Dataset · Yshh525
FLUX.2 klein latents with structural generation errors
Yshh525/rho-internalize-v1-20261021
Latents, states, and images from FLUX.2 klein for studying anatomical and structural generation artifacts.
What's inside
Collected internally are latent representations, intermediate states, and decoded images obtained during 22 GPU generation sessions. The volume ranges from 1k to 10k records with fixed seeds and prompts. The data covers typical model structural errors: extra fingers on hands, incorrect number of limbs in horses and kangaroos, as well as object artifacts such as a cat's tail, railings, or headphones.
What you can do with it
- Analysis of internal states of flow models when anatomical errors occur
- Training classifiers or detectors of generation artifacts
- Researching methods for correcting structural defects in text-to-image models
- Tasks
- Text → image
- Size
- 1K–10K rows
- Downloads this month
- 104.3K
- Likes
- 0
How to load it
from datasets import load_dataset
ds = load_dataset("Yshh525/rho-internalize-v1-20261021")