SDXL-Lightning 4-Step ConvRot (INT8, INT6 & INT4) for ComfyUI

ConvRot-quantized single-file ComfyUI checkpoints of ByteDance/SDXL-Lightning in three clean bit widths: Full INT8, Full INT6 (w6a8_int8), and Full INT4 (asym_w4a8_int8).

Built with ConvRot (block Hadamard rotation) + Act-Order GPTQ Hessian error compensation + Hessian-weighted per-row clip search (INT8) / through-grid FP8 group-scale search / Lloyd-Max codebooks + calibration bias correction, keeping Conv2d layers, text encoders (CLIP-L / OpenCLIP-G), and VAE unquantized in FP16.

Related SDXL ConvRot repositories:


Available Checkpoints

Precision Tier File (ComfyUI/models/checkpoints/) Size (vs 6.94 GB FP16) Quantized UNet Linears
Full INT8 ConvRot sdxl-lightning-4step-int8-convrot.safetensors 4.72 GB (-32.0%) 739 / 739 INT8
Full INT6 ConvRot sdxl-lightning-4step-int6-convrot.safetensors 4.31 GB (-37.8%) 658 / 739 INT6 (W6A8) + 81 INT8 (K=640/det)
Full INT4 ConvRot sdxl-lightning-4step-int4-convrot.safetensors 3.77 GB (-45.6%) 658 / 739 INT4 (W4A8) + 81 INT8 (K=640/det)

All files load directly in stock ComfyUI (Load Checkpoint, ComfyUI >= v0.45.0) using native comfy-kitchen CUDA kernels โ€” no extra nodes or unpacking required. (For SDXL Base/Lightning/Turbo, 80 small K=640 linear layers plus the single architecture-detection to_k key remain in ConvRot INT8 because comfy-kitchen w6a8/w4a8 kernels require K % 256 == 0.)


Visual Comparisons (4 steps, sgm_uniform, 1024x1024)

Lightning Comparison p0 Lightning Comparison p3 Lightning Comparison p6


Evaluation Metrics (8 prompts x 1024x1024, ComfyUI v0.45.3)

Measured via scripts/verify_sdxl_quant.py directly against unquantized FP16 (sdxl_lightning_4step.safetensors):

  • Unquantized BF16 Reference (vs FP16): Cosine 0.99348 (min 0.97744), SSIM 0.94868 (min 0.90096), PSNR 32.76 dB (min 29.26 dB)
Precision Tier Cosine Mean / Min SSIM Mean / Min PSNR Mean / Min Warm Speed/Img Report
Full INT8 ConvRot 0.99498 / 0.98800 0.95632 / 0.91929 33.38 / 27.99 dB 0.39s g4_lightning_int8_summary.json
Full INT6 ConvRot 0.98989 / 0.95676 0.93520 / 0.84038 32.03 / 26.34 dB 0.44s g4_lightning_int6_summary.json
Full INT4 ConvRot 0.96818 / 0.93316 0.85894 / 0.78216 25.44 / 22.99 dB 0.43s g4_lightning_int4_summary.json

How to Use in ComfyUI

  1. Download INT8, INT6, or INT4 .safetensors into ComfyUI/models/checkpoints/.
  2. Load directly with ComfyUI's built-in Load Checkpoint node (ComfyUI >= v0.45.0).

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

All checkpoints in this repository are Derivatives of the Model under the CreativeML Open RAIL++-M License (Copyright ยฉ 2024 Bytedance Inc.; 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
122
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SUSSYEATING/SDXL-Lightning-ConvRot

Quantized
(3)
this model