YuqingWang1029

YuqingWang1029 / CubiD

Public

[CVPR2026] Cubic Discrete Diffusion: Discrete Visual Generation on High-Dimensional Representation Tokens

46
0
100% credibility
Found Mar 23, 2026 at 46 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

CubiD is a PyTorch codebase for training and evaluating a cubic discrete diffusion model that generates images from high-dimensional discrete visual tokens.

How It Works

1
📖 Discover CubiD

You stumble upon a cool research project that creates realistic images using smart building blocks from pictures.

2
🖼️ Gather picture collection

You download a large folder of everyday photos with labels, like animals and objects, to help it learn.

3
📥 Grab ready-made helpers

You pick up pre-trained image understanding and rebuilding tools from a trusted sharing site.

4
⚡ Save quick insights

You process your photos once to store useful summaries, speeding up everything later.

5
🎓 Train the image maker

You start the learning process on powerful computers, watching it get better at inventing new pictures.

6
🖌️ Create new images

You give it labels like 'a sunny beach' and let it generate fresh, detailed pictures.

🎉 Enjoy amazing results

You see high-quality generated images that look real, check their sharpness scores, and feel proud of your creations.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 46 to 46 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is CubiD?

CubiD delivers cubic discrete diffusion in Python and PyTorch for generating images from high-dimensional representation tokens, tackling the gap where visual diffusion sticks to low-dim tokens unlike LLMs. It processes 768-dim DINOv2 features via RAE encoding, applying cubic masking across spatial and channel axes for CVPR2026-level discrete generation. Users run torchrun scripts to train on ImageNet, cache latents, evaluate FID/IS on 50k 256x256 images, or generate with CFG and pre-trained Hugging Face models.

Why is it gaining traction?

It stands out by modeling dependencies in full high-dimensional token space—think cubidesign for cubidi regenbogenball-like visuals—without collapsing to 1D sequences, yielding sharper semantic generation than pixel or VQ-based diffusion. Devs dig the flexible sampling (top-k/p, temperature, 256-1536 steps), EMA support, and distributed eval that spits out metrics fast. Early arXiv hype around discrete high-dim tokens draws diffusion tinkerers.

Who should use this?

Vision researchers prototyping discrete generation beyond pixels, like cubidi magic snake or cubidi ball anleitung visuals in high-dim spaces. Diffusion model devs needing cubic/cubidyn token pipelines for ImageNet-scale eval. Academic teams chasing CVPR2026 cubic discrete benchmarks with RAE integration.

Verdict

Promising for discrete diffusion experiments, but at 46 stars and 1.0% credibility, it's raw—docs lean on external RAE/TokenBridge setups, no tests visible. Fork and eval pre-trains before heavy training commits.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.