SANA
NVIDIASANA is NVIDIA's family of efficient diffusion models for image and video generation. The image models, SANA, SANA-1.5 and the few-step SANA-Sprint, generate images up to 4K and are small enough that the 0.6B model runs on a 16 GB laptop GPU. SANA-Video 2.0, released in August 2026, generates eight-second 720p clips from text or from text and an image.
SANA-WM, the family's camera-controlled world model, belongs with world models and is not part of this entry, and the repository's Sol engine branch, which accelerates other vendors' video models, is not scored here either. The code base moved to Apache 2.0 in January 2025. Every SANA model uses Google's Gemma 2 2B as its text encoder, and SANA-Video 2.0 decodes with the LTX-2.3 video autoencoder.
Openness
3 medium confidence- weights
- open (SANA-Video 2.0 and the SANA image checkpoints download from the Hub without a gate)
- data
- described (the SANA paper describes its caption labeling and selection; no training image or video set is released)
- code
- open (training and inference code for the image models and for SANA-Video 2.0 is published)
- license
- Apache-2.0 (the SANA weights and code)+Gemma-License (the Gemma 2 2B IT text encoder every SANA model requires)+LTX-2-Community-License (the LTX-2.3 video autoencoder SANA-Video 2.0 decodes with)
SANA's own code and weights are Apache 2.0 and its training code is published, but every model runs on Google's Gemma 2 text encoder, whose terms apply, and SANA-Video 2.0 also needs the LTX-2.3 autoencoder, which is free for commercial use only below USD 10M in annual revenue. The training data is described but not released.
- https://arxiv.org/abs/2410.10629 recorded 2026-11-12
SANA paper abstract: "efficient caption labeling and selection to accelerate convergence"; "Sana-0.6B can be deployed on a 16GB laptop GPU"; no training set is released.
- https://huggingface.co/api/models/Efficient-Large-Model/SANA-Video_2.0_5B_720p?expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=gated&expand[]=lastModified&expand[]=createdAt&expand[]=siblings recorded 2026-11-12
Hugging Face API record for SANA-Video_2.0_5B_720p: license apache-2.0, gated false, created 2026-08-20.
- https://huggingface.co/api/models/Efficient-Large-Model/SANA1.5_4.8B_1024px?expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=gated&expand[]=lastModified&expand[]=createdAt&expand[]=siblings recorded 2026-11-12
Hugging Face API record for SANA1.5_4.8B_1024px: license apache-2.0, gated false.
- https://huggingface.co/api/models/google/gemma-2-2b-it?expand[]=cardData&expand[]=gated recorded 2026-11-12
Hugging Face API record for google/gemma-2-2b-it: license gemma, gated "manual" behind Google's usage license.
- https://huggingface.co/Efficient-Large-Model/SANA-Video_2.0_5B_720p/raw/main/config.yaml recorded 2026-11-12
SANA-Video 2.0 config: vae_type LTX2VAE_diffusers loaded from LTX-2.3-Diffusers; text_encoder_name gemma-2-2b-it.
- https://huggingface.co/Efficient-Large-Model/SANA-Video_2.0_5B_720p/raw/main/LICENSE recorded 2026-11-12
SANA-Video 2.0 LICENSE body: Apache License, Version 2.0, "Copyright 2024 Nvidia".
- https://huggingface.co/Efficient-Large-Model/SANA-Video_2.0_5B_720p/raw/main/README.md recorded 2026-11-12
SANA-Video 2.0 card: "Text encoder: google/gemma-2-2b-it"; "VAE: LTX 2.3, 128 latent channels"; "License: Apache 2.0"; the checkpoint holds only the merged transformer.
- https://huggingface.co/Efficient-Large-Model/SANA1.5_4.8B_1024px/raw/main/README.md recorded 2026-11-12
SANA-1.5 card: "License: Apache License 2.0. Additional Information: Gemma Terms of Use ... for Gemma-2-2B-IT"; it "uses one fixed, pretrained text encoders (Gemma2-2B-IT)".
- https://ungh.cc/repos/Lightricks/LTX-2/files/main/LICENSE-2 recorded 2026-11-12
LTX-2 Community License: "Entities with annual revenues of at least $10,000,000" need a commercial license.
- https://ungh.cc/repos/NVlabs/Sana/files/main/LICENSE recorded 2026-11-12
LICENSE body: Apache License, Version 2.0.
- https://ungh.cc/repos/NVlabs/Sana/files/main/README.md recorded 2026-11-12
README news: "[2026/08] SANA-Video 2.0 training, inference, model architecture, and 5B 720p checkpoint are released"; "[2025/3/16] SANA-1.5 code & weights are released"; "[2025/1/11] Sana code-base license changed to Apache 2.0."
Adoption
2 high confidenceAdoption is measured as Hugging Face downloads summed across the NVIDIA SANA image and video checkpoints, most of them the two SANA-Video 2.0 repositories; the SANA-WM and SANA-Streaming repositories are left out.
- https://huggingface.co/api/models?author=Efficient-Large-Model&search=Sana&sort=downloads&direction=-1&limit=100 recorded 2026-11-12
Listing of 48 Efficient-Large-Model SANA repositories sorted by downloads: SANA-Video_2.0_5B_720p 7462, SANA-Video_2.0_5B_720p_4step 4142, Sana_Sprint_0.6B_1024px 693, Sana_600M_512px 543, the rest under 300 each; 14,618 in total, 14,464 without the SANA-WM and SANA-Streaming repositories.
Capability
1 medium confidenceNo SANA model is ranked on the public image or video arenas, while FLUX.2 [dev] sits in the upper half of the image board. SANA's emphasis is speed and size rather than top-end quality.
- https://artificialanalysis.ai/image/leaderboard/text-to-image recorded 2026-11-12
The overall text-to-image board as served lists 112 models and no SANA entry.
- https://artificialanalysis.ai/video/leaderboard/image-to-video recorded 2026-11-12
The image-to-video page serves lists of 35 and 40 models, neither with a SANA-Video entry.
- https://artificialanalysis.ai/video/leaderboard/text-to-video recorded 2026-11-12
The text-to-video page serves lists of 34 and 39 models, neither with a SANA-Video entry.
Verified 2026-11-12