TurboT2AV

TurboT2AV

Fast text-to-audio-video generation distilled from LTX-2 19B.

TurboDiffusion-Style Acceleration

Measured on a single NVIDIA H20 at 1024x1792:

Stage	Latency	Speedup vs previous	Speedup vs teacher	What changes
LTX-2-19B teacher (40 steps)	318.7405s	-	1.00x	Full teacher baseline with dense attention.
+ W8A8 & FastNorm	233.3424s	1.37x	1.37x	Add TileLang W8A8 Linear and FastNorm to the teacher.
+ rCM (4-step student)	12.1655s	19.18x	26.20x	Switch to the distilled student while retaining W8A8/FastNorm.
+ SageSLA final	5.8505s	2.08x	54.48x	Add SageSLA topk=0.3 self-attention and text-context trimming.

At this resolution the video latent is [1,16,128,32,56], corresponding to 28,672 video self-attention tokens. The stages are cumulative: rCM keeps the W8A8/FastNorm stack, and the final stage adds SageSLA and text-context trimming. For reference, the pure 4-step student without these inference optimizations takes 16.5245s/video, so the final path is also 2.82x faster than the pure student.

See TurboDiffusion Integration Notes for the reused components, LTX-2-specific adaptations, and interpretation of these results.

Overview

TurboT2AV generates synchronized audio-video from text prompts in 4 steps. The demo compares the 40-step teacher with the 4-step student. This repository provides single-GPU inference for the distilled checkpoint. On an NVIDIA H20 at 1024x1792, generator-only latency falls from 318.74 seconds/video for the 40-step teacher to 5.85 seconds/video for the accelerated 4-step student.

Main contributions:

• Combines the diversity of consistency models (DCM/SCM) with the high perceptual quality of score-model distillation (DMD), taking advantage of both families of methods by using CM as a forward-divergence offline method that complements DMD as a reverse-KL on-policy method.
• First extends this combined distillation strategy to a large-scale joint audio-video generation model at the 14B-video + 5B-audio scale.
• Integrates a TurboDiffusion-style inference stack with SageSLA, FastNorm, and TileLang W8A8 Linear. On a single NVIDIA H20 at 1024x1792, the final accelerated student is 54.48x faster than the 40-step teacher and 2.82x faster than the pure 4-step student.

1. Setup

cd TurboDiffusion/TurboT2AV/LTX-2
pixi install
pixi run install-acceleration

This single task installs the local LTX packages, CUDA 12.8 PyTorch, SageAttention, SpargeAttn, and TileLang. It provides everything required by the recommended SageSLA + FastNorm + TileLang W8A8 inference path.

Development Tests (Optional)

Run the unit tests in the development environment with:

pixi run -e dev test

The test task installs CUDA 12.8 PyTorch and the local LTX packages in the development environment before running the test suite. Inference-only users do not need this step.

2. Download Weights

Model Name	Checkpoint Link
TurboT2AV-14BVideo-5BAudio	Hugging Face Model
LTX-2-19B	Hugging Face Model
Gemma-3-12B-IT-QAT-Q4_0	Hugging Face Model

Gemma is a gated Hugging Face model. Before downloading, visit the model page, accept the access terms, and export a Hugging Face token with access permission:

export HF_TOKEN=your_huggingface_token

Base model weights:

pixi run hf download Lightricks/LTX-2 ltx-2-19b-dev.safetensors --local-dir /path/to/checkpoints/LTX-2
pixi run hf download google/gemma-3-12b-it-qat-q4_0-unquantized --local-dir /path/to/checkpoints/gemma-3-12b-it-qat-q4_0-unquantized

TurboT2AV main checkpoint:

pixi run hf download luyu1021/TurboT2AV \
  --include "checkpoints/turbot2av_main/*" \
  --local-dir /path/to/turbo-t2av-weights

3. Run Inference

Run the following commands from TurboDiffusion/TurboT2AV/LTX-2:

export TURBO_CHECKPOINT_PATH=/path/to/ltx-2-19b-dev.safetensors
export TURBO_GEMMA_PATH=/path/to/gemma-3-12b-it-qat-q4_0-unquantized
export PYTHONPATH=../..:../../turbodiffusion:$PYTHONPATH

--prompts_file accepts a text file with one prompt per line or a CSV file with a prompt column.

Accelerated Student (Recommended)

CUDA_VISIBLE_DEVICES=0 pixi run python -m ltx_distillation.tools.run_av_inference_eval \
  --config_path packages/ltx-distillation/configs/bidirectional_rcm.yaml \
  --prompts_file /path/to/prompts.csv \
  --output_dir /path/to/student_output \
  --model_kind student \
  --student_checkpoint /path/to/turbo-t2av-weights/checkpoints/turbot2av_main/model.pth \
  --student_param auto \
  --num_prompts 8 \
  --video_height 1024 \
  --video_width 1792 \
  --attention_type sagesla \
  --attention_scope self \
  --sla_topk 0.3 \
  --trim_text_context \
  --fast_norm \
  --quant_linear \
  --quant_linear_scope all \
  --quant_linear_backend tilelang_postscale

Teacher Baseline (40 Steps)

CUDA_VISIBLE_DEVICES=0 pixi run python -m ltx_distillation.tools.run_av_inference_eval \
  --config_path packages/ltx-distillation/configs/bidirectional_rcm.yaml \
  --prompts_file /path/to/prompts.csv \
  --output_dir /path/to/teacher_output \
  --model_kind teacher \
  --teacher_mode native_rf \
  --teacher_steps 40 \
  --num_prompts 8 \
  --video_height 1024 \
  --video_width 1792

Demos

Teacher (40 steps)	Student (4 steps)
sample_0000_teacher.mp4	sample_0000_student.mp4
sample_0001_teacher.mp4	sample_0001_student.mp4
sample_0002_teacher.mp4	sample_0002_student.mp4
sample_0003_teacher.mp4	sample_0003_student.mp4
sample_0004_teacher.mp4	sample_0004_student.mp4
sample_0005_teacher.mp4	sample_0005_student.mp4
sample_0006_teacher.mp4	sample_0006_student.mp4