strong-followup
SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation
Junsong Chen, Shuchen Xue, Yuyang Zhao, Han Cai, Song Han · NVIDIA, MIT · 2025-03 · arXiv:2503.09641
Why it matters
Combines continuous-time consistency distillation with an adversarial term to reach 1-4 step text-to-image generation, reported faster than FLUX-schnell at comparable quality. Distillation-based, so it depends on a strong multi-step teacher — the contrast case against MeanFlow and Drifting, which train few-step behavior from scratch.
What this paper does
read: full textBefore this
Before this, few-step text-to-image distillation split into two camps. Consistency distillation such as LCM kept training stable but eroded quality below four steps, especially semantic alignment to the prompt. Adversarial approaches such as ADD and LADD reached strong single-step fidelity but suffered oscillatory GAN dynamics and needed careful re-tuning for every new backbone, while variational score distillation methods required training a second diffusion model alongside the student, adding memory and compute overhead.
The problem
The paper asks whether a single distillation recipe can combine the training stability of consistency distillation with the sample fidelity of adversarial distillation, and whether one model can serve 1 to 4 inference steps without training a separate model per step count.
The idea
Reformulate a pretrained flow-matching teacher into the TrigFlow parameterization so continuous-time consistency distillation (sCM) applies without training from scratch, then add a latent adversarial loss (LADD) on top. sCM anchors the student to the teacher's trajectory and keeps training stable, while LADD sharpens single-step sample fidelity where sCM alone is weakest.
How it works
SANA-Sprint operates in the VAE latent space inherited from the SANA teacher, not pixels. Training regresses the TrigFlow-parameterized consistency function, which predicts the clean latent x0 along the probability-flow ODE, using the sCM loss weighted by e^{w(t)/D} between the student's own output at two nearby noise levels. The teacher is a pruned and fine-tuned version of SANA-1.5 (4.8B); before distillation it is briefly fine-tuned for 5,000 iterations to add QK-normalization and a dense timestep embedding c_noise(t) = t, then frozen and used only to supply the velocity field and, through its frozen features, the discriminator for the LADD term. Guidance is distilled directly into the student via a CFG embedding rather than run as a separate classifier-free-guidance forward pass at inference. Student sizes of 0.6B and 1.6B parameters are trained with total loss sCM plus 0.5 times LADD, with 50% of training timesteps forced to t = pi/2 (max-time weighting) to stabilize the hardest single-step regime. A single step-adaptive model is trained once and sampled at 1, 2 or 4 steps at inference, rather than training a separate model per step count.
Evidence
On MJHQ-30K at 1 inference step, with guidance distilled into the model rather than run as test-time CFG, SANA-Sprint-1.6B reaches FID 7.59 and GenEval 0.74 in 0.1s on H100, against 1-step FLUX-schnell's FID 7.26 / GenEval 0.69 at 0.68s and 4-step FLUX-schnell's FID 7.94 / GenEval 0.71 at 2.10s. At 4 steps SANA-Sprint-1.6B improves to FID 6.66 and GenEval 0.77 in 0.31s. An ablation at matched settings isolates the two loss terms, sCM alone gives FID 8.93, LADD alone gives FID 12.20 and is unstable, and the combined loss gives FID 8.11, showing the hybrid corrects a real weakness of each term rather than just averaging them.
Limitations
The authors do not include a dedicated limitations section, but several constraints are conceded in the method description. The recipe depends on a strong pretrained flow-matching teacher (SANA-1.5 4.8B) rather than learning few-step behavior from scratch, so the student's quality ceiling is set by the teacher. Training used 32 A100 GPUs across 4 DGX nodes plus a separate 5,000-iteration teacher fine-tuning stage before distillation could begin. The authors also note Flash Attention's JVP kernel is unavailable in PyTorch, so they fall back to linear attention specifically to keep the sCM Jacobian-vector product tractable.
Why it matters
It shows a single step-adaptive model, distilled once, can match or beat a leading few-step baseline across 1, 2 and 4 steps without per-step retraining, and that hybridizing consistency distillation with an adversarial term fixes a specific quality gap each leaves open alone. Because it depends entirely on a strong pretrained teacher, it marks the distillation pole of the one-step landscape, against which training-from-scratch approaches like MeanFlow and Drifting Models are the contrasting case.
Abstract, in the authors' own words
This paper presents SANA-Sprint, an efficient diffusion model for ultra-fast text-to-image (T2I) generation. SANA-Sprint is built on a pre-trained foundation model and augmented with hybrid distillation, dramatically reducing inference steps from 20 to 1-4. We introduce three key innovations: (1) We propose a training-free approach that transforms a pre-trained flow-matching model for continuous-time consistency distillation (sCM), eliminating costly training from scratch and achieving high training efficiency. Our hybrid distillation strategy combines sCM with latent adversarial distillation (LADD): sCM ensures alignment with the teacher model, while LADD enhances single-step generation fidelity. (2) SANA-Sprint is a unified step-adaptive model that achieves high-quality generation in 1-4 steps, eliminating step-specific training and improving efficiency. (3) We integrate ControlNet with SANA-Sprint for real-time interactive image generation, enabling instant visual feedback for user interaction. SANA-Sprint establishes a new Pareto frontier in speed-quality tradeoffs, achieving state-of-the-art performance with 7.59 FID and 0.74 GenEval in only 1 step - outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10x faster (0.1s vs 1.1s on H100). It also achieves 0.1s (T2I) and 0.25s (ControlNet) latency for 1024 x 1024 images on H100, and 0.31s (T2I) on an RTX 4090, showcasing its exceptional efficiency and potential for AI-powered consumer applications (AIPC). Code and pre-trained models will be open-sourced.
Research line
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a strong-followup entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.