hunyuan video 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
HunyuanVideo is a systematic framework for large video generation models, released as open-source by Tencent. The repository includes PyTorch model definitions, pre-trained weights, and inference code. The documentation does not address content moderation or NSFW filtering policies.
HunyuanVideo is presented as a novel open-source video foundation model that exhibits performance comparable to leading closed-source models. The framework adopts key technologies including data curation, image-video joint model training, and an efficient infrastructure for large-scale training and inference. Hunyuan video nsfw has over 13 billion parameters, making it the largest among all open-source video generation models at the time of release. According to professional human evaluation results cited in the repository, HunyuanVideo outperformed previous state-of-the-art models including Runway Gen-3, Luma 1.6, and three top-performing Chinese video generative models. The evaluation used 1,533 text prompts, generating an equal number of video samples in a single run without cherry-picking. Videos were assessed on Text Alignment, Motion Quality, and Visual Quality by more than 60 professional evaluators. The repository releases code and weights of the foundation model and its applications to bridge the gap between closed-source and open-source video foundation models.
The HunyuanVideo architecture introduces a Transformer design with Full Attention mechanism for unified image and video generation. It employs a 'Dual-stream to Single-stream' hybrid model design. 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 multimodal information fusion. Hunyuan video nsfw uses a pre-trained Multimodal Large Language Model (MLLM) with a Decoder-Only structure as its text encoder, which provides better image-text alignment, superior ability in image detail description and complex reasoning, and zero-shot learning capability by following system instructions. An extra bidirectional token refiner enhances text features. The 3D VAE with CausalConv3D 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. A prompt rewrite model fine-tuned from Hunyuan-Large adapts user prompts to model-preferred prompts in Normal mode and Master mode.
The repository documents technical capabilities including text-to-video generation, image-to-video generation, and various inference configurations. Access to hunyuan video nsfw is through open-source code and weights, with no documented content filtering or NSFW policy in the technical documentation.
The HunyuanVideo repository supports multiple generation tasks. The primary text-to-video model generates videos from text prompts. An image-to-video model (HunyuanVideo-I2V) was released on March 6, 2025, based on HunyuanVideo. The HunyuanVideo-Avatar model, released on May 28, 2025, is an audio-driven human animation model. HunyuanCustom, released on May 9, 2025, is a multimodal-driven architecture for customized video generation. The repository also includes a prompt rewrite model that fine-tunes Hunyuan-Large to adapt user prompts to model-preferred prompts, with Normal mode for enhancing comprehension and Master mode for improving composition, lighting, and camera movement descriptions. Hunyuan video nsfw supports various resolutions including 720px1280px129f (recommended), 544px960px129f, and multiple aspect ratios (9:16, 16:9, 4:3, 3:4, 1:1) at 540p and 720p. The FP8 quantized weights save about 10GB of GPU memory. Community contributions include integrations with ComfyUI, FastVideo, GGUF version, Enhance-A-Video, TeaCache, and other tools.
The repository specifies system requirements for running HunyuanVideo. For a batch size of 1, generating 720px1280px129f videos requires 60GB GPU peak memory, while 544px960px129f requires 45GB. An NVIDIA GPU with CUDA support is required, tested on a single 80G GPU. The minimum GPU memory is 60GB for 720px1280px129f and 45GB for 544px960px129f. Recommended GPU memory is 80GB for better generation quality. The tested operating system is Linux. Installation requires CUDA versions 12.6 or 11.8, Python 3.10.9, PyTorch 2.7.1, and flash attention v2.8.3. Pre-built Docker images are available for CUDA 12.4 and CUDA 11.8. Single-GPU inference supports multiple resolutions and frame lengths. Parallel inference on multiple GPUs is supported via xDiT with Unified Sequence Parallelism APIs, with configurations for 2 to 8 GPUs. For example, generating 1280x720 videos with 129 frames and 50 steps on 8 GPUs achieves 337.58 seconds latency (5.64x speedup over single GPU). FP8 inference saves about 10GB of GPU memory.
HunyuanVideo is distributed as open-source software through GitHub. The repository provides model weights, inference code, and documentation. The technical documentation does not establish an adult-content policy or NSFW filtering.
The HunyuanVideo repository is hosted on GitHub at https://github.com/Tencent-Hunyuan/HunyuanVideo. It contains PyTorch model definitions, pre-trained weights, and inference/sampling code. The open-source plan includes inference checkpoints, multi-GPU sequence parallel inference, web demo (Gradio), Diffusers integration, FP8 quantized weights, Penguin Video Benchmark, and ComfyUI support. The repository has been integrated into Diffusers as of December 17, 2024. Parallel inference code powered by xDiT was released on December 7, 2024. The initial inference code and model weights were released on December 3, 2024. HunyuanVideo-1.5, a highly efficient and powerful new foundation model, was released on November 21, 2025. The repository also includes a prompt rewrite model with weights available for download. Community contributions include various tools and integrations such as ComfyUI wrappers, FastVideo, GGUF quantization, Enhance-A-Video, TeaCache, and others. The repository provides a BibTeX citation for academic use.
The repository includes comprehensive documentation covering abstract, architecture overview, key features, comparisons with other models, requirements, dependencies and installation guides, download instructions for pretrained models, single-GPU and multi-GPU inference commands, FP8 inference, and BibTeX citation. The documentation also provides links to community contributions and integration projects. Users can join WeChat and Discord channels for community support. The project page offers additional visualizations. The repository acknowledges contributions from SD3, FLUX, Llama, LLaVA, Xtuner, diffusers, and HuggingFace repositories, as well as the Tencent Hunyuan Multimodal team for help with the text encoder. The documentation does not include any content moderation policy, NSFW filtering guidelines, or terms of service. Users seeking information about content policies should refer to the current service terms of the hosting platform or hunyuan video nsfw provider's official policies.
The repository documents various inference configurations including resolution settings, sampling parameters, and parallel processing options. These technical details are available for users to configure their own generation workflows.
Single-GPU inference supports multiple resolution and frame configurations. Supported resolutions include 544px960px129f, 960px544px129f, 624px832px129f, 832px624px129f, 720px720px129f for 540p, and 720px1280px129f, 1280px720px129f, 1104px832px129f, 832px1104px129f, 960px960px129f for 720p (recommended). The inference command uses sample_video.py with arguments for video size, video length (129 frames), inference steps (default 50), prompt, flow-reverse flag, CPU offload option, and save path. Additional configurations include embedded classifier-free guidance scale (default 6.0), flow shift factor for flow matching schedulers (default 7.0), random seed (optional), and CPU offload for high-resolution video generation. A Gradio server can be run using gradio_server.py with flow-reverse flag. The server name and port can be set manually. The prompt rewrite model can be deployed and inferred using Hunyuan-Large original code, with weights available for download.
Multi-GPU parallel inference is supported via xDiT, a scalable inference engine for Diffusion Transformers on multi-GPU clusters. The repository adopted Unified Sequence Parallelism (USP) APIs for parallel inference. Parallel configurations are controlled by ulysses-degree and ring-degree parameters. Supported configurations include 8x1, 4x2, 2x4, 1x8 for 8 GPUs; 1x5 for 5 GPUs; 4x1, 2x2, 1x4 for 4 GPUs; 3x1, 1x3 for 3 GPUs; and 2x1, 1x2 for 2 GPUs. These configurations apply to various resolutions including 1280x720, 720x1280, 1104x832, 832x1104, 960x960, 960x544, 544x960, 832x624, 624x832, and 720x720. Latency benchmarks for 1280x720 with 129 frames and 50 steps show 1904.08 seconds on 1 GPU, 934.09 seconds (2.04x) on 2 GPUs, 514.08 seconds (3.70x) on 4 GPUs, and 337.58 seconds (5.64x) on 8 GPUs. FP8 inference is also supported, requiring explicit specification of the FP8 weight path and using the --use-fp8 flag.
Users should be aware of the technical requirements and limitations documented in the repository. The documentation focuses on technical capabilities and does not address content policies or NSFW filtering.
Running HunyuanVideo requires significant hardware resources. The minimum GPU memory is 60GB for 720px1280px129f generation and 45GB for 544px960px129f. A GPU with 80GB memory is recommended for better quality. Hunyuan video nsfw is tested on Linux operating system only. Installation requires CUDA versions 12.6 or 11.8, Python 3.10.9, and PyTorch 2.7.1. Flash attention v2.8.3 is required for acceleration. For parallel inference, xDiT version 0.4.0 is recommended with torch 2.7.1 and flash-attn 2.8.3. Users may encounter float point exceptions on specific GPU types; solutions include installing CUDA 12.4 with CUBLAS 12.4.5.8 and CUDNN 9.00, or forcing CUDA 11.8 compiled versions. Pre-built Docker images are available for CUDA 12.4 and CUDA 11.8 to simplify setup. The repository provides detailed installation guides for Linux, including conda environment creation and pip dependency installation.
The repository lists numerous community contributions that extend HunyuanVideo's capabilities. These include ComfyUI-HunyuanVideoWrapper for FP8 inference, V2V and IP2V generation; ComfyUI-HunyuanVideo native support; FastVideo for consistency distilled model and sliding tile attention; HunyuanVideo-gguf for GGUF version and quantization; Enhance-A-Video for improved generated video; TeaCache for cache-based acceleration; HunyuanVideoGP for GPU-poor version; RIFLEx for video length extrapolation; keyframe control LoRA; Sparse-VideoGen for accelerated video generation; FramePack for packing input frame context; Jenga for dynamic token carving; DCM for dual-expert consistency model; and NaviCache for offline calibration-free cache-based acceleration. Users should note that the repository documentation does not include content moderation policies, NSFW filtering, or terms of service. The evaluation results are based on the high-quality version of HunyuanVideo, which differs from the released fast version. Users seeking information about acceptable content should consult the current service terms of the platform or provider.
Common questions about HunyuanVideo based on the official repository documentation.
The official HunyuanVideo repository does not explicitly establish an adult-content policy or built-in NSFW filtering. The documentation focuses on technical capabilities, system requirements, and open-source access. Users seeking information about content policies should refer to the current service terms of the hosting platform or hunyuan video nsfw provider's official policies. The repository provides model weights, inference code, and documentation for video generation tasks including text-to-video, image-to-video, and audio-driven animation.
HunyuanVideo requires an NVIDIA GPU with CUDA support. The minimum GPU memory is 60GB for generating 720px1280px129f videos and 45GB for 544px960px129f videos. A GPU with 80GB memory is recommended for better generation quality. Hunyuan video nsfw is tested on Linux operating system. Installation requires CUDA versions 12.6 or 11.8, Python 3.10.9, and PyTorch 2.7.1. Flash attention v2.8.3 is required for acceleration. Pre-built Docker images are available for CUDA 12.4 and CUDA 11.8.
HunyuanVideo supports text-to-video generation as its primary task. Additional models include HunyuanVideo-I2V for image-to-video generation, HunyuanVideo-Avatar for audio-driven human animation, and HunyuanCustom for customized video generation. Hunyuan video nsfw supports multiple resolutions and aspect ratios including 9:16, 16:9, 4:3, 3:4, and 1:1 at 540p and 720p. A prompt rewrite model fine-tuned from Hunyuan-Large adapts user prompts in Normal mode and Master mode.
