Three papers landed on arXiv this month that attack the cost of diffusion modeling from angles that barely overlap. One fixes where semantics flow inside a transformer. One cuts the price of converting a video diffusion model into something you can actually run in real time. One throws out the Gaussian assumption entirely and rebuilds noising around counting processes.

What connects them is a shared move: stop treating the diffusion stack as one monolithic thing, split it, and give each piece the treatment it needs.

Representation conditioning: the semantics leak out of the later layers

REPA established a trick worth stealing. Take a pretrained vision encoder, supervise an intermediate layer of your diffusion transformer through an MLP projection head, then discard the head once training finishes. The alignment loss back-propagates into the early layers and imports semantics from the encoder into the diffusion backbone.

Here's the gap. That loss only reaches the early layers. Later layers see nothing but the velocity prediction objective, and nothing forces them to use the semantics that got imported upstream. Representation conditioning, or RECON, closes the gap with a direct move. During training, feed the pretrained vision encoder's tokens into the later layers as well. At inference the clean image is gone, so you reuse the MLP head as a stand-in for the encoder, projecting its output tokens into those same later layers.

The method adds no parameters and converges 63x faster than SiT and 25x faster than REPA. Those numbers describe training convergence, not sampling speed.

If you train diffusion transformers from scratch, an order-of-magnitude cut to convergence is an order-of-magnitude cut to the GPU bill. Fine-tuning an already-trained model gets less out of it, because the later layers have settled into whatever representations they have.

Quick Take: The biggest wins in diffusion right now come from rerouting where information flows, not from scaling up parameter counts.

Glance forcing: distilling video diffusion for one GPU-hour

Autoregressive generation is what makes real-time, causal video synthesis possible, and it's the backbone of video world models. The standard path to it distills a bidirectional diffusion model into a causal few-step student. The problem is the pipeline: the conversion requires offline probability flow ODE trajectories that cost tens of thousands of GPU-hours and terabytes of storage. Most teams can't pay that.

Glance Forcing reduces trajectory supervision to a single sample, the strictest setting possible, then deals with what breaks. A single monolithic student collapses in quality when it has to learn from one trajectory. The authors trace the failure to the teacher's trajectory geometry. High noise governs semantic layout and motion. Low noise governs fine texture. Those two regimes have different curvature and different optimization behavior, so forcing one network to fit both causes the representations to interfere.

The fix splits the denoising trajectory into two complementary experts, one slow and one fast, each trained with a stage-specific objective. The teacher stays frozen. Using one PF-ODE trajectory, Glance Forcing pushes the VBench Total score from 84.04 to 84.22 and cuts trajectory-collection cost from thousands of GPU-hours to about one A100-GPU-hour.

84.04 to 84.22 looks like noise on a leaderboard. The cost column is where the paper earns its keep, and the aggregate hides the interesting part. The authors report larger gains in dynamic degree, VisionReward, and instruction-following, with human raters preferring it over strong baselines.

A claim in here generalizes past this method. PF-ODE distillation mainly straightens the geometry of the few-step trajectory. The distribution-matching stage that follows is what governs distributional quality. Two separate jobs, usually treated as one.

Mixed Poisson diffusion: when your data is counts

Gaussian noising assumes your data lives in a continuous space. A lot of interesting data doesn't. Poisson inverse problems show up in PET imaging, photon counting, and low-light sensing, where observations are discrete counts. Most posterior sampling methods for inverse problems assume a deterministic forward operator with additive Gaussian noise, which is the wrong model for counting data.

Mixed Poisson Diffusion Models replace the noising dynamics with multivariate counting processes. MPDM introduces the framework, builds posterior sampling algorithms on top of it for Poisson inverse problems, and evaluates on realistic cases. I wanted a benchmark table here and the paper doesn't hand me one in the material I have. What it gives instead is a generative framework that matches the observation model rather than forcing a Gaussian approximation onto discrete data. If you work in one of those domains, that mismatch has been costing you at every denoising step.

The part that took me two passes to get

The RECON inference trick took me two passes to get. During training you have the clean image, so you can run the vision encoder and feed its tokens into the later layers. At inference there's no clean image, so it looks like you've lost the signal. You haven't. The MLP projection head that REPA discards after training is still a working map from latent space to encoder-token space. Keep it, run it on the current latent, and you get a stand-in for the encoder output. The proxy doesn't reproduce the encoder exactly, but the downstream layers only ever learned to condition on encoder-shaped tokens, and the proxy has that shape. No new parameters, no architecture change. Just don't throw the head away.

Glance Forcing's decoupling took a second read too. The mental model that finally clicked: high-noise steps decide what's in the frame and how it moves, low-noise steps decide what each surface looks like. Those want different objectives and different training signal. A single student network ends up averaging between them, and the average is worse than either half.

Where the three approaches diverge

Paper	Target	Core mechanism	Added parameters	Headline result
RECON	Slow convergence in diffusion transformer training	Condition later layers on vision-encoder tokens, reuse MLP head at inference	0	63x faster than SiT, 25x faster than REPA
Glance Forcing	Cost of distilling video diffusion into a causal few-step model	Split denoising into slow and fast experts with stage-specific objectives	Two small experts	VBench Total 84.04 to 84.22 at roughly one A100-GPU-hour
Mixed Poisson diffusion	Diffusion on discrete, Poisson-distributed data	Multivariate counting-process noising plus posterior sampling	New model class	Well suited to Poisson inverse problems

Key numbers

RECON converges 63x faster than SiT and 25x faster than REPA while adding zero parameters. Glance Forcing cuts trajectory collection to roughly one A100-GPU-hour and lifts VBench Total from 84.04 to 84.22. MPDM replaces Gaussian noising with multivariate counting processes for discrete, Poisson-distributed observations.

Common pitfalls

1. Reading the RECON speedup as a sampling speedup. It's a convergence number, and your inference cost after training is unchanged. If you budgeted for a 63x faster product, you'll be disappointed.
2. Collecting your own PF-ODE trajectories out of habit. Glance Forcing's whole claim is that you need one. If you're generating thousands because that's what the old pipeline did, you're paying thousands of GPU-hours for a result one A100-hour would have matched.
3. Reaching for a Gaussian diffusion prior on Poisson counts. The noising process and the observation model need to match. On PET or photon-counting data, a Gaussian prior fights the data at every step, and posterior sampling inherits that fight.
4. Throwing away the projection head after REPA-style training. That's exactly the artifact RECON needs at inference. If you've been discarding it, you'll have to retrain to get it back.
5. Assuming one student can absorb both noise regimes in few-step distillation. The quality collapse at a single trajectory isn't a tuning problem. It's representation interference between two objectives that genuinely want different things.

One thing to remember

When a diffusion training run stalls, the useful question is which single step in the pipeline is actually two steps glued together. RECON found it in the layer stack, where early layers got semantics and later layers didn't. Glance Forcing found it in the noise schedule, where high noise and low noise want separate networks. MPDM found it at the very bottom, in the assumption that noising has to be Gaussian. Each fix was to split the step and give each half its own objective. None of them needed a bigger model to do it.

The bottom line

1. If you're training a diffusion transformer from scratch, wire RECON in before you buy more GPUs. Zero new parameters and a 63x convergence speedup over SiT changes what a run costs, with no inference penalty afterward.
2. If you're converting a video diffusion model to causal autoregressive, don't pay for offline ODE trajectories. Glance Forcing beats the fully-trained baseline on one A100-hour, and the slow/fast expert split is the part to copy even if you build your own pipeline.
3. If your data is counts, stop pretraining a Gaussian prior. MPDM gives you a noising process that matches the observation model, and posterior sampling falls out of it. Watch for standardized benchmarks on Poisson inverse problems within the next year, and expect Gaussian priors to quietly leave PET and photon-counting pipelines before that.