The paper measures the local dependency ratio (LDR) of FLUX1-dev and Stable Diffusion 3.5-medium when prompted to generate images containing text (e.g. 'a blackboard with formulas'). Both models show high LDR values across most denoising timesteps, indicating that each text region is denoised primarily from its own local noise rather than from global context. Correspondingly, generated images contain misspelled words, missing characters, and nonsensical symbol assemblies. The LDR for FLUX1 stays above 0.79 across all five measured timesteps, while SD 3.5 starts at 0.9692 and drops to 0.4534 only at the final timestep.
Evidence
observational
Key metric
LDR (prompt 'a blackboard with formulas', 50 total reverse timesteps): FLUX1 0.9253/0.8078/0.7992/0.8208/0.7994 at timesteps 50/40/30/20/10; SD 3.5 0.9692/0.8283/0.7912/0.5726/0.4534 at timesteps 50/40/30/20/10
Caveat
The authors note that SD 3.5 has a small LDR near the end of denoising (timestep 10) and state 'we do not know why.' LDR was computed via zeroth-order approximation due to memory constraints of the full Jacobian. Only one prompt ('a blackboard with formulas') was used for the LDR measurement, though multiple prompts and seeds (0-2) were used for the qualitative hallucination examples.