{"id":5897,"date":"2026-03-26T17:36:05","date_gmt":"2026-03-26T09:36:05","guid":{"rendered":"https:\/\/crepal.ai\/blog\/?p=5897"},"modified":"2026-03-26T17:36:07","modified_gmt":"2026-03-26T09:36:07","slug":"ltx-2-3-ic-lora-guide","status":"publish","type":"post","link":"https:\/\/crepal.ai\/blog\/aivideo\/ltx-2-3-ic-lora-guide\/","title":{"rendered":"How to Use IC-LoRA in LTX 2.3: Identity-Consistent Video"},"content":{"rendered":"\n

HEY fellows! I`m Dora, as usual. Lately, one problem that sent me down this rabbit hole was embarrassingly simple: I was making an episodic short-form series for a client, and every time I generated a new clip, the main character looked like a different person. Same prompt, same seed range, same style LoRA \u2014 different cheekbones, different eyes, different voice tone. Consistent identity across video clips has been the hardest unsolved problem in AI video generation for a while.<\/p>\n\n\n\n

IC-LoRA in LTX 2.3<\/a> is the most practical answer I’ve found so far. It’s not magic, but it’s real, it just got proper native ComfyUI support. This guide covers what it actually does, how to train one, and what you should realistically expect from your output.<\/p>\n\n\n\n

What IC-LoRA Is and Why It Matters for Video<\/h2>\n\n\n\n

IC-LoRA stands for In-Context <\/strong>LoRA<\/strong> \u2014 a training approach where the LoRA learns to condition generation on reference inputs rather than just text. Standard LoRAs teach the model “what this concept looks like” through captions and training clips. IC-LoRA teaches it something more structured: “given this reference image and this reference audio, produce a video where the subject looks and sounds like the reference.”<\/p>\n\n\n\n

The practical payoff for creators: you provide a single reference image (a face, a character, a product) and a short audio clip (a voice sample), and LTX 2.3 generates a video where that person is speaking, moving, and looking consistent with the reference \u2014 synchronized audio included. That’s genuinely different from text-prompt consistency tricks, which always drift across generations.<\/p>\n\n\n\n

\n
\"\"<\/figure>\n<\/figure>\n\n\n\n

This matters most for anyone building recurring characters: virtual hosts, brand mascots, indie animation, AI-assisted narration series, or product spokespersons. The goal isn’t photorealistic deepfakes \u2014 it’s controllable, repeatable identity that holds up across a content pipeline.<\/p>\n\n\n\n

IC-LoRA vs Standard LoRA in LTX 2.3<\/h2>\n\n\n\n

Understanding the difference prevents a lot of wasted training time:<\/p>\n\n\n\n

<\/td>Standard LoRA<\/td>IC-LoRA<\/td><\/tr>
Activation<\/td>Text trigger token<\/td>Reference image + audio at inference<\/td><\/tr>
What it learns<\/td>Style, concept, motion pattern<\/td>Identity mapping from reference to output<\/td><\/tr>
Dataset type<\/td>Video clips + text captions<\/td>Video clips + paired reference frames\/audio<\/td><\/tr>
Training complexity<\/td>Lower<\/td>Higher \u2014 paired data required<\/td><\/tr>
Use case<\/td>Style transfer, character aesthetics<\/td>Face\/voice identity preservation<\/td><\/tr>
Inference input<\/td>Text prompt only<\/td>Text + reference image + (optionally) audio<\/td><\/tr>
Training steps<\/td>1000\u20132000 typical<\/td>4000\u20136000 for identity (ID-LoRA rank 128)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

The key distinction: a standard LoRA activates on a text token. IC-LoRA activates a reference input at inference time. You don’t need a trigger word \u2014 you hand it a photo and voice clip, and the model does the rest.<\/p>\n\n\n\n

Prerequisites (Dataset, VRAM, ltx-trainer Version)<\/h2>\n\n\n\n

Hardware:<\/strong> An H100 80GB is the officially documented reference. In practice, an RTX 4090 (24GB) works with gradient checkpointing enabled and resolution capped at 768\u00d7432. IC-LoRA at rank 128 (the ID-LoRA configuration) requires more memory than rank 32 standard LoRA \u2014 budget for slower training or cloud GPU time. A100 80GB on RunPod or vast.ai runs the full job in under 3 hours at reasonable cost.<\/p>\n\n\n\n

ltx-trainer version:<\/strong> Use the version from the Lightricks\/LTX-2 monorepo<\/a>, specifically the ltx-trainer<\/code> package. For the identity-audio variant (ID-LoRA), use the ID-LoRA\/ID-LoRA repository which builds on top of ltx-trainer <\/a>with the audio_ref_only_ic<\/code> training strategy. As of March 24, 2026, native ComfyUI support for ID-LoRA was merged upstream (PR #13111) \u2014 you no longer need a custom node fork.<\/p>\n\n\n\n

Model assets required:<\/strong><\/p>\n\n\n\n

LTX-2.3 base checkpoint (~44 GB)       \u2192 models\/checkpoints\/\nGemma text encoder (~6 GB)             \u2192 models\/text_encoders\/\nSpatial upscaler (~700 MB)             \u2192 models\/latent_upscale_models\/\nTemporal upscaler                      \u2192 models\/latent_upscale_models\/\nDistilled LoRA (~900 MB)               \u2192 models\/loras\/<\/code><\/pre>\n\n\n\n
Download all assets from HuggingFace<\/a> before starting. The distilled LoRA is required for the two-stage pipeline that IC-LoRA uses at inference. The official LTX-2 model card on HuggingFace<\/a> lists all required assets with direct download links and notes that IC-LoRA training in many settings takes under an hour.<\/p>\n\n\n\n
\n
\"\"<\/figure>\n<\/figure>\n\n\n\n
Training an IC-LoRA with ltx-trainer<\/h2>\n\n\n\n
Dataset Requirements for Consistent Identity<\/h3>\n\n\n\n
This is where IC-LoRA training succeeds or fails. The dataset requirement is more demanding than standard LoRA because you’re teaching a paired relationship \u2014 what goes in (reference) must clearly correspond to what comes out (generated video).<\/p>\n\n\n\n
Minimum viable dataset:<\/strong><\/p>\n\n\n\n