hunyuan nsfw on Anyken: compare the model examples with your intended input and output. Review the capabilities documented below, then use the controls actually offered in the workspace. Available settings and credits are shown before submission.
Read source
The HunyuanVideo framework is a systematic approach to large-scale video generation, released as open-source by Tencent. The following items describe its core architecture and key features as documented in the official repository.
HunyuanVideo introduces a Transformer-based design with a Full Attention mechanism for unified image and video generation. Hunyuan nsfw employs a 'Dual-stream to Single-stream' hybrid architecture. In the dual-stream phase, video and text tokens are processed independently through multiple Transformer blocks, allowing each modality to learn its own modulation mechanisms without interference. In the single-stream phase, video and text tokens are concatenated and fed into subsequent Transformer blocks for effective multimodal information fusion. This design captures complex interactions between visual and semantic information, enhancing overall model performance. The model is trained on a spatial-temporally compressed latent space using a Causal 3D VAE, which compresses pixel-space videos and images into a compact latent space with compression ratios of 4 for video length, 8 for space, and 16 for channel. This significantly reduces the number of tokens for the subsequent diffusion transformer model, enabling training at original resolution and frame rate.
Unlike previous text-to-video models that typically use pre-trained CLIP and T5-XXL as text encoders, HunyuanVideo utilizes a pre-trained Multimodal Large Language Model (MLLM) with a Decoder-Only structure as its text encoder. This approach offers several advantages: MLLM after visual instruction finetuning has better image-text alignment in the feature space, alleviating the difficulty of instruction following in diffusion models; MLLM has demonstrated superior ability in image detail description and complex reasoning; and MLLM can act as a zero-shot learner by following system instructions prepended to user prompts, helping text features pay more attention to key information. Because MLLM uses causal attention while T5-XXL utilizes bidirectional attention that produces better text guidance for diffusion models, the framework introduces an extra bidirectional token refiner to enhance text features. Additionally, hunyuan nsfw includes a prompt rewrite system fine-tuned from the Hunyuan-Large model, offering two modes: Normal mode, designed to enhance comprehension of user intent, and Master mode, which enhances descriptions of composition, lighting, and camera movement for higher visual quality.
The official documentation details hunyuan nsfw's technical capabilities, system requirements, and access methods. The following items cover performance benchmarks, hardware requirements, and inference options.
The HunyuanVideo model was evaluated against five strong baselines from closed-source video generation models using 1,533 text prompts, generating an equal number of video samples in a single run without cherry-picking. Videos were assessed by more than 60 professional evaluators based on three criteria: Text Alignment, Motion Quality, and Visual Quality. According to the published results, HunyuanVideo demonstrated the best overall performance, particularly excelling in motion quality. The evaluation is based on the high-quality version of hunyuan nsfw, which differs from the currently released fast version. The model has over 13 billion parameters, making it the largest among all open-source video generation models at the time of release. The official comparison table shows HunyuanVideo outperforming closed-source models including Runway Gen-3, Luma 1.6, and three top-performing Chinese video generative models in overall ranking.
The official documentation specifies hardware requirements for running the HunyuanVideo model with a batch size of 1. For generating videos at 720px by 1280px with 129 frames, the GPU peak memory requirement is 60GB. For 544px by 960px with 129 frames, the requirement is 45GB. An NVIDIA GPU with CUDA support is required, and hunyuan nsfw is tested on a single 80G GPU. The recommended GPU memory is 80GB for better generation quality. The tested operating system is Linux. The repository provides installation instructions for CUDA versions 12.6 or 11.8, along with a pre-built Docker image. For multi-GPU parallel inference, the repository supports xDiT, a scalable inference engine for Diffusion Transformers, with various parallel configurations documented for different video sizes and frame counts. FP8 inference is also supported, saving approximately 10GB of GPU memory compared to full precision.
The HunyuanVideo project is released as open-source with code, pre-trained weights, and inference tools available. The following items describe the release history and community ecosystem.
The HunyuanVideo project has seen multiple releases expanding its capabilities. The initial inference code and model weights were released, followed by FP8 model weights to save GPU memory. Subsequent releases include the HunyuanVideo-I2V image-to-video model, HunyuanCustom for customized video generation, HunyuanVideo-Avatar for audio-driven human animation, and HunyuanVideo-1.5 as a new foundation model. The repository also includes the Penguin Video Benchmark for evaluation. Hunyuan nsfw has been integrated into Diffusers, and parallel inference code powered by xDiT is available. The open-source plan includes inference checkpoints, multi-gpu sequence parallel inference, a Gradio web demo, FP8 quantized weights, and ComfyUI support. The repository provides PyTorch model definitions, pre-trained weights, and inference/sampling code for the paper exploring HunyuanVideo.
The HunyuanVideo repository lists numerous community contributions that extend hunyuan nsfw's functionality. These include ComfyUI wrappers for FP8 inference, video-to-video and image-to-video generation, and native ComfyUI support. FastVideo provides a consistency distilled model and sliding tile attention. GGUF version and quantization tools are available through HunyuanVideo-gguf. Enhance-A-Video offers improved generated video quality without additional cost. TeaCache provides cache-based acceleration, and RIFLEx enables video length extrapolation. Other contributions include keyframe control LoRA, Sparse-VideoGen for accelerated generation with high pixel-level fidelity, FramePack for packing input frame context, Jenga for training-free efficient video generation, DCM for dual-expert consistency model, and NaviCache for offline calibration-free cache-based acceleration. These community tools demonstrate the active ecosystem around the model.
The official documentation provides detailed configuration options for video generation. The following items describe supported resolutions, inference parameters, and parallel processing setups.
The HunyuanVideo model supports multiple resolution configurations with aspect ratios including 9:16, 16:9, 4:3, 3:4, and 1:1. For 540p resolution, supported settings include 544x960 with 129 frames for 9:16, 960x544 for 16:9, 624x832 for 4:3, 832x624 for 3:4, and 720x720 for 1:1. For 720p resolution, which is recommended, supported settings include 720x1280 for 9:16, 1280x720 for 16:9, 1104x832 for 4:3, 832x1104 for 3:4, and 960x960 for 1:1. All configurations generate 129 frames. The model uses 50 inference steps by default, with an embedded classifier-free guidance scale of 6.0 and a flow shift factor of 7.0. The flow-reverse parameter determines whether sampling learns from t=1 to t=0. CPU offload is available to save memory for high-resolution video generation.
For multi-GPU parallel inference using xDiT, the repository documents specific parallel configurations for different video sizes and GPU counts. For 1280x720 or 720x1280 videos with 129 frames, supported configurations include 8x1, 4x2, 2x4, and 1x8 for 8 GPUs; 1x5 for 5 GPUs; 4x1, 2x2, and 1x4 for 4 GPUs; 3x1 and 1x3 for 3 GPUs; and 2x1 and 1x2 for 2 GPUs. Similar configurations are provided for other resolutions including 1104x832, 960x960, 960x544, 832x624, and 720x720. The latency measurements for 1280x720 videos with 129 frames and 50 steps on 8 GPUs show scaling from 1904.08 seconds on 1 GPU to 337.58 seconds on 8 GPUs, representing a 5.64x speedup. Users can control the ulysses-degree and ring-degree parameters to optimize performance for their specific hardware setup.
Developers working with the HunyuanVideo model should be aware of several practical aspects documented in the official repository. The following items cover installation, memory management, and content policy considerations.
The official documentation provides detailed installation instructions for Linux systems. The recommended CUDA versions are 12.6 or 11.8. The installation process involves creating a conda environment with Python 3.10.9, installing PyTorch 2.7.1 with the appropriate CUDA version, installing pip dependencies from requirements.txt, and installing flash attention v2 for acceleration. For parallel inference, xDiT version 0.4.0 is required. The repository also provides pre-built Docker images for both CUDA 12.4 and CUDA 11.8, which can help avoid float point exceptions on specific GPU types. The Docker images include necessary environment configurations such as seccomp settings and memory limits. For users experiencing float point exceptions, the documentation suggests either using CUDA 12.4 with specific CUBLAS and CUDNN versions, or forcing the use of CUDA 11.8 compiled versions of PyTorch and all other packages.
The official HunyuanVideo documentation focuses on technical capabilities, architecture, and usage instructions. The GitHub repository and research paper describe model training, inference, and evaluation but do not establish a specific adult-content policy or built-in NSFW filtering mechanism. The documentation mentions a prompt rewrite system that can be used to enhance or modify prompts, but does not specify content restrictions. The model's training data curation process is mentioned as a key technology for model learning, but specific filtering criteria are not detailed in the available documentation. Users seeking information about content policies should refer to the current service terms and conditions of the platforms or services through which they access the model. The open-source nature of the model means that users can deploy it locally with their own content policies and moderation systems as needed.
Common questions about the HunyuanVideo model and its documentation.
The search query 'hunyuan nsfw' typically refers to questions about adult content support in Tencent's HunyuanVideo model. The official technical documentation, including the GitHub repository and research paper, describes the model's architecture, training methodology, inference code, and system requirements. However, these sources do not establish a specific adult-content policy or built-in NSFW filtering mechanism. The documentation mentions a prompt rewrite system and data curation processes but does not detail content restrictions. Users should consult the current service terms of the platforms through which they access the model for content policy information. The open-source release allows users to deploy the model locally with their own content moderation systems.
The official documentation specifies that an NVIDIA GPU with CUDA support is required. For generating videos at 720px by 1280px with 129 frames, the minimum GPU memory requirement is 60GB. For 544px by 960px with 129 frames, the requirement is 45GB. The recommended GPU memory is 80GB for better generation quality. The model is tested on a single 80G GPU, and the tested operating system is Linux. For multi-GPU setups, the repository supports parallel inference using xDiT with various configurations documented for different video sizes and GPU counts.
HunyuanVideo supports multiple resolution configurations with aspect ratios including 9:16, 16:9, 4:3, 3:4, and 1:1. For 540p resolution, supported settings include 544x960, 960x544, 624x832, 832x624, and 720x720, all with 129 frames. For the recommended 720p resolution, supported settings include 720x1280, 1280x720, 1104x832, 832x1104, and 960x960, also with 129 frames. The model uses 50 inference steps by default and supports various generation parameters including classifier-free guidance scale and flow shift factor.
