Research map / Image generation / Fast diffusion sampling
The Drift Kernel: Why Diffusion Models Change Even When Told Not To
Gokul Srinath Seetha Ram, Rashmi Elavazhagan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase noise strength from 0.1 to 0.4 (ablation) | SD15 under null prompts · strength sweep | MSE drift: 0.0027 -> 0.0083 | Hurt |
| replace null prompt with strict copy prompt (ablation) | SD15 · strength sweep | drift coefficient k_M: 0.0345 -> 0.0346 | No clear effect |
| replace null prompt with strict copy prompt (ablation) | SD21 · strength sweep | drift coefficient k_M: 0.0685 -> 0.0801 | No clear effect |
| replace null prompt with strict copy prompt (ablation) | SDXL · strength sweep | drift coefficient k_M: 0.0710 -> 0.0770 | No clear effect |
| use a larger model (SDXL vs SD15) | null-prompt drift · NoOp-Bench strength 0.3 | MSE drift: 0.0059 -> 0.0085 | Hurt |
| use instruction-tuned InstructPix2Pix instead of SD15 | null-prompt drift · NoOp-Bench strength 0.3 | MSE drift: 0.0059 -> 0.0052 (std 0.0061 -> 0.0071) | Mixed |
| raise strength for InstructPix2Pix (ablation) | InstructPix2Pix · strength sweep | MSE drift: 8% vs 207% increase for SD15 | No clear effect |
| switch image domain from aerial to artwork | null-prompt drift · domain comparison | MSE drift: 0.0022 -> 0.0124 | Hurt |
| use detailed copy instructions in InstructPix2Pix (ablation) | InstructPix2Pix · strength sweep | quadratic fit R2: 0.883 -> 0.059 | No clear effect |
Closest papers by meaning
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.89)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.89)
- Generalization through variance: how noise shapes inductive biases in diffusion models (ICLR 2025 · similarity 0.89)
- Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · similarity 0.88)
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.88)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
