A joint paper from Stanford and Stability AI hit arXiv yesterday — "Single-Step Rectified Flow with Latent Adversarial Distillation" — describing a technique that generates 1024x1024 images in a single diffusion step, taking 90ms on an H100. Conventional diffusion needs 25-50 steps and 600-1200ms.
The 13x speedup is the result of two combined techniques: rectified flow (already in FLUX and SD3) plus a new latent adversarial distillation that preserves quality while collapsing the multi-step process. Code is open-source on GitHub.
What single-step diffusion enables
The implications go beyond "diffusion is faster":
- **Real-time AI video**: 30 fps generation becomes feasible on consumer hardware (a single 5090 Ti could generate 720p at 30 fps using this technique)
- **Interactive image generation**: the diffusion model becomes a primary input device — you adjust prompts and see results in less than a frame
- **Mobile-tier inference**: the technique compresses to 8-bit quantization without quality loss, enabling on-phone diffusion at usable quality
- **Cost reduction**: API providers can drop image generation prices by 10x because compute is 13x cheaper
Stable Diffusion XL Lightning (a similar technique from 2024) hit 4 steps. SDXL Turbo hit 1 step but with significant quality degradation. The new paper hits 1 step with quality matching multi-step generation — that's the breakthrough.
How it works (without the math)
The technique trains a smaller "student" model to mimic the output of a larger "teacher" diffusion model — but with two innovations:
- **Adversarial loss in latent space**: the student is judged not on pixel similarity but on perceptual similarity in a learned latent space, preventing the blurry output common in distillation
- **Reflow training**: the model is iteratively retrained on its own outputs in a way that straightens the diffusion path, making single-step generation natural rather than forced
Combined, you get a model that generates the final image directly from noise in a single forward pass — no iterative denoising required.
Where this hits production first
Three concrete deployments expected in the next 8 weeks:
- **FLUX Pro 2 Turbo** — Black Forest Labs already announced single-step variant in their Q1 2026 roadmap
- **Stability AI SD-Turbo 2** — direct application of the paper, expected mid-May
- **Adobe Firefly real-time mode** — Adobe confirmed working on single-step diffusion for Premiere integration
For builders: if you maintain a diffusion pipeline, the techniques in this paper will be production-ready within 60 days. Switching cost is minimal — same model architecture, retrained weights.
Why this matters strategically
For two years, diffusion latency has been the constraint blocking AI image and video from competing with traditional rendering. Real-time image generation has been "almost feasible" since SDXL Turbo. Single-step generation at full quality finishes that arc.
The next frontier: multi-modal generation in real time. Generate a 4K image, animate it to 16-second video, add voice acting and score, all in under one minute on consumer hardware. We're 12-18 months from that being routine.
Sources
- Stanford CS / Stability AI (April 27, 2026): Single-Step Rectified Flow paper on arXiv
- GitHub repo with reference implementation (April 27, 2026)
- HuggingFace blog (April 28, 2026): What single-step diffusion changes