SANA is NVIDIA Labs' open-source framework for high-resolution image and video generation built around linear diffusion transformers. The project, hosted at NVlabs/Sana, crossed 6,500 GitHub stars by May 2026 and has become a reference implementation for efficient diffusion at high resolutions. The repository is released under the Apache-2.0 license and ships complete training and inference pipelines, which makes it usable for both research and production deployment. The core thesis of SANA is that the quadratic attention used in mainstream diffusion transformers like Flux and SD3 is the main bottleneck at 4K resolutions. By replacing it with linear attention and aggressive latent compression, the project reaches a published claim of "20 times smaller and 100 times faster than Flux-12B" for comparable image quality. That positioning has made SANA a frequently cited baseline for efficient generative models in 2026. ## Architecture SANA is built on three architectural ideas. The first is linear attention inside the diffusion transformer, which reduces complexity from O(N squared) to O(N) in the sequence length and allows the model to scale to 4K image resolutions without prohibitive memory cost. The second is DC-AE, a deep compression autoencoder that achieves 32-times image compression, so the latent grid the transformer operates on is much smaller than in conventional VAEs. The third is a decoder-only text encoder with in-context learning, which replaces the older CLIP-style encoders and improves long-prompt understanding. Sampling efficiency is addressed by the Flow-DPM-Solver, a custom solver tuned for the flow-matching formulation, and by sCM distillation that compresses many-step generation into a small number of inference steps. The combination is what enables SANA-Sprint, the one-step or few-step variant that produces a 1024-pixel image in roughly 0.1 seconds on an H100. ## Model Family The SANA family covers several use cases. The original SANA is a text-to-image model up to 4K. SANA-1.5 adds compute scaling for training and inference. SANA-Sprint is the few-step variant focused on interactive applications. SANA-Video and LongSANA extend the framework to video generation using block causal linear attention for long sequences. Sol-RL provides reinforcement learning training infrastructure built on the SANA base. SANA-WM is a 2.6-billion-parameter world model for video generation released in May 2026, which signals that the project is moving from pure image synthesis into agentic and simulation use cases. ## Deployment SANA is designed to be deployable on consumer GPUs. Four-bit quantization brings memory use down to roughly 8 GB of VRAM, which makes the model runnable on RTX 4060-class hardware for inference. The framework integrates with the diffusers library, ComfyUI, and SGLang, so existing pipelines can use SANA without a custom serving stack. A Replicate API endpoint is available for users who do not want to self-host. The repository includes full training code, which sets SANA apart from many model releases that ship inference-only weights. Researchers can fine-tune the model on custom datasets, reproduce ablations, or extend the linear attention design to other domains. Block causal linear attention in particular has been picked up by other video generation projects as a generic building block. ## Strengths The combination of linear attention, 32-times VAE compression, and modern text encoding is the main reason SANA stays competitive with much larger models. The Apache-2.0 license removes the legal friction that has slowed adoption of some other strong open models. Active maintenance, including the SANA-WM release in May 2026 and the Sol-RL update in April 2026, signals that the project is not a one-shot release. Discord support and continued documentation updates make it accessible to teams without internal generative modeling expertise. ## Limitations CUDA-capable hardware is effectively required for usable performance. The 4-bit quantization that enables 8 GB deployment trades some image quality for memory and speed, and applications that need maximum fidelity should expect to run the full-precision variant on stronger hardware. Documentation for the newest models, especially SANA-WM, is still catching up with the code. SANA's research focus means some of the deployment paths, particularly RL training and world modeling, are less polished than the base text-to-image pipeline. ## Outlook SANA's influence is visible in the way 2026's diffusion research has shifted toward linear and sub-quadratic attention. The framework's openness, combined with NVIDIA Labs' continued investment, makes it likely to remain a reference point for efficient generative models. For teams building image or video products, SANA is a practical choice today, and the SANA-WM and Sol-RL components hint at a broader role in agent and simulation systems as those areas mature.