SDXL Turbo INT8 ConvRot for ComfyUI

Powered by Stability AI

Quantized single-file ComfyUI checkpoint derived from stabilityai/sdxl-turbo using ConvRot (block Hadamard rotation + row-wise symmetric INT8 quantization + calibration bias correction) while keeping the 80 most sensitive UNet linear layers preserved in FP16.

โš ๏ธ License Separation Notice: This model is hosted in a dedicated repository because it is governed by the Stability AI Community License Agreement rather than Open RAIL++-M or Apache-2.0. Please review the license terms below before commercial or downstream use.


Checkpoint Details

File Quantized Layers File Size Baseline FP16 Size Reduction Bias Corrections
sdxl-turbo-1.0-int8-convrot.safetensors 659 / 739 INT8 (80 protected FP16) 5.18 GB 6.94 GB -25.3% 256 linear biases
  • Drop straight into ComfyUI/models/checkpoints/ and load with the standard Load Checkpoint node (ComfyUI >= v0.45.0).
  • Conv2d layers, text encoders (CLIP-L / OpenCLIP-G), and VAE remain unquantized.

Visual Comparisons (512ร—512, 2 steps, SDTurboScheduler)

Turbo Prompt 3 Compare Turbo Prompt 7 Compare


Evaluation Metrics (A100 GPU, ComfyUI v0.45.0, 24 Renders: 8 Prompts ร— 3 Seeds)

Measured via scripts/verify_sdxl_quant.py at 512ร—512, 2 steps, comparing directly against the unquantized FP16 checkpoint (sd_xl_turbo_1.0_fp16.safetensors), with unquantized BF16 (b_*) included as a baseline reference:

Metric vs FP16 INT8 ConvRot (659/739 INT8) Unquantized BF16 Reference Notes
Latent / Feature Cosine Mean 0.99805 (min 0.98357) 0.99802 (min 0.99386) Higher mean than BF16
SSIM Mean 0.96659 (min 0.84238) 0.92631 (min 0.88360) +0.04028 higher mean SSIM than BF16
PSNR Mean 34.39 dB (min 26.97 dB) 31.18 dB (min 27.16 dB) +3.21 dB higher mean PSNR than BF16
Warm Mean Time / Image 0.69s 0.22s Hadamard dequant overhead traded for smaller size

Full summary JSON: metrics/a100_turbo_int8_summary.json.


Stability AI Community License & Acceptable Use Policy

This model is a Derivative Work of stabilityai/sdxl-turbo and is distributed under the Stability AI Community License Agreement (sai-nc-community), Copyright ยฉ Stability AI Ltd. All Rights Reserved. See also the included NOTICE file.

Key terms under LICENSE.md:

Downloads last month
14
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SUSSYEATING/SDXL-Turbo-ConvRot

Quantized
(11)
this model