ltx 2.3 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 LTX-Video model, developed by Lightricks, is a DiT-based video generation model that integrates multiple core capabilities. This section covers ltx 2.3 nsfw's architecture and key features.
LTX-Video is described as the first DiT-based video generation model that contains all core capabilities of modern video generation in one model. These include synchronized audio and video, high fidelity, multiple performance modes, production-ready outputs, API access, and open access. Ltx 2.3 nsfw can generate up to 50 FPS videos at native 4K resolution with synchronized audio in one pass. It is trained on a large-scale dataset of diverse videos and can generate high-resolution videos with realistic and diverse content. The model supports image-to-video, multi-keyframe conditioning, keyframe-based animation, video extension (both forward and backward), video-to-video transformations, and any combination of these features. The official repository provides examples of image-to-video and controlled video generation.
The repository documents several model versions and updates. The initial release (v0.9.0) in November 2024 supported text-to-video and image-to-video generation. Subsequent releases added features like improved quality, support for keyframes and video extension, higher resolutions, improved prompt understanding, and a new VAE. The v0.9.5 release introduced a new license for commercial use (OpenRail-M). The v0.9.6 release included a distilled model with 15x faster inference. The v0.9.7 release added new control models for depth, pose, and canny control. The v0.9.8 release supported up to 60 seconds of video. The repository also announced LTX-2, the next generation with synchronized audio+video generation, though its weights, code, and benchmarks were planned for release later in 2025.
This section details the technical capabilities of the LTX-Video model and the various ways users can access it for video generation tasks.
Ltx 2.3 nsfw supports a range of technical parameters for video generation. It works on resolutions divisible by 32 and frame counts divisible by 8 plus 1 (e.g., 257). The model works best on resolutions under 720 x 1280 and frame counts below 257. Recommended guidance scale values are 3-3.5. Inference steps can be adjusted: more steps (40+) for quality, fewer steps (20-30) for speed. The model supports stochastic inference, which can improve visual quality when using the distilled model. Prompt engineering guidelines recommend detailed, chronological descriptions of actions and scenes, starting with the main action and including specific movements, appearances, camera angles, and environmental details. Automatic prompt enhancement can be enabled by setting enhance_prompt=True when using LTXVideoPipeline directly.
Ltx 2.3 nsfw is accessible through multiple methods. Online inference is available via LTX-Studio, Fal.ai, and Replicate. For local installation, the codebase was tested with Python 3.10.5, CUDA version 12.2, and supports PyTorch >= 2.1.2. On macOS, MPS was tested with PyTorch 2.3.0. The repository provides inference scripts for image-to-video generation, video extension, and multi-condition video generation. ComfyUI integration is built into ComfyUI core for seamless workflows. Diffusers integration is also supported, including an 8-bit version of LTX-Video. The repository provides model user guides, parameter guides, and community contributions like ComfyUI-LTXTricks, LTX-VideoQ8, and TeaCache for LTX-Video.
This section outlines the different ways users can access the LTX-Video model, including online platforms and local installation options.
Ltx 2.3 nsfw is accessible right away via several online links. LTX-Studio offers image-to-video generation with both the 13B-mix and 13B distilled models. Fal.ai provides image-to-video generation with the 13B full and 13B distilled models. Replicate also offers image-to-video generation. These platforms allow users to generate videos without local installation. The repository also mentions an online web app in LTX-Studio with automatic prompt enhancement. For users who prefer local installation, the repository provides detailed instructions for setting up the environment and running inference scripts.
For local installation, users need to clone the repository and set up a Python environment. The codebase was tested with Python 3.10.5, CUDA version 12.2, and supports PyTorch >= 2.1.2. On macOS, MPS was tested with PyTorch 2.3.0. The repository provides inference scripts for various tasks: image-to-video generation, video extension, and multi-condition video generation. Users can also use the model as a library with the infer function. FP8 kernels are available for performance boost on supported graphics cards (Ada architecture and later). The repository recommends using ComfyUI workflows for best results and is working on updating the inference.py script to match ComfyUI's quality.
This section describes the different model variants available, including their characteristics and recommended use cases.
The repository lists several model variants with different characteristics. The ltxv-13b-0.9.8-dev offers the highest quality but requires more VRAM. The ltxv-13b-0.9.8-distilled is faster with less VRAM usage and slight quality reduction, ideal for rapid iterations. The ltxv-2b-0.9.8-distilled is a smaller model with slight quality reduction compared to the 13B distilled, ideal for fast generation with light VRAM usage. Quantized versions (FP8) are available for both the 13B dev and distilled models, as well as the 2B distilled model. The repository also provides a mix variant that combines the 13B dev and distilled models in a multi-scale rendering workflow for balanced speed and quality.
The repository released three control models for LTX-Video on HuggingFace: Depth Control, Pose Control, and Canny Control. These enable precise control over video generation. Community contributions include ComfyUI-LTXTricks, which provides additional nodes for enhanced control, including implementations of RF-Inversion, RF-Edit, FlowEdit, and more. LTX-VideoQ8 is an 8-bit optimized version for faster performance on NVIDIA ADA GPUs, offering up to 3X speed-up with no accuracy loss. TeaCache for LTX-Video is a training-free caching approach that can accelerate inference by up to 2x without significant visual quality degradation. The repository also provides training tools for fine-tuning, including LoRA training capabilities.
This section provides practical guidance for users working with the LTX-Video model, including prompt engineering tips and parameter recommendations.
The repository provides detailed prompt engineering guidelines for best results. Users should focus on detailed, chronological descriptions of actions and scenes. Include specific movements, appearances, camera angles, and environmental details in a single flowing paragraph. Start directly with the action and keep descriptions literal and precise. The recommended structure includes: start with main action in a single sentence, add specific details about movements and gestures, describe character/object appearances precisely, include background and environment details, specify camera angles and movements, describe lighting and colors, and note any changes or sudden events. Prompts should be kept within 200 words for best results.
The repository provides parameter recommendations for different use cases. Higher resolutions are recommended for detailed scenes, while lower resolutions are better for faster generation and simpler scenes. The model works best on resolutions under 720 x 1280 and frame counts below 257. Recommended guidance scale values are 3-3.5. For inference steps, more steps (40+) are recommended for quality, while fewer steps (20-30) are suitable for speed. Users can save seed values to recreate specific styles or compositions. The model supports stochastic inference, which can improve visual quality when using the distilled model. For advanced parameters, users can run python inference.py --help.
Common questions about the LTX-Video model and its capabilities.
The official LTX-Video repository documents the model's technical capabilities, access methods, and configurations. It does not establish an adult-content policy or built-in NSFW filtering. Users should refer to current service terms for content policies. The model supports various video generation tasks including image-to-video, video extension, and multi-keyframe conditioning.
The model is accessible through multiple methods. Online inference is available via LTX-Studio, Fal.ai, and Replicate. For local installation, users can clone the repository and set up a Python environment. ComfyUI integration is built into ComfyUI core for seamless workflows. Diffusers integration is also supported.
The codebase was tested with Python 3.10.5, CUDA version 12.2, and supports PyTorch >= 2.1.2. On macOS, MPS was tested with PyTorch 2.3.0. FP8 kernels are available for performance boost on supported graphics cards (Ada architecture and later). The repository provides detailed installation instructions.
