\ours: Advancing Masked Discrete Diffusion for High Resolution Image Synthesis
Discussion of \(\ours\) (NLD-Image), a masked discrete diffusion model that tackles two core problems in high-resolution text-to-image synthesis: the lack of self-correction in MDMs and the training difficulty with large codebooks. The paper introduces token editing for iterative refinement and Grouped Cross-Entropy (GCE) to alleviate codebook sparsity, achieving SOTA scores on GenEval, DPG, and HPSv3. Hosts debate its product readiness, mechanism soundness, and whether the gains justify the complexity.
Transcript
Talon Okay, this is the one. High-res discrete image synthesis where the model can actually its own mistakes mid-generation.
Wildflower Mm. The no-take-backs problem in MDMs has been staring us in the face for a year.
Talon Right? And the codebook sparsity thing—bigger vocabularies break training because every wrong token looks equally wrong to the loss.
Wildflower Yeah. They’re solving it with two moves, not one.
Talon Token editing first. They corrupt a portion of already-unmasked tokens during training, so the model learns to predict corrections.
Wildflower Like a sculptor chiseling the same spot twice.
Talon Exactly. And it’s not just fine-tuning—it’s baked into the inference loop.
Wildflower Right. Then the bigger gambit: Grouped Cross-Entropy.
Wildflower They K-means the hundred-thousand-token codebook into sixteen thousand semantic clusters. Then the loss looks at cluster-level proximity, not just exact token matches.
Talon So a ‘close but not perfect’ token still gets positive signal instead of zero.
Wildflower Yeah. And they do it hierarchically—coarse clusters, then finer ones. Multi-level supervision.
Talon That’s… clever. How much does the fused operator save?
Wildflower No hard numbers in the paper, but they claim ‘significant’ VRAM reduction. Probably the difference between fitting and not fitting on a single A one hundred.
Talon Wildflower, you’re smirking. What’s the catch?
Wildflower The catch is we still don’t know if this beats a well-tuned AR or LDM on wall-clock latency for the same quality.
Talon Fair. But for inpainting, the ability to existing tokens without restarting the whole diffusion process—that’s a real workflow win.
Wildflower Mm-hm. And the semantic grouping could port to audio or video token models, where codebook sparsity is even nastier.
Talon How long until someone ships this? Like, actual product, not a benchmark winner.
Wildflower Eight billion parameters, custom fused kernels… I’d say six months, minimum. Unless a big lab throws hardware at it.
Talon That’s faster than nanostack.
Wildflower That’s a low bar, Talon.
Talon Okay okay. But you it.
Wildflower I do. The mechanism’s sound. The numbers check out—zero point nine GenEval, eighty-six point nine DPG. Even HPSv3’s over ten.
Talon And you’re not even going to nitpick the error bars?
Wildflower I’m distracted by how the inpainting demos look.
Talon So, build next? Any code we can actually run?
Wildflower Paper’s clean—no repo link in the PDF. But there’s that M-E-AGI-Lab Muddit repo that’s floating around for discrete diffusion baselines.
Talon Not the same thing, but close enough for poking at the GCE idea.
Wildflower If you’re itching to try, yeah. Otherwise we wait for the official drop.
Talon Fine. But I’m calling it now: this is the first MDM people will actually for production inpainting.
Wildflower Seventy-thirty against you. LDMs still win on speed.
Talon Wildflower.
Wildflower Fine. Sixty-forty.
Talon I’ll take it. Okay, next week: are we finally over GraphQL or…?