Every single week we get 20+ diffusion papers on arXiv. Most are incremental. Most will never run on production hardware.

This week was different. 9 drops landed that actually change what you can build, and how much it will cost you to run it. We will go through each one, with hard numbers, implementation status, and exactly when you should deploy each.

DiffusionGemma breaks the autoregressive monopoly

This is the biggest news this month, and almost no one is talking about the correct implication.

Google did not release a faster LLM. They proved that diffusion works better than autoregressive generation for text, for every workload that does not require streaming token output.

DiffusionGemma 26B is an MoE, activates 3.8B parameters per step. On RTX 5090 it hits 700 tokens per second. On H100 it crosses 1100 tps. That is 4.1x faster than Gemma 4 27B at the same perplexity.

It does this by generating all 256 output tokens in parallel over 12 denoising steps. There is no left to right bias. It solves Sudoku correctly 89% of the time, where the best autoregressive 70B model sits at 41%.

You will not use this for chat. Users want tokens one at a time. You will use this for every batch job, every summarization, every classification, every code generation run that you do offline. This cuts inference cost by 75% overnight for those workloads.

This is not a one off model. This is an architecture shift. Every major lab will have a diffusion text model out by Q4.

GF-DiT: 6x throughput for DiT serving without retraining

If you are running DiT models in production today, stop reading and go pull the vLLM-Omni branch. This is not an incremental improvement.

All existing DiT serving systems assign a fixed parallelism configuration for the entire lifetime of a request. That was always stupid. DiT workloads change drastically across denoising steps. Early steps are memory bound. Late steps are compute bound. Load changes every second.

GF-DiT treats parallelism as a schedulable resource. It can reallocate GPUs between running requests between every single denoising step. It does this with group-free collectives that drop reconfiguration overhead from 778ms down to 60 microseconds. That is four orders of magnitude improvement.

Benchmarks show 6.01x higher throughput, 95% lower average latency, 90% fewer SLO violations. No changes to the model. No retraining. No quantization. Just better scheduling.

This is the single largest serving efficiency gain for diffusion models we have ever seen published. It will become the default serving runtime by the end of July.

BudCache fixes the biggest problem with diffusion caching

Step level caching was supposed to cut inference cost in half. Everyone tried it. Everyone rolled it back.

The problem with all existing caching implementations is that they use per step error thresholds. You get wildly variable runtime per request. You cannot give customers a fixed latency guarantee. You cannot budget compute correctly. Half the time you run 20 steps, half the time you run 50. No one accepts that in production.

BudCache inverts this. You specify exactly how many steps you want to run. It then runs an offline simulated annealing search once, to find the optimal set of steps to cache that will produce the best possible output quality for that exact budget.

No online overhead. No thresholds. No variable latency. You get exactly the run time you asked for, and better quality than any heuristic caching approach at the same cost.

On FLUX.1-dev, BudCache delivers equivalent FID at 32 steps that heuristic caching delivers at 42 steps. That is a 24% cost reduction with zero downsides. Code is already public. You can deploy this this week.

Block verification brings speculative decoding to diffusion

Speculative decoding gave LLMs 2-3x speedup three years ago. Everyone has been trying to port it correctly to diffusion ever since. No one got it right until now.

The problem was always the residual distribution. Discrete spaces make this trivial. Continuous diffusion spaces do not. All prior implementations either broke the output distribution, or added so much overhead they cancelled out any speed gain.

This paper implements the correct original speculative sampling mechanism for continuous diffusion. Most importantly, it ports block verification from LLMs.

Block verification lets you accept or reject multiple steps at once. This delivers an additional 6.3% speedup over the best existing speculative diffusion implementations, with no extra training, no extra parameters, and zero distribution drift.

This is a pure sampling algorithm change. Drop it into any existing diffusion pipeline. It will just run faster.

Flex4DHuman: 4D humans from phone video, no priors

Everyone working on 4D reconstruction has been stuck on the same dead end for two years: you need skeleton fits, depth maps, normals, or calibrated cameras. None of those exist for casual phone video.

Flex4DHuman throws all of that out. It only conditions on relative camera pose. No explicit geometry priors at all.

It is fine tuned on Wan 2.1 1.3B. Feed it a 10 second monocular video shot on a phone. It will output synchronized dense multi view video. Feed that straight into an off the shelf 4D Gaussian Splatting pipeline. You get a full dynamic 4D asset that you can render from any angle.

It beats every prior state of the art on DNA-Rendering and ActorsHQ. It even works on dogs and horses if you add 10% animal data to the training set.

This is the first system that actually makes scalable user generated 4D content practical. This will be in every AR platform 12 months from now.

A2D2: Reward fine tuning for any length discrete diffusion

Discrete diffusion is the quiet workhorse no one talks about. It powers almost all modern molecular generation, circuit design, and structured output systems.

Until now there was no good way to do reward guided fine tuning on variable length discrete diffusion models. All existing methods only worked for fixed length sequences.

A2D2 solves this. It derives the correct Radon-Nikodym derivative for the joint insertion unmasking path. This gives guaranteed convergence to the reward tilted distribution, no required target samples.

This is not a minor detail. This removes the last major blocker for deploying diffusion models for production structured generation. If you are building anything that generates non text sequences, you will build it on this framework.

Dual constrained diffusion compression hits the full RDP frontier

Image compression has been stuck on the same tradeoff for ten years. You can have sharp realistic images that do not match the original, or accurate blurry images. There was no way to slide between them from the same bitstream.

DCIC fixes this. It adds two simple constraints to the diffusion decoder: distortion bound, and idempotence.

From a single compressed bitstream, you can adjust two scalar values at decode time. Slide continuously all the way from maximum fidelity archival mode, to maximum perceptual realism social media mode. No extra bits. No re-encoding.

On ImageNet-1K DCIC beats every existing perceptual codec on BD-PSNR, and matches dedicated realism codecs on BD-FID. This will replace every neural image codec used on the internet within three years.

Long tail sampling without retraining

Classifier guided diffusion produces very average samples. It converges hard on the mean of the distribution. It almost never generates rare, unusual, interesting examples.

All prior fixes for this required training an extra auxiliary network. No one ever did that in production.

This paper presents a one line change to the sampler. You apply guidance on the predicted image instead of the predicted noise. That is it. No retraining. No extra parameters. Just change where you multiply the guidance gradient.

On ImageNet 256x256 this improves recall by 18% while keeping FID almost identical. You get far more diverse outputs, for free, from every existing pretrained conditional diffusion model.

This is the most underrated paper this week. Almost no one has noticed it yet. Everyone will be running this sampler by the end of the month.

Molecular diffusion finally gets uncertainty estimates

You can generate 10,000 candidate molecules with a diffusion model. 9000 of them are garbage. Until now you had no way to tell which ones without running expensive DFT simulations.

This work adds a post hoc Laplace approximation to the denoising network. It measures noise prediction variance across the entire generation trajectory.

The resulting uncertainty score correlates -0.78 with actual binding affinity. You can filter out 70% of bad samples before you ever run a simulation. This cuts the cost of molecular discovery pipelines by 60% overnight.

This is not flashy. No one will write news articles about it. It will save pharmaceutical companies billions of dollars.

Deployment priority order

Prioritize in this order for this quarter:

1. GF-DiT. Deploy this first. 6x throughput for zero work.
2. BudCache. 25% cost reduction for all image generation.
3. The modified long tail sampler. Drop this into every pipeline.
4. Block verification speculative sampling. Another 6% free speedup.

Wait 3 months for:

• DiffusionGemma for batch workloads
• DCIC image compression

Keep an eye on, but do not deploy yet:

• Flex4DHuman
• A2D2
• Molecular uncertainty estimation

None of this is theoretical. All of this works today. All of this runs on hardware you already have.

Diffusion stopped being just for images a long time ago. This week it stopped being an experimental technology. It is now the best general purpose generative architecture for almost every workload.