Tencent HunyuanVideo-1.5: 8.3B Video Model Runs on 14GB GPUs

Tencent’s HunyuanVideo-1.5 punches well above its weight class. This 8.3-billion-parameter model delivers visual fidelity and motion coherence that previously required data-center-grade hardware. It now runs on consumer GPUs with as little as 14GB of VRAM, significantly lowering the barrier to professional video synthesis.

The architecture features a Diffusion Transformer (DiT) paired with a 3D causal VAE, compressing spatial dimensions by 16x and temporal dimensions by 4x. The primary differentiator is Selective Sliding Tile Attention (SSTA). This mechanism prunes redundant spatiotemporal KV blocks to slash compute overhead for long sequences. When generating a 10-second 720p clip, SSTA performs 1.87 times faster than FlashAttention-3.

The model supports native text-to-video (T2V) and image-to-video (I2V) workflows. An integrated lightweight super-resolution network upscales output to 1080p, sharpening edges and correcting distortion without requiring a heavy, separate post-processing step. Community adoption is already robust, with native optimizations available in ComfyUI and LightX2V.

Requirements

• GPU: NVIDIA with CUDA support. 14GB VRAM minimum (with model offloading enabled).
• OS: Linux.
• Python: 3.10 or newer.
• CUDA: Must match your PyTorch version.

Additional VRAM allows you to disable offloading to increase inference speeds.

Setup and Installation

Clone the repository and enter the directory:

git clone https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5.git
cd HunyuanVideo-1.5

Install base dependencies:

pip install -r requirements.txt
pip install -i https://mirrors.tencent.com/pypi/simple/ --upgrade tencentcloud-sdk-python

Select one of the following attention libraries for optimal performance:

• Flash Attention: Reduces memory usage and increases inference speed. Refer to official documentation for installation.
• Flex-Block-Attention: Required for sparse attention.

  git clone https://github.com/Tencent-Hunyuan/flex-block-attn.git
  cd flex-block-attn
  git submodule update --init --recursive
  python3 setup.py install
  

• SageAttention: Note that this disables Flex-Block-Attention when active.

  git clone https://github.com/cooper1637/SageAttention.git
  cd SageAttention
  export EXT_PARALLEL=4 NVCC_APPEND_FLAGS="--threads 8" MAX_JOBS=32
  python3 setup.py install
  

Model Download

Distilled and sparse attention models are forthcoming. Check the Hugging Face model card for updates. Follow checkpoints-download.md to pull the pretrained weights before attempting video generation.

Prompt Rewriting

Detailed prompts yield the best results. HunyuanVideo-1.5 includes a rewrite mechanism that refines user input via a vLLM-compatible language model. Qwen3 models are recommended:

• T2V: Qwen3-235B-A22B-Thinking-2507
• I2V: Qwen3-VL-235B-A22B-Instruct

Set environment variables to point to your vLLM endpoint:

export T2V_REWRITE_BASE_URL="<your-vllm-server>"
export T2V_REWRITE_MODEL_NAME="<your-model-name>"
export I2V_REWRITE_BASE_URL="<your-vllm-server>"
export I2V_REWRITE_MODEL_NAME="<your-model-name>"

Rewriting is enabled by default; use --rewrite false to bypass it.

Running a Generation

Example command for T2V or I2V:

PROMPT='a girl holding a paper reading "Hello, world!"'
IMAGE_PATH=none
SEED=1
ASPECT_RATIO=16:9
RESOLUTION=480p
OUTPUT_PATH=./outputs/output.mp4

REWRITE=true
N_INFERENCE_GPU=8
CFG_DISTILLED=true
SPARSE_ATTN=false
SAGE_ATTN=true
OVERLAP_GROUP_OFFLOADING=true
ENABLE_CACHE=true
CACHE_TYPE=deepcache
ENABLE_SR=true
MODEL_PATH=ckpts

torchrun --nproc_per_node=$N_INFERENCE_GPU generate.py \
 --prompt "$PROMPT" \
 --image_path $IMAGE_PATH \
 --resolution $RESOLUTION \
 --aspect_ratio $ASPECT_RATIO \
 --seed $SEED \
 --rewrite $REWRITE \
 --cfg_distilled $CFG_DISTILLED \
 --sparse_attn $SPARSE_ATTN --use_sageattn $SAGE_ATTN \
 --enable_cache $ENABLE_CACHE --cache_type $CACHE_TYPE \
 --overlap_group_offloading $OVERLAP_GROUP_OFFLOADING \
 --sr $ENABLE_SR --save_pre_sr_video \
 --output_path $OUTPUT_PATH \
 --model_path $MODEL_PATH

Troubleshooting

• OOM errors with >14GB VRAM: Run export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:128 before execution.
• CPU memory OOM: Disable offloading with --overlap_group_offloading false.

Optimal Configurations

Model Variants

Performance Benchmarks

Tested on 8 H800 GPUs with standard engineering acceleration:

The model’s multi-stage training employs a Muon optimizer to accelerate convergence. The pipeline transitions from initial pretraining to post-training refinement, improving motion continuity and visual quality while aligning outputs with user preferences. Evaluation metrics emphasize text-video alignment, structural stability, and frame-level aesthetics.

HunyuanVideo-1.5 represents a major shift: 8 billion parameters running on 14GB of VRAM with a single command. The gap between high-end research labs and local workstations continues to narrow.