SANA — Efficient High-Resolution Image and Video Generation family (NVlabs/Sana, sol-engine branch)

TL;DR

NVlabs/Sana is the umbrella repository for the SANA family — a series of efficiency-oriented diffusion models developed by NVIDIA Research and MIT HAN Lab. It currently packages seven distinct projects behind one codebase: SANA (T2I up to 4K), SANA-1.5 (inference-time scaling), SANA-Sprint (one/few-step distillation), SANA-Video / LongSANA (video with block linear attention + minute-length real-time), SANA-WM (world model), SANA-Streaming (real-time V2V editing), and Sol-RL (NVFP4 rollout + BF16 training RL). The sol-engine branch highlighted in the Slack post is the code companion to the Sol Video Inference Engine paper (Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation) — the agent-native acceleration framework applied to Cosmos3-Super, LTX-2.3, and SANA-Video.

Key claims

The repository packages seven distinct projects under one codebase, framed as “a fully open-source framework integrating efficient training, fast inference, and flexible deployment” [README §Introduction, §Getting Started].
SANA T2I is claimed 20× smaller and 100× faster than Flux-12B at up to 4K resolution [README §Introduction; ICLR 2025 Oral].
SANA-Sprint achieves one/few-step generation via sCM continuous-time consistency distillation, 0.1s per 1024px image on H100 and 0.3s on RTX 4090 [README §Introduction; ICCV 2025 Highlight].
SANA-Video (2B) achieves VBench Total 84.05 (Quality 84.63, Semantic 81.73) at 36s latency, vs Wan-2.1-14B at 83.73 / 1897s and Wan-2.1-1.3B at 83.38 / 400s — better VBench and >10× faster [README §Performance, Video Generation table; ICLR 2026 Oral].
SANA-WM is a 2.6B controllable world model producing 720p 1-minute video with 6-DoF camera control [README §Introduction; already filed as SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer].
SANA-Streaming is a 2B real-time streaming V2V editing model on Wan2.2 5B for 720p minute-scale video [README §Introduction; arXiv:2605.30409].
Sol-RL is a diffusion RL post-training recipe using NVFP4 rollouts + BF16 optimization, claiming 4.64× faster convergence, with training recipes for SANA / FLUX.1 / SD3.5-L [README §Introduction, News 2026/04; arXiv:2604.06916].
Model deployment surface: HuggingFace collection, ComfyUI nodes, SGLang serving, NVIDIA Cosmos-RL integration for RL post-training, MLX support via MFLUX for Apple Silicon [README §Getting Started, News].
4-bit quantization via SVDQuant/Nunchaku puts a 4K-capable SANA under 8GB VRAM [README §News 2025/01/24].
The sol-engine branch is specifically the SANA repository state that accompanies the Sol Video Inference Engine paper — i.e. the acceleration framework code for the three-model demonstration (Cosmos3-Super, LTX-2.3, SANA-Video) [tree/sol-engine URL, paper Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation §5].
Licensing: Apache 2.0 from 2025-01-11 onward [README §News 2025/1/11].
Companion repository mit-han-lab/dc-ae publishes the DC-AE (Deep Compression Autoencoder) that gives SANA its 32× spatial compression (vs traditional 8× LDM VAEs) — the tokenizer choice that makes 4K feasible on consumer hardware [README §Introduction §Key Techniques].

Method

The repository consolidates NVIDIA + MIT HAN Lab’s SANA line into one codebase spanning image, video, and world-model generation. The architectural through-line is linear attention replacing softmax attention in the DiT backbone (later formalized as Block Causal Linear Attention + Causal Mix-FFN for video), combined with a high-compression tokenizer (DC-AE, 32× compression) and a decoder-only LLM text encoder (Gemma) with in-context learning. Efficiency levers layered on top: sCM distillation (Sprint), Flow-DPM-Solver sampling, Sol-RL low-precision rollouts + high-precision optimization, and multi-scale training with FSDP + TAR-file WebDataset. The sol-engine branch adds the agent-native acceleration framework from Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation on top of the existing SANA-Video deployment.

Results

SANA-0.6B: 39.5× speedup vs FLUX-dev, FID 5.61, GenEval 0.68, DPG 84.2 at 1024² [README Image Generation table].
SANA-1.6B: 23.3× speedup, GenEval 0.69, DPG 84.5 [README Image Generation table].
SANA-1.5 4.8B: GenEval 0.82 (best in table), DPG 84.7, CLIP 29.23 — beats FLUX-dev on GenEval by 0.15 [README Image Generation table].
SANA-Video-2B: VBench Total 84.05 vs Wan-2.1-14B 83.73, at 36s vs 1897s (52.7× wall-clock) [README Video Generation table].
Sol-RL: 4.64× faster convergence claimed on diffusion RL post-training [README §Introduction].

Why it’s interesting

NVlabs/Sana is one of the most contentful multi-project umbrella releases in Open foundation-model releases — same code, same authorship line, seven distinct model releases spanning image → video → world → streaming editing → RL post-training. It fits the “single-model series with deep variants” pattern that the concept page tracks for LTX-2 and Qwen3-VL, but pushes it further: instead of variants of one model, it’s an architectural family (all linear-attention DiT + DC-AE) with each release exploring a different task. The sol-engine branch specifically closes the loop with Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation — NVLabs is now publishing not just the model family but also an agent-native acceleration framework that is evaluated on that model family, tying the Diffusion serving optimization frontier to the Open foundation-model releases frontier in one repo. Contrasts sharply with Genesis World — Simulation platform for general-purpose robotics & embodied AI learning which does the same “one repo, many artifacts” pattern for the physics simulation stack rather than the generative model stack.