Most of the ML world spent last week arguing about Sora clone benchmark scores and parameter count announcements. While that was happening, five papers dropped on arXiv that actually change how we will build diffusion systems for the next two years.

None have flashy demo videos. None claim AGI adjacent capabilities. All of them fix fundamental, unacknowledged problems that every production diffusion engineer has been working around for years.

The discrete diffusion identity crisis

For three years the discrete diffusion field has been conducting a very expensive, very loud argument about nothing.

Every six months a new paper would come out proposing a new parameterization, a new noise schedule, a new loss function, and demonstrate 2 bits better ELBO. Everyone would rush to implement it. Six months later another paper would prove the old one was better.

This month that argument ended.

The authors of What Does a Discrete Diffusion Model Learn? proved that denoisers, score ratio predictors, and bridge plug-in predictors are all exactly the same mathematical object. They are just written in different coordinate systems. If you read the output of your trained neural network using the wrong conversion, you are not just getting worse performance. You are running an entirely different stochastic process than the one you trained.

Most importantly they proved the Oracle Distance theorem. Everyone had treated the ELBO as an upper bound. It is not. It is exact equality.

Negative ELBO = data entropy + path KL from learned reverse process to the true oracle reverse process.

There is no slack. There is no trick noise schedule that will give you a better lower bound. Every possible noising process has exactly the same achievable minimum ELBO. All prior papers that claimed ELBO improvements from schedule changes were only measuring how close their parameterization let them get to the universal hard limit.

This is the most important theoretical result for diffusion models since the original score matching paper. If you are building token, text, or discrete diffusion systems you should stop what you are doing and read this paper before writing another line of code.

What breaks when you erase concepts

Every engineer that has deployed a public text-to-image service has run concept erasure. Every one of you has seen the collateral damage.

Erase one celebrity and suddenly every human face gets distorted. Erase watermarks and skies turn grey. Erase violent content and fire stops generating correctly.

Everyone assumed this was an unavoidable tradeoff. It was not. It was a bug in the core design of every prior erasure method.

All existing erasure operators project out the target concept vector from cross attention values. This does not just remove the target. It removes every component of every other concept that shares any basis direction with the target. Collateral damage was not a side effect. It was guaranteed by the math.

The new CARE method fixes this. Instead of projecting out the target, it first constructs an orthonormal basis of all concepts you explicitly want to keep. It then builds an erase direction that is strictly orthogonal to this entire kept subspace. The edit is closed form, requires no fine tuning, and adds 3ms of one time offline computation per target concept.

Concept erasure benchmark comparison

All numbers below are measured on the standard 2026 concept erasure evaluation suite:

Method	Erasure F1	Non-target retention	Runtime per prompt	Fine tuning required
Baseline vector projection	0.92	0.61	0ms	No
SEER	0.87	0.74	210ms	No
Unconstrained Concept Erasure	0.91	0.71	47ms	Yes
CARE	0.90	0.89	3ms	No

CARE matches the erasure strength of the best existing methods while retaining almost all unrelated concept performance. This is not a marginal improvement. This makes training free concept erasure actually usable for production.

RADIANCE: fixing diffusion's most embarrassing failure mode

You have seen this bug a thousand times. You prompt for "a blue cat sitting on a red chair". You get a red cat sitting on a blue chair. Or just a cat. Or just a chair.

This is not bad training. This is not bad prompting. This is a fundamental failure mode of the entire open loop denoising trajectory.

Dominant concepts commit very early in the diffusion process. Once the model has decided there is a cat at timestep 800, it will never go back and add the blue attribute. All existing guidance methods fight this after the fact. None work reliably for rare attribute pairings.

RADIANCE fixes this without any retraining. It runs a closed loop controller inside the denoising loop. Every 20 steps it extracts intermediate latents, runs a lightweight CLIP probe to verify which requested concepts are actually present, and adjusts IP-Adapter scales only for missing components.

It does not crank global guidance. It applies targeted restoring force exactly where the trajectory has drifted.

This is the first general solution to the composition problem that works on existing unmodified base models. It adds 15% inference overhead.

MV-Forcing: long multi-view video is now solved

Until last week there was an unbreakable tradeoff for video diffusion. You could generate 10 second multi-view consistent video. Or you could generate 2 minute single view video. You could not do both.

MV-Forcing breaks this tradeoff completely.

The core insight is extremely simple, and extremely embarrassing that no one did this earlier. Stop trying to teach diffusion models geometry.

Diffusion models are very good at generating detail and appearance. They are catastrophically bad at rigid geometric consistency. MV-Forcing does not attempt to train geometry into the diffusion model. It inserts an off the shelf monocular 3D reconstruction model as an interface between generation steps.

When generation completes for one viewpoint, the system reconstructs full 3D structure, renders a perfect geometric prior for the next viewpoint, and only asks the diffusion model to add appearance detail.

The second critical breakthrough is spatio-temporal self forcing. This eliminates the exposure bias gap that caused every prior autoregressive video model to fall apart after 12 frames. The student model is trained exclusively on its own outputs during distillation. Train and inference distributions are identical.

The authors demonstrate 120 frame, 8 viewpoint consistent video with zero geometric drift. Generation length is effectively unbounded.

EquiEdit: breaking the consistency / editability tradeoff

Every video editing method lies exactly on one line. If you want perfect temporal consistency you can barely change anything. If you want strong edits the video flickers like a 1990s animated gif. Everyone accepted this was a fundamental unavoidable tradeoff.

EquiEdit breaks this line.

They did two things. First they replaced temporal attention blocks with a directional temporal Mamba module. That was the obvious improvement everyone expected.

The clever part is the noise injection. All existing editing methods add uniform noise across the entire latent. EquiEdit performs Fourier decomposition on the input latent, and injects edit noise only into high frequency components. All low frequency components that carry motion, camera path, and object timing remain completely untouched.

You can replace every object in the video, change the entire artistic style, and the original timing, movement and camera motion will remain exactly identical. No flickering. No warping. No temporal drift.

The architecture shift no one has noticed yet

All five of these papers share one extremely important property. None of them make the base model bigger. None of them train on more data.

Every single improvement this month came from:

1. Correcting theoretical mistakes that everyone accepted for years
2. Inserting narrow, specialized modules between existing components
3. Running closed loop control during inference
4. Stopping trying to make one neural network learn everything

This is the end of the pure scaling era for diffusion. We are not going to get 10x better models by making them 10x bigger. We are going to get 10x better models by fixing the stupid mistakes we have all been making this entire time.

Open questions

None of these results are final.

The discrete diffusion identity proof only applies to factorizing noise schedules. We still do not know if this holds for general CTMC processes.

CARE only operates on cross attention values. It will not erase concepts that have been baked into UNet convolution weights.

RADIANCE overhead can be optimized further, but it will always carry some cost over open loop generation.

MV-Forcing requires a working monocular reconstruction model. This works perfectly for real world scenes. It will fail for abstract, stylized or impossible geometry.

Closing

This is what good ML research looks like. No flashy demos. No press releases. No claims of paradigm shifts. Just people digging into the fundamentals, finding the bugs that everyone else ignored, and fixing them.

Most of you will not implement any of this this month. All of you will be using every single one of these techniques by the end of the year.