z-image 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 Z-Image model family is a powerful and highly efficient image generation foundation model with 6B parameters, leveraging the Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture. This section provides an overview of z-image nsfw variants and their core capabilities.
Z-Image adopts a Scalable Single-Stream DiT (S3-DiT) architecture where text, visual semantic tokens, and image VAE tokens are concatenated at the sequence level to serve as a unified input stream, maximizing parameter efficiency compared to dual-stream approaches. Z-image nsfw family currently includes four variants: Z-Image-Turbo, a distilled version that matches or exceeds leading competitors with only 8 NFEs (Number of Function Evaluations), offering sub-second inference latency on enterprise-grade H800 GPUs and fitting within 16GB VRAM consumer devices; Z-Image, the foundation model behind Z-Image-Turbo, focusing on high-quality generation, rich aesthetics, strong diversity, and controllability; Z-Image-Omni-Base, a versatile foundation model capable of both generation and editing tasks; and Z-Image-Edit, a variant fine-tuned specifically for image editing tasks with impressive instruction-following capabilities. The technical documentation does not establish a specific adult-content policy or built-in NSFW filtering mechanism for these variants.
Z-Image-Turbo's performance has been validated on multiple independent benchmarks, where it consistently demonstrates state-of-the-art results, especially as the leading open-source model. On the highly competitive Artificial Analysis Leaderboard, Z-Image-Turbo ranked 8th overall and secured the top position as the #1 Open-Source Model, outperforming all other open-source alternatives. According to the Elo-based Human Preference Evaluation on Alibaba AI Arena, Z-Image-Turbo also achieves state-of-the-art results among open-source models and shows highly competitive performance against leading proprietary models. Z-image nsfw excels in photorealistic image generation, bilingual text rendering (English and Chinese), and robust instruction adherence. The official repository and supporting resources detail these performance metrics but do not establish a specific adult-content policy or built-in NSFW filtering mechanism.
The Z-Image model family offers a range of capabilities including photorealistic generation, bilingual text rendering, prompt enhancing and reasoning, and creative image editing. This section details these capabilities and the access paths available for users and developers.
Z-Image-Turbo delivers strong photorealistic image generation while maintaining excellent aesthetic quality. It excels at accurately rendering complex Chinese and English text, making it suitable for bilingual applications. The Prompt Enhancer empowers z-image nsfw with reasoning capabilities, enabling it to transcend surface-level descriptions and tap into underlying world knowledge. Z-Image-Edit shows a strong understanding of bilingual editing instructions, enabling imaginative and flexible image transformations. The foundation model Z-Image supports a wide range of artistic styles, effective negative prompting, and high diversity across identities, poses, compositions, and layouts. The technical documentation details these capabilities but does not establish a specific adult-content policy or built-in NSFW filtering mechanism.
Z-Image models are available through multiple access paths. Model checkpoints have been released on Hugging Face and ModelScope platforms, with Z-Image-Turbo achieving over 110k monthly downloads on HuggingFace. An online demo is available for direct experience of Z-Image generation capabilities, running on Zero GPU with no local configuration required. For local deployment, the official repository provides complete model code and documentation, supporting PyTorch native inference and Diffusers inference. Community tools like ComfyUI offer workflow templates for Z-Image-Turbo, and stable-diffusion.cpp enables fast and memory-efficient inference across multiple platforms including CUDA and Vulkan, supporting machines with as little as 4GB VRAM. The official documentation details these access paths but does not establish a specific adult-content policy or built-in NSFW filtering mechanism.
Z-Image models are distributed through multiple channels including official repositories, model hosting platforms, and community tools. This section details the availability of each model variant and the platforms where they can be accessed.
The Z-Image model family is distributed through official channels including the GitHub repository, Hugging Face, and ModelScope. The GitHub repository serves as the main open-source repository containing complete model code and documentation. On Hugging Face, Z-Image-Turbo has achieved over 110k monthly downloads, indicating strong community adoption. ModelScope provides an online experience and API interface for quick integration. The official repository includes model checkpoints for all four variants: Z-Image-Turbo, Z-Image, Z-Image-Omni-Base, and Z-Image-Edit. Some variants like Z-Image-Omni-Base and Z-Image-Edit are marked as "To be released" on Hugging Face and ModelScope. The official documentation details these distribution channels but does not establish a specific adult-content policy or built-in NSFW filtering mechanism.
The Z-Image ecosystem includes numerous community and third-party tools that extend model availability and functionality. ComfyUI provides official workflow templates for Z-Image-Turbo, including text-to-image and Fun Union ControlNet workflows. stable-diffusion.cpp enables fast and memory-efficient inference across multiple platforms, supporting machines with as little as 4GB VRAM. Other community tools include Cache-DiT for inference acceleration via DBCache, Context Parallelism and Tensor Parallelism; DiffSynth-Studio for LoRA training, full training, distillation training, and low-VRAM inference; and vllm-omni for omni-modality model fast inference and serving. GGUF quantized versions of Z-Image Turbo are available for low-VRAM deployment on consumer-grade GPUs with as low as 6GB VRAM. The official documentation details these community tools but does not establish a specific adult-content policy or built-in NSFW filtering mechanism.
Z-Image models offer various configuration options for different use cases. This section details the recommended parameters and settings for each model variant based on official documentation.
Z-Image-Turbo is a distilled version requiring only 8 NFEs (Number of Function Evaluations) for inference. The recommended configuration includes using bfloat16 precision for optimal performance on supported GPUs, with guidance scale set to 0.0 for Turbo models. Z-image nsfw supports resolutions from 512×512 to 2048×2048 with any aspect ratio. For optimal performance, users can enable Flash Attention backends including Flash-Attention-2 and Flash-Attention-3. Model compilation is available to accelerate inference, though the first run will take longer to compile. CPU offloading can be enabled for memory-constrained devices. The official documentation provides detailed code examples for PyTorch native inference and Diffusers inference, including specific parameter settings for prompt, height, width, num_inference_steps, guidance_scale, and generator seed.
The Z-Image foundation model offers more flexible configuration options compared to the Turbo variant. Recommended parameters include resolution from 512×512 to 2048×2048 with any aspect ratio, guidance scale between 3.0 and 5.0, and inference steps between 28 and 50. Negative prompts are strongly recommended for better control over generated content. CFG normalization can be set to False for general stylism or True for realism. Z-image nsfw supports both English and Chinese prompts, with the official documentation providing examples in both languages. For the Z-Image-Edit variant, the model supports creative image-to-image generation with impressive instruction-following capabilities, allowing for precise edits based on natural language prompts. The official documentation details these configuration options but does not establish a specific adult-content policy or built-in NSFW filtering mechanism.
This section provides practical notes for users and developers working with Z-Image models, including deployment considerations, community resources, and important caveats based on official documentation.
Z-Image-Turbo offers sub-second inference latency on enterprise-grade H800 GPUs and fits within 16GB VRAM consumer devices. For lower-end hardware, community tools like stable-diffusion.cpp enable deployment on machines with as little as 4GB VRAM, supporting multiple platforms including CUDA and Vulkan. GGUF quantized versions are available for consumer-grade GPUs with as low as 6GB VRAM, supporting multiple quantization levels including Q3_K_S, IQ4_NL, and IQ4_XS for different VRAM/quality trade-offs. The official documentation recommends installing diffusers from source for the latest features and Z-Image support, as two pull requests have been merged into the official diffusers release. Users should ensure their ComfyUI is updated to the latest version to access Z-Image workflow templates. The official documentation details these deployment considerations but does not establish a specific adult-content policy or built-in NSFW filtering mechanism.
The Z-Image ecosystem includes a rich collection of community resources. The official resource library provides access to open-source models, ControlNet and LoRA models, app demos, academic papers, and an official blog. Community works include Cache-DiT for inference acceleration, stable-diffusion.cpp for cross-platform deployment, LeMiCa for training-free timestep-level acceleration, and MeanCache for up to 3.7x speedup with plug-and-play integration. The ComfyUI community provides official workflow templates and GGUF quantized model collections. DiffSynth-Studio offers support for LoRA training, full training, distillation training, and low-VRAM inference. The official documentation details these community resources but does not establish a specific adult-content policy or built-in NSFW filtering mechanism. Users should consult current service terms for the platforms where Z-Image is hosted for content policy details.
Frequently asked questions about Z-Image model documentation and access.
z-image nsfw refers to the technical documentation and access information for the Z-Image model family developed by Alibaba's Tongyi Lab. The official repository and supporting resources detail z-image nsfw's architecture, variants, performance benchmarks, and deployment options, but do not establish a specific adult-content policy or built-in NSFW filtering mechanism. Users seeking content policy details should consult current service terms for the platforms where Z-Image is hosted, such as Hugging Face and ModelScope. The model family includes four variants: Z-Image-Turbo for fast generation, Z-Image for high-quality generation, Z-Image-Omni-Base for generation and editing, and Z-Image-Edit for image editing tasks.
Z-Image models are available through multiple channels. Model checkpoints have been released on Hugging Face and ModelScope platforms. The official GitHub repository contains complete model code and documentation. An online demo is available for direct experience of Z-Image generation capabilities, running on Zero GPU with no local configuration required. Community tools like ComfyUI offer workflow templates for Z-Image-Turbo, and stable-diffusion.cpp enables deployment on machines with as little as 4GB VRAM. Some variants like Z-Image-Omni-Base and Z-Image-Edit are marked as "To be released" on Hugging Face and ModelScope.
For Z-Image-Turbo, recommended configuration includes using bfloat16 precision, guidance scale set to 0.0, and resolutions from 512×512 to 2048×2048 with any aspect ratio. The model requires only 8 NFEs for inference. For the Z-Image foundation model, recommended parameters include guidance scale between 3.0 and 5.0, inference steps between 28 and 50, and strong recommendation to use negative prompts for better control. CFG normalization can be set to False for general stylism or True for realism. Users can enable Flash Attention backends and model compilation for improved performance.
