Discrete Distribution Networks
AI
generative models
SHOW HN2025
Legend
Discrete Distribution Networks
I invented a new generative model and got accepted to ICLR
discrete-distribution-networks.github.io#generative model#discrete distribution#zero-shot conditional generation#tree latent
How HN took it
16 positive, 24 neutral or asking, 1 mixed, 7 critical, out of 48 comments about the project. Thread score +0.22.
Idea: 5 for, 1 againstPerformance: 0 for, 3 againstDesign: 3 for, 0 againstExecution: 1 for, 1 againstDocs: 0 for, 2 against
The thread, comment by comment
Each row is a top-level comment and its replies, in HN's order. Select a cell to open that comment on HN.
positive neutral or mixed critical about another commenter or something else the maker
What HN said
What people liked
Novel approach, a genuinely new generative method
From 3 comments
Super cool, I spent a lot of time playing with representation learning back in the day and the grids of MNIST digits took me right back :) A genuinely interesting and novel…
moconnor on HNImpressive, congrats.
nothrowaways on HNFirst, I think this is really cool. Its great to see novel generative architectures. Here are my thoughts on the statistics behind this. First, let D be the data sample. Start…
elchananHaas on HN
Objections
Results are preliminary and far from state of the art
From 1 comment
The first version of DDN was developed in less than three months, almost entirely by one person. Consequently, the experiments were preliminary and the results far from SoTA. The…
diyer22 on HN
Relation to normalizing flows and VQ-VAE questioned
From 2 comments
This sounds somewhat like a normalizing flow from a discrete space to a continuous space. I think there's a way you can rewrite your DDN layer as a normalizing flow which avoids…
cooljoseph on HNisn't this kind of like an 80% vq-vae?
serf on HN
Requests
Scaling to larger models and non-image domains
From 2 comments
Super cool, I spent a lot of time playing with representation learning back in the day and the grids of MNIST digits took me right back :) A genuinely interesting and novel…
moconnor on HNHow does it compare to state of the art models? Does it scale?
p1esk on HN
Reproduction details such as hardware and training curves
From 1 comment
Do you have any details on the experiment procedures? E.g. hardware, training time, loss curves? It is difficult to confidently reproduce research without at least some of these…
highd on HN
Questions
How does it compare to state of the art models
From 1 comment
How does it compare to state of the art models? Does it scale?
p1esk on HN
Can it be used for object detection or text-to-speech
From 2 comments
Can you train this model to detect objects (e.g detect a fish in the picture)?
FitchApps on HNCould this be used to train a text -> audio model? I'm thinking of an architecture that uses RVQ. Would RVQ still be necessary?
cellis on HN
Alternatives named
Diffusion models
From 2 comments
isn't this kind of like an 80% vq-vae?
serf on HNI believe DDN is exceptionally well-suited to the “generative models for discriminative tasks” paradigm for object detection. Much like DiffusionDet, which applies diffusion…
diyer22 on HN
VQ-VAE
From 2 comments
isn't this kind of like an 80% vq-vae?
serf on HNNo, DDN and VQ-VAE are clearly different. Similarities: - Both map data to a discrete latent space. Differences: - VQ-VAE needs an external prior over code indices (e.g. PixelCNN…
diyer22 on HN
Normalizing flows
From 1 comment
This sounds somewhat like a normalizing flow from a discrete space to a continuous space. I think there's a way you can rewrite your DDN layer as a normalizing flow which avoids…
cooljoseph on HN