Ada Görgün

PhD Student @cvml_mpiinf

[Ada Görgün @AdaGorgun]

Huge thanks to my fantastic co-authors @fawaz_sammani (equal contribution), Nikos Deligiannis, Bernt Schiele, and @JonasFischerML ! Paper: https://t.co/y6nEdfRSao Page: https://t.co/QViROAdoSL: https://t.co/24NnOUjMOo #ICLR #ICLR2026 #Diffusion #DiffusionModels #XAI📷

[Ada Görgün @AdaGorgun]

Diffusion models create stunning images, but when and how do specific concepts actually emerge during the denoising process?

🎉 Excited to share that our paper “Temporal Concept Dynamics in Diffusion Models via Prompt-Conditioned Interventions” has been accepted to ICLR 2026! 😊

[Ada Görgün @AdaGorgun]

✨ Key contributions:
1. PCI: a simple controllable framework to insert concepts mid-generation.
2. CIS: a temporal metric revealing concept emergence and stability.
3. Applications: identifying failure modes, discovering concept timelines, and enabling temporally-aware editing.

[Ada Görgün @AdaGorgun]

Until now, most analyses looked only at the final timestep.
PCI (Prompt-Conditioned Intervention) flips the perspective: we intervene during denoising to measure Concept Insertion Success (CIS) across timesteps, revealing how different concepts form, stabilize, or fail.

[Ada Görgün @AdaGorgun]

💡 Why this matters:
PCI offers a new perspective for understanding and diagnosing diffusion models through their temporal behavior. It is training free, model agnostic, and requires no fine tuning or external supervision.