efficient high-resolution image/video generation via diffusion transformers
NVlabs/Sana is an efficiency-oriented PyTorch codebase for high-resolution text-to-image and text-to-video generation using linear diffusion transformers, with training and inference pipelines for Sana, Sana-1.5, Sana-Video, Sana-Streaming, and related variants.
Solves the same problem with a different architecture or at a different layer. Expect to rewrite the integration.
Choose when a bitwise autoregressive high-resolution image-generation approach is acceptable and you do not need Sana's video/RL variants.
[CVPR 2025 Oral]Infinity ∞ : Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis
Choose when you want one unified multimodal model for understanding, text-to-image generation, and editing rather than Sana's dedicated diffusion stack.
An unified model that seamlessly integrates multimodal understanding, text-to-image generation, and image editing within a single powerful framework.
Choose when you need LTX-Video's real-time text-to-video model family and weights instead of Sana-Video.
Official repository for LTX-Video
Choose when real-time long-video generation is the goal and Helios's model/codebase fits better than Sana-Video.
Helios: Real Real-Time Long Video Generation Model
Choose when you want Kandinsky 5's MIT-licensed image/video diffusion model family and weights rather than Sana's architecture.
Kandinsky 5.0: A family of diffusion models for Video & Image generation
Choose when you want VideoCrafter2 text/image-to-video weights and training instead of Sana-Video.
VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models
Choose when you want causal-forcing distillation for real-time interactive video generation rather than Sana's streaming variant.
[ICML 2026] Official codebase for "Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation" & Causal Forcing++
Choose when you need CogVideoX/CogVideo text/image-to-video model weights; Apache-2.0 but it does not cover Sana's image-generation side.
text and image to video generation: CogVideoX (2024) and CogVideo (ICLR 2023)
An earlier or more limited way to do the job. Still the right call when you need something small, proven, or CPU-only.
Choose when you want LCM-distilled few-step Stable Diffusion inference rather than Sana's linear-DiT codebase; older and tied to SD.
Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference
Choose when you need the legacy Stable Diffusion v1 latent-diffusion pipeline/weights; lower resolution and no Sana efficiency or video support.
A latent text-to-image diffusion model
Choose when you want the original LDM research/training codebase rather than a modern text-to-image/text-to-video system.
High-Resolution Image Synthesis with Latent Diffusion Models
Choose when you want to train/experiment with an unCLIP/DALL-E-2-style architecture instead of Sana; research code with no maintained pretrained weights.
Implementation of DALL-E 2, OpenAI's updated text-to-image synthesis neural network, in Pytorch
Choose when you already have Stable Diffusion/SDXL weights and only need training-free high-resolution extension, not a full Sana-class generator.
Let us democratise high-resolution generation! (CVPR 2024)
Choose when you want the older VQGAN+Transformer image-synthesis research codebase; no modern text-to-image/video pipeline.
Taming Transformers for High-Resolution Image Synthesis
Choose when you need one-shot video personalization from a single example video built on image diffusion; old and narrow.
[ICCV 2023] Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation
Choose when your text-to-video task requires identity-preserving generation of a specific face, not general-purpose Sana-Video.
[CVPR 2025 Highlight🔥] Identity-Preserving Text-to-Video Generation by Frequency Decomposition
Choose when you want to distill an existing conditional diffusion model to few steps; it is an acceleration layer, not an end-to-end high-resolution system.
Flash Diffusion — accelerating conditional diffusion models (AAAI 2025 Oral)
Choose when you want the earlier PixArt-α diffusion-transformer codebase and checkpoints rather than Sana; no video and not actively updated.
PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis
Choose when you want a training-free zero-shot video generator built on pretrained image-diffusion models; old, proof-of-concept, and not Sana-Video quality.
[ICCV 2023 Oral] Text-to-Image Diffusion Models are Zero-Shot Video Generators
These projects were analysed and named Sana among their alternatives. The relationship is not symmetric — how Sana rates them is a separate judgement, made when Sana is analysed in its own right.
calls Sana “efficient high-resolution diffusion transformer with current checkpoints and faster sampling than the original LDM pipeline”
High-Resolution Image Synthesis with Latent Diffusion Models
calls Sana “Efficient high-resolution linear-diffusion transformer; choose for modern architecture and Apache-2.0 licensing.”
Implementation of DALL-E 2, OpenAI's updated text-to-image synthesis neural network, in Pytorch
calls Sana “Choose for text-to-image high-resolution synthesis with a linear diffusion transformer instead of VQGAN+AR.”
Taming Transformers for High-Resolution Image Synthesis