Alibaba and Tsinghua Make Their Move: Diffusion Model Inference in Just a Few Steps, ICML Outstanding Paper Truly Deserved

The recently concluded ICML 2026 announced its list of Outstanding Papers, and the diffusion inference acceleration work jointly developed by Alibaba and Tsinghua University’s Zhu Jun team was prominently featured. This is not a paper obsessed with stacking SOTA numbers. Instead, it feels more like a counterstrike—using the simplest possible idea to push diffusion model inference speed records forward by another major leap.

For engineers who have worked on image/video generation deployment, this news is worth paying attention to. Diffusion models being slow is an old problem, but it is also becoming an increasingly fatal one. As Sora-like video models and successive generations of T2I models such as Flux continue scaling up parameter counts, inference cost has become the hardest obstacle to commercialization.

Where exactly are diffusion models “slow”?

First, some background. A standard diffusion model starts from pure Gaussian noise and gradually “denoises” it into a clear image, typically requiring 50 to 100 steps. Compared with generative models like GANs and VAEs that produce results in a single forward pass, diffusion models are inherently 50 to 100 times more expensive.

Over the past few years, academia has done extensive work around “how to reduce the number of steps,” with roughly three major directions:

• ODE/SDE solver optimization: Zhu Jun’s own DPM-Solver is a representative example, squeezing the structural information of ODEs to generate high-quality images within 10–15 steps
• Distillation approaches: Consistency Models and Progressive Distillation compress multi-step teachers into few-step students, even achieving one-step generation
• Training-time redesigns: Rectified Flow directly changes the training objective to straighten trajectories, naturally reducing the number of sampling steps

Each route has trade-offs. The ceiling of solver optimization is constrained by the mathematical structure itself—below 10 steps, image quality degrades noticeably. Distillation requires additional training, and student models often lose prompt responsiveness. Rectified Flow requires retraining from scratch, making existing models difficult to reuse.

What did this “minimalist solution” get right?

The cleverness of this Alibaba + Tsinghua work lies in cutting complexity down rather than piling it on. The paper introduces no new network branches, no extra alignment losses, and no multi-stage distillation pipeline. Its core idea can be summarized bluntly in one sentence: reexamine redundancy in the diffusion sampling process, merge what should be merged, skip what should be skipped, and trade minimal mathematical modifications for maximum acceleration.

Based on information disclosed by QbitAI, the key characteristics of this approach include:

1. Training-free or extremely low-cost fine-tuning: It can be plugged directly into existing pretrained models without retraining the base model. This is especially friendly to industry—you do not need to rebuild the SD/Flux/DiT model you spent millions training
2. Extreme compression of sampling steps: While keeping metrics like FID essentially intact, it reduces effective step counts into a range previously difficult for existing methods to reach
3. Architecture agnostic: Works for both U-Net-based and DiT-based systems

To be honest, when I first saw the phrase “minimalist,” I was skeptical. Diffusion models have already been optimized repeatedly by every possible group over the past few years, so there should not have been many low-hanging fruits left. But the fact that this work won an ICML Outstanding Paper award suggests the reviewers agreed on one thing: the method is clean, fundamental, and reproducible. In recent years, ICML Outstanding Papers have increasingly favored work that “makes complex problems simple,” and this paper is no exception.

Compared with DPM-Solver and Consistency Models, where is the advantage?

A side-by-side comparison makes things clearer.

| Method | Requires Training | Typical Steps | Generality | |------|-----------|---------|--------| | DDIM | No | 50 | High | | DPM-Solver++ | No | 10–20 | High | | Consistency Model | Distillation required | 1–4 | Medium | | LCM/LCM-LoRA | Distillation required | 4–8 | Medium | | Rectified Flow | Retraining required | 1–4 | Low | | This paper’s method | No/lightweight | Single-digit | High |

Pay attention to the last row—training-free + single-digit steps + high generality. That combination was previously almost nonexistent. Consistency Models achieve low step counts, but at the cost of distillation; DPM-Solver is training-free, but image quality starts dropping at single-digit steps. This new work effectively combines the sweet spots of both curves.

For teams building inference services, what does this mean? It means no need to maintain piles of LoRA or distilled weights, and no need to worry about acceleration schemes breaking when users switch base models. One sampler implementation, usable everywhere.

What can this bring from an engineering perspective?

Translating academic metrics into engineering gains, the rough accounting looks like this:

• Single-image T2I generation: Reducing from 50 steps to around 5 steps cuts latency on an A100 from 2–3 seconds down to roughly 300–500ms, fundamentally changing interactivity
• Video generation: Sampling-step bottlenecks are even more painful for video models. If the same acceleration ratio transfers, the commercialization threshold for Sora-like models drops significantly
• Edge deployment: Compressing step counts directly reduces memory-operator round trips, making it friendlier for mobile and edge devices
• Cost: For API-based service providers, inference cost can almost be divided by 10 directly

Interestingly, Zhu Jun’s team has consistently worked on diffusion model efficiency. From Analytic-DPM to DPM-Solver to U-ViT, nearly every step landed right before industry demand exploded. This collaboration with Alibaba is, in a sense, a fusion of academic insight and engineering scale—Alibaba has access to massive real-world inference workloads, which can in turn guide algorithm design toward the most cost-effective step-count regimes.

Deployment timeline and ecosystem impact

Based on past experience, once methods like this are open-sourced, the community absorbs them extremely quickly. Back when DPM-Solver was released, it took less than three months to go from paper publication to inclusion in the diffusers library and integration into Stable Diffusion WebUI. If this new method follows a similar open-source pace, it will likely appear in mainstream inference frameworks’ sampler lists within the year.

For developers, several points are worth thinking about in advance:

• Migration cost for existing pipelines: If this really is a sampler-layer replacement, diffusers users may only need to change one line of scheduler code to benefit
• Combination with TensorRT / ONNX compiler optimizations: Algorithm-level step compression and compiler-level operator fusion are orthogonal, and together they may deliver 20–30× end-to-end acceleration
• Impact on upper-layer products: “Real-time preview” in AI art products may become genuinely real-time instead of today’s halfway progressive refresh

Another point worth mentioning is the model distribution landscape. The more universal acceleration methods become, the weaker the moat around models themselves. If any open-source base model can be cheaply accelerated to under five steps, then “whose API is faster” rapidly stops being a meaningful technical barrier. Competition then shifts toward model quality and orchestration efficiency.

This is also why, recently, more developers have been moving toward multi-model aggregation approaches. Platforms like OpenAI Hub place GPT, Claude, Gemini, DeepSeek, Qwen, and many others behind a single OpenAI-compatible interface, accessible domestically with one key. This aligns perfectly with the trend: as underlying inference becomes faster and cheaper, freedom of model choice becomes the scarcer resource instead.

One conclusion

The “slowness” problem of diffusion models has been attacked repeatedly ever since DDPM was proposed in 2020. After every “major breakthrough,” people think we are near the limit, and then another piece of work pushes things further. The fact that this minimalist solution won an ICML Outstanding Paper award suggests that reviewers also recognize: this is not an incremental tweak, but a paradigm-level simplification.

But I do not think this is the endpoint. The real endpoint should be diffusion models matching—or even surpassing—GANs in inference cost, and matching autoregressive models in per-token cost. Only then will video generation truly have the material foundation for mass adoption.

Looking six months to a year ahead from this work, several directions are worth watching:

• Whether this minimalist idea naturally extends to video diffusion (where temporal redundancy may be even greater)
• Integration with newer paradigms like Flow Matching
• Adaptation within unified multimodal models (such as hybrid architectures like Transfusion and Show-o)

For teams constantly fighting inference costs every day, this paper is worth reading in full. Sometimes shortening the codebase by a few lines is more satisfying than doubling the number of GPUs.

References

• Tsinghua and Shanghai AI Lab Publish Survey on Diffusion Model Efficiency (Zhihu) - A systematic overview of diffusion model efficiency optimization techniques; referenced for the background section of this article
• NeurIPS 2025 Best Paper Awards Announced (BAAI Community) - A recent reference point for trends in top conference awards, used for comparison with ICML review tendencies