Research map / Image generation / Fast diffusion sampling
Unified Backbone Refinement for Diffusion Models via Internal-Latent Analysis
HAKSU LIM, Myeongjin Lee, Wonjoon Chang, Jaesik Choi
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 |
|---|---|---|---|
| apply DUNE detect-suppress refinement | SDXL · COCO human prompts | HADM-L: 0.227 -> 0.074 | Improved |
| apply DUNE | SDXL · COCO prompts | FID: 18.86 -> 18.75 | Improved |
| apply FreeU | SDXL · COCO prompts | FID: 18.86 -> 31.22 | Hurt |
| apply PAG | LCM · COCO prompts | FID: 22.53 -> 35.87 | Hurt |
| apply ASCED | PixArt-Sigma / Hunyuan-DiT · COCO prompts | FID: N/A (pure noise) | Hurt |
| apply DUNE | PixArt-Sigma · COCO human prompts | HADM-L: 0.316 -> 0.254 | Improved |
| apply DUNE | Hunyuan-DiT · COCO human prompts | HADM-L: 0.294 -> 0.283 | No clear effect |
| intervene at 3/4-depth self-attention layer instead of last layer (ablation) | DUNE · PixArt-Sigma | FID: 32.68 -> 27.80 | Improved |
| replace detector mask with random mask in h-space (50%) (ablation) | DUNE · SDXL | FID: 18.86 -> 20.59 | Hurt |
| replace detector mask with random mask in h-space (10%) (ablation) | DUNE · SDXL | FID: 18.86 -> 18.92 | No clear effect |
| suppress shallow skip/upsampling features instead of h-space (ablation) | DUNE · U-Net hallucinated regions | score deviation | Hurt |
Closest papers by meaning
- Backbone-Equated Diffusion OOD via Sparse Internal Snapshots (NeurIPS 2026 · similarity 0.88)
- Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 2024 · similarity 0.88)
- Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.88)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.87)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.87)
- Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models (NeurIPS 2026 · similarity 0.87)
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.
