dustin2050/lexy-vox-ltx2.3-lora

Lexy_Vox Character LoRA for LTX-2.3

A character-specific LoRA for Lightricks LTX-2.3 (22B), trained on 29 reference images of the AI-generated character "Lexy_Vox".

Character

Lexy_Vox — young woman with long dark wavy hair, green-hazel eyes, defined cheekbones, olive/tan skin, athletic build.

Trigger word: lexy_vox

Checkpoints

Three checkpoints are provided from one training run (rank 32, FFN + attention modules):

File	Steps	Best for
lexy_vox_v3_step1000.safetensors	1000	Subtle character influence, less stylization
lexy_vox_v3_step1500.safetensors	1500	Recommended — best identity/quality balance
lexy_vox_v3_step2000.safetensors	2000	Maximum identity sharpness, slightly more baked-in style

Recommended Inference Setup

Daily content (I2V with motion)

• Base: Lightricks/LTX-2.3 dev checkpoint
• Required: ltx-2.3-22b-distilled-lora-384-1.1 at strength 1.0 (enables 8-step inference)
• Character: lexy_vox_v3_step1500.safetensors at strength 1.0
• Optional motion booster: ltx-2.3-22b-distilled-1.1_lora-dynamic_rank_111_bf16 at strength 0.6 (works in I2V mode)
• Resolution: 768x768 or 768x1360
• Frames: 97 (4s @ 24fps) up to 145 (6s @ 24fps)
• Steps: 8-12 (with distilled-LoRA)

Clean studio T2V (no reference image)

• Same setup, omit Dynamic-LoRA (it leaks party/setting bias in T2V mode)
• Use lexy_vox_v3_step1500 directly

Avoid

• T2V + Dynamic-LoRA → brings party/lights/extra-people background from Dynamic-LoRA training data
• Stacking Dynamic-LoRA at strength > 0.6 → causes facial aging artifacts

Training Details

Parameter	Value
Base model	Lightricks/LTX-2.3-22B-dev
Text encoder	Gemma-3-12B-IT (DreamFast heretic-v2 mirror)
Training framework	ltx-trainer from Lightricks/LTX-2 monorepo
Hardware	NVIDIA RTX PRO 6000 Blackwell (96GB, sm_120)
Time	49.6 minutes (2000 steps, 1.49 s/step)
Dataset	23 face-cropped 768×768 images + 23 horizontal-flip augmentations = 46 samples
Filtered out	Multi-person photos (red carpet/photographers), party/restaurant/cafe settings, watermarked images
Captions	Pure-identity prompts (5 rotating templates) — no settings or activities
LoRA rank	32
LoRA alpha	32
Target modules	to_k, to_q, to_v, to_out.0, ff.net.0.proj, ff.net.2
Trainable params	276M
Optimizer	AdamW
Learning rate	3e-5
LR scheduler	Cosine
Mixed precision	bfloat16
Batch size	1 (grad_accum=4 → effective 4)
Final loss	0.81
First-frame conditioning	p=0.5 (works in both T2V and I2V)

Version History

This is v3 — version notes from the iteration:

• v1 (rank 32, 500 steps, raw captions): aging artifacts when used with Dynamic-LoRA; T2V showed 2 people + restaurant setting due to "photographers in background" caption baked in
• v2 (rank 16, 1500 steps, identity-stripped captions): aging fixed; T2V cleaner but still had subtle Setting-leak (e.g. "5-Minute Crafts" watermark from training image)
• v3 (this version, rank 32 + FFN, face-crops, augmentation, cosine schedule): all setting bias eliminated, identity stable in I2V (>4s), clean studio output in T2V

ComfyUI Usage (API)

workflow_partial = {
    "ckpt": {"class_type": "CheckpointLoaderSimple",
             "inputs": {"ckpt_name": "ltx-2.3-22b-dev.safetensors"}},
    "lora_distilled": {"class_type": "LoraLoaderModelOnly",
                       "inputs": {"model": ["ckpt", 0],
                                  "lora_name": "ltx-2.3-22b-distilled-lora-384-1.1.safetensors",
                                  "strength_model": 1.0}},
    "lora_character": {"class_type": "LoraLoaderModelOnly",
                       "inputs": {"model": ["lora_distilled", 0],
                                  "lora_name": "lexy_vox_v3_step1500.safetensors",
                                  "strength_model": 1.0}},
    # ... continue with LTXAVTextEncoderLoader, CLIPTextEncode, LTXVImgToVideo, etc.
}

License

Apache-2.0 for the LoRA weights. The base LTX-2.3 model is subject to its own Lightricks license — review at Lightricks/LTX-2.3.

Acknowledgements

• Lightricks for the LTX-2.3 model and ltx-trainer framework
• DreamFast for the non-gated Gemma-3-12B-IT mirror
• Trained on vast.ai cloud GPU