kailasa-ngpt/sph_ltx2_5

SPH - LTX-2 LoRA

A LoRA fine-tune of Lightricks/LTX-2 trained using Ostris AI Toolkit.

Model Details

Parameter	Value
Base Model	Lightricks/LTX-2
Type	LoRA
Trigger Word	SPH
Resolution	768x768
Frames	121 (~5s at 24fps)
Training Steps	2000
Precision	bf16

LoRA Configuration

Parameter	Value
Linear Rank	32
Linear Alpha	16
Conv Rank	16
Conv Alpha	16

Training Details

• Hardware: NVIDIA A100 SXM 80GB (RunPod)
• Optimizer: AdamW 8-bit (weight decay: 0.0001)
• Learning Rate: 1e-4
• Batch Size: 1 (gradient accumulation: 4, effective batch size: 4)
• Noise Scheduler: Flow Matching
• Timestep Type: Weighted
• Content/Style: Balanced
• EMA: Enabled (decay: 0.99)
• Gradient Checkpointing: Enabled
• Loss: MSE
• Caption Dropout: 10%

The model was trained on both text-to-video and image-to-video datasets with 2 repeats each.

Usage

Use the trigger word SPH in your prompt to activate the LoRA.

Example Prompts

SPH, smiling and blessing with both hands up, seated in golden ceremonial attire, looking at camera.

SPH sitting on the golden throne with both legs folded and hands in a blessing posture, ornate jewelry, devotional setting.

Files

• sph_ltx2_5.safetensors - LoRA weights
• finetune-config.yaml - Full training configuration

Framework

• Training Framework: Ostris AI Toolkit
• Inference: Diffusers