Researchers have developed Chameleon, a novel post-training quantization (PTQ) framework designed to optimize diffusion models for memory-constrained hardware. Unlike previous methods that fix the number format, Chameleon dynamically selects optimal formats per weight channel and activation layer, and optionally per timestep bucket. This dynamic approach allows for better encoding of varying activation distributions throughout the diffusion process. Experiments on models like SDXL, SDXL-Turbo, and PixArt-α demonstrate that Chameleon achieves superior FID scores and smaller weight sizes compared to existing quantized methods, while maintaining close fidelity to FP16 references. AI
IMPACT Enables more efficient deployment of large diffusion models on resource-constrained hardware, potentially broadening access and application.
RANK_REASON This is a research paper detailing a new technical method for optimizing AI models. [lever_c_demoted from research: ic=1 ai=1.0]
- Arnab Sanyal
- Chameleon
- COCO-2014
- FP4 E2M1
- FP8 E4M3
- FP8 E5M2
- INT4
- INT8
- MXFP4
- MXFP8
- MXINT4
- MXINT8
- Pixart
- SDXL
- SDXL-Turbo
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →