SDXL Base 1.0 ConvRot (INT8, W6/W8 Hybrid & INT6) for ComfyUI

ConvRot-quantized single-file ComfyUI checkpoints of stabilityai/stable-diffusion-xl-base-1.0 (full 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-base-1.0-int8-convrot.safetensors 739 / 739 INT8 4.72 GB 6.94 GB -32.0%

Experimental INT6 & W6/W8 Hybrid ConvRot Checkpoints (int6_experiments/)

Every 6-bit variant is provided both as a plug-and-play ComfyUI file (*-w6a8-* / *-w6w8-hybrid-*) 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
W6/W8 Hybrid ConvRot int6_experiments/sdxl-base-1.0-w6w8-hybrid-convrot.safetensors (4.72 GB) int6_experiments/sdxl-base-1.0-w6w8-hybrid-convrot-packed.safetensors (4.28 GB, -38.3%) 619 INT6 + 120 INT8
W6A8 INT6 ConvRot int6_experiments/sdxl-base-1.0-w6a8-int6-convrot.safetensors (4.72 GB) int6_experiments/sdxl-base-1.0-int6-convrot-packed.safetensors (4.16 GB, -40.0%) 739 / 739 INT6

Visual Comparisons (1024ร—1024, 50 steps)

Base p0 4-way Base p2 4-way Base p5 4-way Base 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 visual-equivalence reference baseline. Full JSON reports live in metrics/.

Variant CosMean / Min SSIM Mean / Min PSNR Mean / Min BF16 Ref (Cos / SSIM / PSNR) Warm Mean Speed Report
INT8 ConvRot (739/739) 0.99433 / 0.98653 0.95213 / 0.91965 32.82 / 30.11 dB 0.99186 / 0.94269 / 32.20 dB 4.63s (FP16: 2.90s) g4_base_int8_summary.json
W6/W8 Hybrid ConvRot 0.95784 / 0.89655 0.84762 / 0.72200 24.78 / 20.07 dB 0.99186 / 0.94269 / 32.20 dB 4.69s g4_base_w6w8_hybrid_summary.json
W6A8 INT6 ConvRot (739/739) 0.95743 / 0.89424 0.84064 / 0.72535 24.57 / 20.38 dB 0.99186 / 0.94269 / 32.20 dB 4.62s g4_base_int6_summary.json

Notice why INT8 ConvRot with calibration bias correction beats the unquantized BF16 vs FP16 reference on mean Cosine, SSIM, and PSNR (0.95213 vs 0.94269): INT8 symmetric grid resolution (255 levels) is finer near typical normalized weight distributions than BF16's 7-bit mantissa truncation.


How to Use in ComfyUI

  1. INT8 ConvRot and *-w6a8-* / *-w6w8-hybrid-* 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-base-1.0-int6-convrot-packed.safetensors ComfyUI/models/checkpoints/sdxl-base-1.0-w6a8-int6-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 2023 Stability AI). See NOTICE for full attributions and modification details.

Under Section III, Paragraph 5 & Attachment A of LICENSE.md, all users and downstream recipients are legally bound by the Open RAIL++-M Use-Based Restrictions and must pass through LICENSE.md and NOTICE with any redistribution.

Downloads last month
78
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

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

Quantized
(26)
this model