hu-zijing/D-Fusion
[ICML 25] Denoising trajectory fusion, a method to construct RL-trainable visually consistent samples.
This project helps researchers and practitioners fine-tune diffusion models to better understand what makes a generated image align with a text prompt. It takes a diffusion model and poorly-aligned image-text pairs, and outputs improved, visually consistent image samples that are suitable for further training. This is for machine learning researchers and AI developers working on improving text-to-image generation.
No commits in the last 6 months.
Use this if you are working on aligning diffusion models with text prompts and need to generate visually consistent samples for effective direct preference optimization (DPO) training.
Not ideal if you are a casual user looking to generate images or if you need a solution compatible with the absolute latest versions of diffusion model libraries without manual code adaptation.
Stars
11
Forks
1
Language
Python
License
MIT
Category
Last pushed
Jun 02, 2025
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/hu-zijing/D-Fusion"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
FlorianFuerrutter/genQC
Generative Quantum Circuits
horseee/DeepCache
[CVPR 2024] DeepCache: Accelerating Diffusion Models for Free
Gen-Verse/MMaDA
MMaDA - Open-Sourced Multimodal Large Diffusion Language Models (dLLMs with block diffusion,...
kuleshov-group/mdlm
[NeurIPS 2024] Simple and Effective Masked Diffusion Language Model
Shark-NLP/DiffuSeq
[ICLR'23] DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models