Research map / Architectures and efficiency / Pruning and efficient tuning
Memory-Efficient Fine-Tuning for Quantized Diffusion Model
Hyogon Ryu, Seohyun Lim, Hyunjung Shim
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 multi-channel-wise scale update and 2 timestep experts to PEQA baseline (ablation) | quantized DDIM on CIFAR-10 · CIFAR-10 unconditional (8-bit) | FID: 4.39 -> 3.80 | Improved |
| add multi-channel-wise scale update to baseline (1 expert) (ablation) | quantized DDIM on CIFAR-10 · CIFAR-10 unconditional (8-bit) | FID: 4.39 -> 4.33 | Improved |
| add timestep-aware scales with 2 experts (no MCSU) (ablation) | quantized DDIM on CIFAR-10 · CIFAR-10 unconditional (8-bit) | FID: 4.39 -> 4.24 | Improved |
| increase timestep experts from 2 to 4 (with MCSU) (ablation) | TuneQDM · CIFAR-10 unconditional (8-bit) | FID: 3.80 -> 4.15 | Hurt |
| TuneQDM instead of baseline in 4-bit single-subject generation | quantized Stable Diffusion · Single-subject generation (4-bit) | DINO-I: 0.519 -> 0.551 | Improved |
| TuneQDM instead of baseline in 4-bit single-subject generation | quantized Stable Diffusion · Single-subject generation (4-bit) | CLIP-T: 0.313 -> 0.306 | Hurt |
| TuneQDM instead of baseline in 8-bit single-subject generation | quantized Stable Diffusion · Single-subject generation (8-bit) | DINO-I: 0.581 -> 0.578 | No clear effect |
| TuneQDM instead of baseline in multi-subject generation (4-bit) | quantized Stable Diffusion · Multi-subject generation (4-bit) | DINO-I: 0.275 -> 0.276 | No clear effect |
| TuneQDM vs QLoRA on CIFAR-10 | quantized DDIM · CIFAR-10 unconditional (8-bit) | FID: 4.15 -> 3.80 (QLoRA r=2 vs TuneQDM) | Improved |
| TuneQDM extra scale multiplications at inference | 4-bit linear layers · Linear-layer inference time | time (s): 6.94 -> 6.99 | Hurt |
Closest papers by meaning
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.92)
- Diffusion Tuning: Transferring Diffusion Models via Chain of Forgetting (NeurIPS 2024 · similarity 0.91)
- IntLoRA: Integral Low-rank Adaptation of Quantized Diffusion Models (ICML 2025 · similarity 0.91)
- Modulated Diffusion: Accelerating Generative Modeling with Modulated Quantization (ICML 2025 · similarity 0.90)
- Pioneering 4-Bit FP Quantization for Diffusion Models: Mixup-Sign Quantization and Timestep-Aware Fine-Tuning (CVPR 2025 · similarity 0.90)
- Qua2SeDiMo: Quantifiable Quantization Sensitivity of Diffusion Models (AAAI 2025 · similarity 0.90)
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.
