# Sana | NVIDIA Research Projects | Knowledge Base | Instagit

SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformer

GitHub Stars: 6.8k

Repository: https://github.com/NVlabs/Sana

---

## Articles

### [Implementing Video-to-Video Editing with Sana Video Refiner: A Complete Guide](/NVlabs/Sana/implementing-video-to-video-editing-sana-video-refiner)

Master video-to-video editing with Sana Video Refiner. Learn how to achieve high-fidelity results using diffusion and upsampling with this comprehensive guide.

- Tags: how-to-guide
- Published: 2026-05-19

### [DC-AE Tiling for 4K Resolution Inference with Limited GPU Memory: A Complete Guide](/NVlabs/Sana/dc-ae-tiling-4k-inference-limited-gpu-memory)

Learn DC-AE tiling to perform 4K resolution inference on GPUs with only 8GB VRAM. This guide explains how tiling drastically reduces peak GPU memory usage.

- Tags: how-to-guide
- Published: 2026-05-19

### [Implementing PAG (Progressive Attention Guidance) in Sana Inference](/NVlabs/Sana/implementing-pag-sana-inference)

Implement Progressive Attention Guidance PAG in Sana inference to boost fine detail fidelity. Learn how this plug-in enhances Sana diffusion models without losing global coherence.

- Tags: how-to-guide
- Published: 2026-05-19

### [How to Configure Mixed Precision Training (BF16/FP16) in Sana](/NVlabs/Sana/configuring-mixed-precision-training-sana)

Learn to configure mixed precision training (bf16/fp16) in NVlabs/Sana by updating your YAML config. Accelerate your training with automatic mixed precision (AMP).

- Tags: how-to-guide
- Published: 2026-05-19

### [How to Use the Diffusers Integration with SanaPipeline](/NVlabs/Sana/using-diffusers-integration-sanapipeline)

Learn how to use the diffusers integration with SanaPipeline for seamless loading and inference of native checkpoints. Effortlessly use the familiar pipe API.

- Tags: how-to-guide
- Published: 2026-05-19

### [Debugging Out of Memory (OOM) Issues During High-Resolution Image Training in Sana](/NVlabs/Sana/debugging-oom-high-resolution-image-training-sana)

Learn to debug Out of Memory OOM issues during high-resolution image training in Sana. Reduce GPU memory by 50% with mixed precision, gradient checkpointing, component offloading, and xformers.

- Tags: how-to-guide
- Published: 2026-05-19

### [Implementing SANA-WM for Controllable World Modeling with 6-DoF Camera Control](/NVlabs/Sana/implementing-sana-wm-controllable-world-modeling-6dof)

Implement SANA-WM for controllable world modeling and generate 720p videos with precise 6-DoF camera control. Discover its advanced Linear-attention DiT architecture.

- Tags: tutorial
- Published: 2026-05-19

### [Achieving 27FPS for Minute-Length Video Generation with LongSANA: A Complete Technical Guide](/NVlabs/Sana/achieving-27fps-minute-length-video-longsana)

Unlock 27FPS minute-length video generation with LongSANA. Explore our technical guide on chunk-causal inference, KV-cache reuse, streaming training, and optimized samplers.

- Tags: technical-guide
- Published: 2026-05-19

### [Configuring Multi-Scale WebDataset Training with TAR Files in NVIDIA Sana](/NVlabs/Sana/configuring-multi-scale-webtaset-tar-files)

Train images at multiple scales using TAR files in NVIDIA Sana. Learn to configure heterogeneous resolutions, sample captions by ClipScore, and dynamically select aspect ratios with wids-meta.json.

- Tags: how-to-guide
- Published: 2026-05-19

### [Using Cosmos-RL for Post-Training SANA-Image and SANA-Video: A Complete Guide](/NVlabs/Sana/using-cosmos-rl-post-training-sana-image-video)

Master post-training SANA-Image and SANA-Video with NVlabs Cosmos-RL. Learn efficient supervised fine-tuning and reinforcement learning using PEFT LoRA adapters and asynchronous reward scoring.

- Tags: how-to-guide
- Published: 2026-05-19

### [How Sol-RL Achieves 4.64× Faster Convergence with NVFP4 Rollout and BF16 Training](/NVlabs/Sana/how-sol-rl-achieves-faster-convergence-nvfp4-bf16)

Discover how Sol-RL achieves 4.64x faster convergence. Learn about NVFP4 rollout and BF16 training, which eliminate precision-stability trade-offs for efficient diffusion RL.

- Tags: deep-dive
- Published: 2026-05-19

### [High-Performance Serving of Sana with SGLang and OpenAI-Compatible API](/NVlabs/Sana/high-performance-serving-sana-sglang-openai-api)

Serve NVlabs Sana diffusion models at scale with SGLang. Enjoy unified runtime, command-line, Python SDK, and OpenAI-compatible API for high performance.

- Tags: performance
- Published: 2026-05-19

### [Deploying Sana with ComfyUI for Visual Workflow Editing](/NVlabs/Sana/deploying-sana-comfyui-visual-workflow-editing)

Deploy Sana with ComfyUI for visual workflow editing. Integrate custom nodes for seamless high-resolution diffusion pipeline creation and composition. Explore powerful generative AI.

- Tags: how-to-guide
- Published: 2026-05-19

### [Implementing LoRA Fine-Tuning with DreamBooth for Sana](/NVlabs/Sana/implementing-lora-finetuning-dreambooth-sana)

Learn to implement LoRA fine-tuning with DreamBooth for NVlabs Sana. Create custom adapters efficiently on consumer GPUs without altering base model parameters.

- Tags: how-to-guide
- Published: 2026-05-19

### [Using ControlNet with Sana for Controllable Image Generation: A Complete Implementation Guide](/NVlabs/Sana/using-controlnet-sana-controllable-image-generation)

Unlock controllable image generation with ControlNet and Sana. This guide shows how to use SanaControlNetPipeline to inject VAE-encoded control signals for precise diffusion sampling. Integrate spatial conditioning easily.

- Tags: how-to-guide
- Published: 2026-05-19

### [How to Train Sana with FSDP: A Complete Guide to Fully Sharded Data Parallel](/NVlabs/Sana/training-sana-fsdp)

Learn how to train Sana with FSDP to fit 2B parameter video models on two GPUs. This guide covers sharding parameters, gradients, and optimizer states.

- Tags: how-to-guide
- Published: 2026-05-19

### [Running Sana with 4-bit Quantization using SVDQuant and Nunchaku](/NVlabs/Sana/running-sana-4-bit-quantization-svdquant-nunchaku)

Run Sana with 4-bit quantization using SVDQuant and Nunchaku. Slash GPU memory to 8GB while preserving quality. Easily integrate NunchakuSanaTransformer2DModel for efficient inference.

- Tags: how-to-guide
- Published: 2026-05-19

### [Configuring Inference-Time Scaling in SANA-1.5 for Improved Image Quality](/NVlabs/Sana/configuring-inference-time-scaling-sana-1.5-improved-quality)

Enhance SANA-1.5 image quality using inference-time scaling. Generate numerous candidates and select top-K via tournament for superior GenEval scores without retraining.

- Tags: performance
- Published: 2026-05-19

### [How sCM Distillation Enables One-Step Generation in SANA-Sprint](/NVlabs/Sana/how-scm-distillation-enables-one-step-generation-sana-sprint)

Discover how sCM distillation in SANA-Sprint enables one-step image generation. Train diffusion models to jump from noise to image instantly, bypassing iterative refinement. Learn more about this breakthrough.

- Tags: deep-dive
- Published: 2026-05-19

### [Implementing Block Causal Linear Attention for Long Video Generation with Sana](/NVlabs/Sana/implementing-block-causal-linear-attention-long-video-generation)

Implement Block Causal Linear Attention for long video generation with Sana. This technique reduces complexity to linear, enabling generation of thousands of frames.

- Tags: how-to-guide
- Published: 2026-05-19

### [How a Decoder-Only Text Encoder Enhances Text-Image Alignment in Sana](/NVlabs/Sana/how-decoder-only-text-encoder-enhances-text-image-alignment-sana)

Discover how Sana uses a decoder-only text encoder to improve text-image alignment, extracting richer embeddings for optimized generation.

- Tags: deep-dive
- Published: 2026-05-19

### [What Is DC‑AE and How Does It Enable 32× Image Compression in Sana?](/NVlabs/Sana/what-is-dc-ae-32x-image-compression-sana)

Discover DC-AE, a deep compression auto-encoder that achieves 32x image compression in NVlabs Sana. Learn how this technique slashes memory usage and speeds up inference.

- Tags: deep-dive
- Published: 2026-05-19

### [How Linear Attention Improves Efficiency in DiT for High‑Resolution Images](/NVlabs/Sana/how-linear-attention-improves-efficiency-dit-high-resolution)

Discover how Linear Attention slashes DiT complexity from O(N²) to O(N), enabling 4K image generation on consumer GPUs by overcoming standard self-attention's memory bottleneck.

- Tags: deep-dive
- Published: 2026-05-19