pipeline.

What is ID-LoRA? Identity personalization — a consistent person across generations, with voice cloning, exposed as a native ComfyUI template.

Can I train an LTX 2.3 LoRA on a consumer GPU? The documented floor is a 32 GB GPU with the INT8 low-VRAM config. Standard configs want 80 GB+.

How long does training take? The model card says motion, style, or likeness training can take under an hour in many settings — but that presumes appropriate hardware.

Do I need the distilled LoRA? For most two-stage pipelines, yes. Not for DistilledPipeline, ICLoraPipeline, or LipDubPipeline.

Bottom line

LTX 2.3's LoRA ecosystem is really three ecosystems: one required pipeline component, a set of in-context control adapters, and identity/style adapters you can train yourself — with a documented floor of 32 GB VRAM and a documented ceiling of under an hour per run.

Before you build a dataset, check whether the base model plus a good prompt already gets you there. ltx23.app is the quickest way to answer that question.

Sources

Verified July 2026 against primary documentation:

1. Lightricks/LTX-2 — official GitHub repository — full LoRA catalog, --lora usage, distilled LoRA requirement
2. ltx-trainer package documentation — 80 GB+ requirement, 32 GB INT8 low-VRAM config, LoRA/full-finetune/IC-LoRA support
3. ltx-trainer configs directory — the shipped training configs and Accelerate DDP/FSDP setups
4. Lightricks/LTX-2.3 — Hugging Face model card — distilled LoRA files and sizes, training time, trainable variant
5. LTX-2 available pipelines documentation — IC-LoRA and LipDub pipeline constraints
6. LTX-2 installation & CLI flags — --lora <path> [strength], default 1.0
7. ComfyUI official LTX-2.3 tutorial — ID-LoRA and IC-LoRA templates, geometry estimator
8. LTX-2.3-22b-IC-LoRA-Union-Control
9. LTX-2.3-22b-IC-LoRA-LipDub
10. LTX-2.3-22b-IC-LoRA-HDR
11. LTX-2-19b-LoRA-Camera-Control-Dolly-In — camera-move adapter family
12. LTX-2 technical paper (arXiv 2601.03233)

Training hardware figures are Lightricks' published requirements for their own trainer; third-party training scripts may differ.