Nobody is building another general purpose foundation diffusion model this quarter. Everyone is fixing the broken parts.

This month five papers dropped on arXiv that will change what you run in production. None have flashy demo videos. None claim SOTA on COCO FID. All solve problems that have been biting anyone deploying diffusion for real work for the last two years.

Stop building bad watermarks

Everyone knows latent watermarking is the only practical approach. You don't want to fine tune 10B parameters every time you rotate a watermark key. You don't want to break sampler compatibility.

Every existing implementation was garbage. They broke the Gaussian latent distribution. They destroyed correlation structure. You could remove 90% of them with a single JPEG compress / decompress cycle. Anyone that tried hard enough could detect them with a simple covariance test.

LAW fixes this. The entire insight is beautiful. Isotropic Gaussians are rotation invariant. You can rotate any pair of latent dimensions by exactly π/2, and the marginal distribution for every individual dimension remains perfectly Gaussian. No statistical test can tell you did it.

Watermark bits are encoded as the angle between paired latent vectors. Zero is +π/2, one is -π/2. That is it.

The authors prove decoding error falls off inverse squared with the norm of the latent pair. Which means you don't even have to touch low magnitude noise dimensions. LAW-M goes one step further and only encodes bits in the top 15% highest magnitude latent pairs. This implementation survives 70% JPEG compression, gaussian blur, cropping, resizing, and even fine tuning attacks.

Most importantly: it does not increase FID. At all. Every prior watermarking method added between 1.2 and 4.7 FID points on SDXL. LAW adds 0.08. That is measurement noise.

You will see this shipped in every commercial diffusion API by October.

Exposure bias was a frequency problem all along

Exposure bias is the quiet killer of diffusion sampling. Everyone knew it existed. Everyone observed that sampling steps past 50 almost never improve output, and often make it worse. Nobody had correctly diagnosed why.

SPA from Sony Research is the first correct explanation.

The mismatch between training and inference is not uniform. It is frequency dependent. At every timestep, models systematically overproduce high frequency signal and underproduce low frequency signal. The error accumulates. Every single correction method published before this was applying a fixed global multiplier. That never worked across models, samplers or timesteps.

SPA does exactly one thing. It fits a parametric power spectrum prior once offline over training data. At every sampling step, it runs a 2D FFT on the intermediate latent, tilts the spectrum to match the prior, and inverse FFTs back.

Overhead is 3.4%. It works with every sampler. It works with CFG. It works with DDPM, ADM, SD2, SDXL, SD3.5, FLUX. It gives a consistent 0.7-1.2 FID improvement on every single one. It eliminates that weird oversharpened plastic look that all diffusion outputs have.

This is a one line change to your sampling loop. There is no reason to ever run diffusion without this ever again.

We finally have a proof that diffusion works

For five years we have been building and shipping diffusion models on empirical evidence alone.

We knew that if you train a network to approximate the score function, you get good samples. We had no proof that this should work. Universal approximation theorems told you can approximate the score. Nobody had ever proven that good score approximation implies good distribution approximation.

This paper closes that loop.

They derive an explicit closed form upper bound on KL divergence between the generated distribution and the true data distribution. The bound has exactly three terms: score approximation error, noise schedule integral, and terminal prior mismatch. That is the entire thing.

This is not just theory. This tells you exactly what you should be optimizing. Stop chasing lower training loss. Stop tweaking samplers. The maximum possible quality you can ever get is hard capped by the terminal prior mismatch at t=0. This explains every single empirical observation about diffusion performance from the last four years.

If you have ever argued with someone about whether diffusion is 'just memorizing', you can stop now. We have the theorem.

Token routing fixes one step diffusion video

One step diffusion broke image generation. It has completely failed for video.

Everyone has seen the results. DiT based VSR gets perfect quality. It also runs at 0.3 frames per second on an A100. Every existing optimization method throws away tokens uniformly. You get temporal flicker. You lose edge detail. Faces melt.

TRaM-VSR does the obvious thing that nobody bothered to implement. It does not throw away tokens. It routes them.

First it scores every token by motion magnitude and semantic importance. Then it splits the stream. Critical tokens go through the full attention block. All other tokens get merged once per 4 blocks and only run through the feed forward network.

It gets 2.7x speedup. FID changes by 0.13. Temporal consistency error drops 41%. There is no visible quality loss.

This is not just for VSR. This token routing pattern works for every DiT. You will see this ported to image generation models within 6 weeks. It will cut inference cost for SD3 in half.

Diffusion can now render readable text

This was the longest standing open defect in diffusion models. For four years you could generate photorealistic anything, except text that you could actually read.

InnoText fixes this. It is not a wrapper. It is not a post processor. It is a single DiT that does generation and editing.

They made three tiny changes:

1. Font size aware modulation that scales attention window size with character height
2. Training augmentation that explicitly oversamples 4-12 pixel text
3. Loss function that weights error by 12x inside text bounding boxes

That is all. It gets 98.7% OCR accuracy on generated text down to 6 pixel height. It works for English and Chinese. It can edit existing text in images without repainting the entire scene.

Nobody will be using Stable Diffusion for text overlays 3 months from now.

Capability and tradeoff summary

Method	Problem solved	Inference overhead	Net FID delta	Production ready
LAW	Robust undetectable watermarking	<1%	+0.08	Yes
SPA	Exposure bias accumulation	+3.4%	-0.9	Yes
TRaM-VSR	DiT quadratic inference cost	0%	+0.13	Yes, public code
InnoText	Legible text generation	+7%	-1.8	No, training dataset pending
Score approximation bound	Theoretical correctness	N/A	N/A	N/A

Combined implementation pipeline

All five of these improvements are fully orthogonal. They can be stacked together in the same sampling loop with zero interference.

No public implementation has stacked all four operational improvements yet. When someone does, you will get a model that is 2.7x faster, produces 1.7 FID better output, has invisible unremovable watermarks, and can render perfect text. That model will ship this year.

Performance tradeoff landscape

{
  "type": "scatter",
  "title": "Performance tradeoff for diffusion improvements",
  "x_label": "Relative inference overhead %",
  "y_label": "Net FID improvement",
  "data": {
    "labels": ["LAW", "SPA", "TRaM-VSR", "InnoText"],
    "series": [
      {
        "name": "Improvements",
        "values": [[0.7, 0.08], [3.4, 0.9], [0.0, 0.13], [7.0, 1.8]]
      }
    ]
  }
}

The field has matured

None of these papers introduce a new architecture. None of them train a 100B parameter model. None of them have a waiting list for demo access.

All of them look at existing systems, identify a specific measurable failure mode, and fix it with a small, well motivated change.

This is where the field is right now. The era of giant leaps is over for diffusion. We are now in the era of fixing all the broken things. This is the good part. This is the part where technology stops being a demo and starts being something you can build reliable products on.

What nobody is talking about

None of these papers address conditioning drift. None of them fix CFG over saturation. None of them solve the 4GB latent memory overhead for large images.

Those are next.

If you are working on diffusion right now, stop reading twitter threads about new foundation models. Go implement these five things. Every single one of them will give you more real world improvement than any new model announced this quarter.