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

ConvRot-quantized single-file ComfyUI checkpoints of ByteDance/SDXL-Lightning (4-step UNet linear quantization + calibration bias correction, with Conv2d, CLIP text encoders, and VAE kept unquantized).

Related SDXL ConvRot repositories:


Checkpoint Catalog

Recommended INT8 ConvRot Checkpoint (Direct Drop-In for ComfyUI models/checkpoints/)

File Quantized Layers File Size Baseline FP16 Size Reduction
sdxl-lightning-4step-int8-convrot.safetensors 659 / 739 INT8 (80 protected FP16) 5.03 GB 6.94 GB -27.5%

Experimental INT6 ConvRot Checkpoints (int6_experiments/)

Each 6-bit variant is provided both as a plug-and-play ComfyUI file (*-w6a8-*) and as a smaller bit-packed download (*-packed.safetensors) unpackable with scripts/unpack_int6.py.

Variant Direct ComfyUI File (INT8 container) Packed Download (unpack_int6.py) Quantized Scheme
INT6 Protected ConvRot int6_experiments/sdxl-lightning-4step-w6a8-int6-protected-convrot.safetensors (5.04 GB) int6_experiments/sdxl-lightning-4step-int6-protected-convrot-packed.safetensors (4.57 GB, -34.2%) 659 INT6 + 80 FP16
INT6 All-Linear ConvRot int6_experiments/sdxl-lightning-4step-w6a8-int6-all-convrot.safetensors (4.72 GB) int6_experiments/sdxl-lightning-4step-int6-all-convrot-packed.safetensors (4.16 GB, -40.0%) 739 / 739 INT6

Visual Comparisons (4 steps, sgm_uniform, 1024ร—1024)

Lightning p0 4-way Lightning p3 4-way Lightning p6 4-way


Evaluation Metrics (G4 GPU, ComfyUI v0.45.0, 8 prompts ร— 1024ร—1024)

All quality metrics compare generated images and latents against the unquantized FP16 checkpoint (q_*), alongside unquantized BF16 vs FP16 drift (b_*) as a reference baseline. Sensitivity impact ranking is archived in metrics/impact_lightning.json.

Variant CosMean / Min SSIM Mean / Min PSNR Mean / Min BF16 Ref (Cos / SSIM / PSNR) Warm Mean Speed Report
INT8 ConvRot (659/739)* 0.99401 / 0.97584 0.95126 / 0.89363 33.42 / 29.28 dB 0.99348 / 0.94868 / 32.76 dB 0.35s (FP16: 0.15s) g4_lightning_int8_summary.json
W6A8 INT6 Protected ConvRot 0.97686 / 0.95608 0.88127 / 0.83725 26.20 / 22.48 dB 0.99348 / 0.94868 / 32.76 dB 0.35s g4_lightning_int6_protected_summary.json
W6A8 INT6 All ConvRot 0.97232 / 0.94586 0.86867 / 0.81594 26.20 / 22.48 dB 0.99348 / 0.94868 / 32.76 dB 0.37s g4_lightning_int6_all_summary.json

*Note on INT8 file size: sdxl-lightning-4step-int8-convrot.safetensors (5,030,597,998 B) and the G4-evaluated rebuild (5,041,886,350 B) share identical top-80 FP16 protection (659/739 INT8 layers, calibration bias correction).


How to Use in ComfyUI

  1. INT8 ConvRot and *-w6a8-* files: Drop directly into ComfyUI/models/checkpoints/ and load with the standard Load Checkpoint node in ComfyUI (>= v0.45.0).
  2. *-packed.safetensors (smaller 6-bit downloads): Unpack first via scripts/unpack_int6.py:
    python scripts/unpack_int6.py sdxl-lightning-4step-int6-protected-convrot-packed.safetensors ComfyUI/models/checkpoints/sdxl-lightning-4step-w6a8-int6-protected-convrot.safetensors
    

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
53
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