Research lines

Learn visual features by aligning images to their captions across web-scale data.

The default choice. Most systems you encounter use this.

Learn visual features with no labels, no language and no pixel reconstruction, by having a network predict its own slowly-updated teacher.

The default choice. Most systems you encounter use this.

Compress images with a reconstruction-trained autoencoder, then run diffusion or flow matching in that latent space.

The default choice. Most systems you encounter use this.

Distil several specialist foundation models into one backbone rather than choosing between their objectives.

Gaining ground quickly, with results that are holding up.

Autoregress over continuous vectors, scoring each with a small diffusion head instead of a softmax over a codebook.

Gaining ground quickly, with results that are holding up.

Freeze a vision foundation model and generate directly inside its feature space; train only a decoder back to pixels.

Live disagreement about whether it is the right bet. Not a verdict of failure.

Drop the tokenizer entirely and generate raw pixels with one plain transformer, in one stage, with no cascade.

Live disagreement about whether it is the right bet. Not a verdict of failure.

Take a foundation-model representation and give it back the low-level detail it discards, with a residual branch, detail queries, or a VAE compressing it.

Young, promising, and not yet independently replicated at scale.

Perform the edit inside a foundation-model feature space rather than a reconstruction latent.

Young, promising, and not yet independently replicated at scale.

Keep generation in pixels, but handle resolution progressively: generate small, then refine upward.

Largely displaced, kept for the history it explains.

Regress the velocity field of a straight path between noise and data, rather than reversing a stochastic noising process.

The default choice. Most systems you encounter use this.

Quantize an image into discrete tokens and predict them with a language-model-style autoregressive transformer.

Live disagreement about whether it is the right bet. Not a verdict of failure.

Train for one or few steps from the start, instead of training a many-step model and distilling it afterwards.

Young, promising, and not yet independently replicated at scale.

Return to exact-likelihood invertible models, scaled up with transformers and trained in a latent space.

Young, promising, and not yet independently replicated at scale.

Model generation as discrete-time Markov transitions over a continuous state, unifying flow matching and continuous-token autoregression.

Young, promising, and not yet independently replicated at scale.

Keep the autoencoder, but apply pressure from a vision foundation model so its latent (or the generator's hidden states) become semantically structured.

Gaining ground quickly, with results that are holding up.

Fine-tune generators against reward models or human preference with policy-gradient methods, rather than only maximizing likelihood.

Gaining ground quickly, with results that are holding up.

Change what happens at inference time — guidance formulation, stochasticity, step schedule — without retraining the model.

Gaining ground quickly, with results that are holding up.

Edit by intervening in attention maps and features at sampling time, with no training and often no change to the latent at all.

Gaining ground quickly, with results that are holding up.

Recover the noise that produced a real image so it can be re-generated with a change, and control the error that recovery introduces.

Gaining ground quickly, with results that are holding up.

Measure whether generators actually do what their scores claim, and fix the benchmarks when they stop tracking human judgment.

Gaining ground quickly, with results that are holding up.

Train an editor that reads an instruction and rewrites the image inside a standard reconstruction-trained latent.

The default choice. Most systems you encounter use this.

Edit by putting the reference image into the model's own token sequence and letting attention handle it, rather than through adapters or inversion.

The default choice. Most systems you encounter use this.

Inject visual information through dedicated gated cross-attention layers interleaved into a frozen language model.

Live disagreement about whether it is the right bet. Not a verdict of failure.

Bolt a small trainable side network onto a frozen pretrained diffusion model to inject a new control signal, instead of retraining or restructuring the backbone itself.

Live disagreement about whether it is the right bet. Not a verdict of failure.

Give the generator tools, search, or a planning loop, so it can render subjects its training data never covered.

Young, promising, and not yet independently replicated at scale.

Bolt a frozen vision encoder onto a language model with a small trained projection, and treat the projected patches as prompt tokens.

The default choice. Most systems you encounter use this.

Build one model that both interprets images and produces them, instead of separate understanding and generation systems.

Gaining ground quickly, with results that are holding up.

Process images at their own resolution and aspect ratio, with position encodings general enough to handle the variation.

Gaining ground quickly, with results that are holding up.

Treat editing as one more thing a single multimodal model does, with no editing-specific pipeline.

Gaining ground quickly, with results that are holding up.