SDXL Refiner 1.0 INT8 ConvRot for ComfyUI

ConvRot-quantized single-file ComfyUI checkpoint of stabilityai/stable-diffusion-xl-refiner-1.0 (all 484 UNet linear layers quantized to symmetric INT8 with Hadamard rotation + calibration bias correction; Conv2d, OpenCLIP-G text encoder, and VAE kept in FP16).

Related SDXL ConvRot repositories:


Checkpoint Details

File Quantized Layers File Size Baseline FP16 Size Reduction Bias Corrections
sdxl-refiner-1.0-int8-convrot.safetensors 484 / 484 INT8 4.64 GB 6.08 GB -23.6% All eligible linear biases
  • Drop straight into ComfyUI/models/checkpoints/ and load with the standard Load Checkpoint node (ComfyUI >= v0.45.0).

Visual Comparisons (A100 GPU, 40-step Base+Refiner handoff, 1024ร—1024)

Refiner p0 compare Refiner p2 compare


Evaluation Metrics (A100 GPU, ComfyUI v0.45.0, 8 prompts ร— 1024ร—1024)

Measured via scripts/verify_sdxl_quant.py comparing generated images and latents directly against the unquantized FP16 checkpoint (sd_xl_refiner_1.0_fp16.safetensors), alongside unquantized BF16 vs FP16 as a reference baseline (metrics/a100_refiner_int8_summary.json):

Metric vs FP16 INT8 ConvRot (484/484) Unquantized BF16 Reference Report
Latent Cosine Mean / Min 0.99989 / 0.99982 0.99992 / 0.99986 a100_refiner_int8_summary.json
SSIM Mean / Min 0.99154 / 0.98751 0.99087 / 0.98572 a100_refiner_int8_summary.json
PSNR Mean / Min 42.96 / 40.60 dB 42.88 / 40.00 dB a100_refiner_int8_summary.json

License & Use Restrictions (CreativeML Open RAIL++-M)

This checkpoint is a Derivative of the Model under the CreativeML Open RAIL++-M License (Copyright ยฉ 2023 Stability AI). See NOTICE for full attributions and modification details, and pass through LICENSE.md and NOTICE with any redistribution.

Downloads last month
11
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SUSSYEATING/SDXL-Refiner-1.0-ConvRot

Quantized
(3)
this model