Research map / Architectures and efficiency / Pruning and efficient tuning
FINE: Factorizing Knowledge for Initialization of Variable-sized Diffusion Models
Yucheng Xie, Fu Feng, Ruixiao Shi, Jianlu Shen, Jing Wang, Yong Rui, Xin Geng
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 |
|---|---|---|---|
| add shared-singular-vector knowledge factorization (vs independent per-layer SVD) (ablation) | FINE · ImageNet-1K DiT-B L6 | FID: 62.86 -> 51.58 | Improved |
| add shared-singular-vector knowledge factorization (vs independent per-layer SVD) (ablation) | FINE · ImageNet-1K DiT-L L6 | FID: 56.42 -> 44.38 | Improved |
| use trainable Sigma init instead of random (ablation) | FINE · ImageNet-1K DiT-B L12 | FID: 77.70 -> 42.74 | Improved |
| use trainable Sigma init instead of identical init (ablation) | FINE · ImageNet-1K DiT-B L12 | FID: 47.84 -> 42.74 | Improved |
| use trainable Sigma init instead of linear init (ablation) | FINE · ImageNet-1K DiT-B L12 | FID: 46.71 -> 42.74 | Improved |
| use trainable Sigma init instead of random (ablation) | FINE · ImageNet-1K DiT-L L12 | FID: 73.58 -> 35.59 | Improved |
| use trainable Sigma init instead of linear init (ablation) | FINE · ImageNet-1K DiT-L L12 | FID: 39.34 -> 35.59 | Improved |
| initialize with FINE instead of TLEG | DiT-B L10 · ImageNet-1K | FID: 47.22 -> 42.33 | Improved |
| initialize with FINE instead of He-Init | DiT-B L6 · ImageNet-1K | FID: 80.37 -> 51.58 | Improved |
| initialize with FINE instead of fine-tuning the full pre-trained model | DiT-B · LSUN-Bedroom | FID: 24.43 -> 17.83 | Improved |
| initialize with FINE instead of fine-tuning the full pre-trained model | DiT-L · CelebA | FID: 8.65 -> 8.41 | Mixed |
| initialize with FINE instead of Auto-LG | DeiT-Ti L12 · ImageNet-1K classification | accuracy: 66.79 -> 67.02 | No clear effect |
Closest papers by meaning
- Large Language Models to Diffusion Finetuning (ICML 2025 · similarity 0.90)
- Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.89)
- Transfer Learning for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- SaRA: High-Efficient Diffusion Model Fine-tuning with Progressive Sparse Low-Rank Adaptation (ICLR 2025 · similarity 0.89)
- Generalization in VAE and Diffusion Models: A Unified Information-Theoretic Analysis (ICLR 2025 · similarity 0.89)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.89)
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.
