FP Compare

FP32_reduce_block_noise_

Images and Data

FP32

FP32_1 FP32_2 FP32_3 FP32_4 FP32_6 FP32_7 FP32_8
FP32 MAE_Similarity SSIM_Similarity Time (sec)
No.1 100.0 % 100.0 % 374.0
No.2 100.0 % 100.0 % 371.2
No.3 100.0 % 100.0 % 371.1
No.4 100.0 % 100.0 % 370.5
No.5 100.0 % 100.0 % 371.4
No.6 100.0 % 100.0 % 371.2
No.7 100.0 % 100.0 % 370.8
median 100.0 % 100.0 % 371.2

FP16

FP16_1 FP16_2 FP16_3 FP16_4 FP16_6 FP16_7 FP16_8
FP16 MAE_Similarity SSIM_Similarity Time (sec)
No.1 94.0 % 93.4 % 103.4
No.2 92.7 % 86.5 % 98.7
No.3 91.8 % 84.8 % 98.5
No.4 91.8 % 79.8 % 98.5
No.5 92.7 % 87.6 % 98.6
No.6 92.6 % 89.2 % 98.5
No.7 93.4 % 88.0 % 98.6
median 92.7 % 87.6 % 98.6

BF16

BF16_1 BF16_2 BF16_3 BF16_4 BF16_6 BF16_7 BF16_8
BF16 MAE_Similarity SSIM_Similarity Time (sec)
No.1 94.0 % 92.8 % 100.6
No.2 93.7 % 90.1 % 96.0
No.3 91.6 % 83.7 % 96.0
No.4 91.7 % 79.7 % 95.9
No.5 92.6 % 87.5 % 95.9
No.6 92.6 % 89.5 % 95.9
No.7 90.4 % 77.4 % 96.0
median 92.6 % 87.5 % 96.0

MXFP8

MXFP8_1 MXFP8_2 MXFP8_3 MXFP8_4 MXFP8_6 MXFP8_7 MXFP8_8
MXFP8 MAE_Similarity SSIM_Similarity Time (sec)
No.1 73.1 % 27.6 % 116.5
No.2 92.7 % 86.5 % 112.8
No.3 91.0 % 82.8 % 112.9
No.4 92.0 % 79.8 % 112.8
No.5 92.9 % 87.8 % 112.8
No.6 92.7 % 90.0 % 112.9
No.7 91.9 % 83.8 % 112.8
median 92.0 % 83.8 % 112.8

FP8_scaled

FP8_scaled_1 FP8_scaled_2 FP8_scaled_3 FP8_scaled_4 FP8_scaled_6 FP8_scaled_7 FP8_scaled_8
FP8_scaled MAE_Similarity SSIM_Similarity Time (sec)
No.1 72.9 % 28.7 % 76.5
No.2 92.6 % 87.3 % 73.6
No.3 91.6 % 86.6 % 73.5
No.4 92.0 % 79.8 % 73.5
No.5 90.7 % 79.9 % 73.6
No.6 90.5 % 85.5 % 73.6
No.7 91.8 % 84.8 % 73.5
median 91.6 % 84.8 % 73.6

FP8

FP8_1 FP8_2 FP8_3 FP8_4 FP8_6 FP8_7 FP8_8
FP8 MAE_Similarity SSIM_Similarity Time (sec)
No.1 90.2 % 86.0 % 94.7
No.2 80.7 % 40.8 % 92.0
No.3 86.0 % 67.8 % 92.0
No.4 85.8 % 59.1 % 92.0
No.5 80.4 % 37.3 % 92.0
No.6 80.1 % 47.9 % 92.1
No.7 86.4 % 64.2 % 92.0
median 85.8 % 59.1 % 92.0

NVFP4

NVFP4_1 NVFP4_2 NVFP4_3 NVFP4_4 NVFP4_6 NVFP4_7 NVFP4_8
NVFP4 MAE_Similarity SSIM_Similarity Time (sec)
No.1 90.2 % 86.0 % 93.4
No.2 87.2 % 65.1 % 90.7
No.3 86.4 % 72.4 % 90.8
No.4 89.7 % 71.9 % 90.8
No.5 85.1 % 58.7 % 90.7
No.6 89.3 % 84.4 % 90.7
No.7 85.0 % 54.3 % 90.7
median 87.2 % 71.9 % 90.7

Int8_ConvRot_HQ

INT8_ConvRot_HQ_1 INT8_ConvRot_HQ_2 INT8_ConvRot_HQ_3 INT8_ConvRot_HQ_4 INT8_ConvRot_HQ_6 INT8_ConvRot_HQ_7 INT8_ConvRot_HQ_8
Int8_ConvRot_HQ MAE_Similarity SSIM_Similarity Time (sec)
No.1 93.3 % 91.2 % 57.1
No.2 91.0 % 81.9 % 53.8
No.3 91.7 % 89.0 % 53.8
No.4 93.2 % 87.0 % 53.9
No.5 92.2 % 85.6 % 53.9
No.6 93.1 % 88.6 % 53.9
No.7 92.6 % 91.0 % 53.9
median 92.6 % 88.6 % 53.9

Graph

Model Quantization: Image Quality Comparison

vs_FP32 MAE_Similarity SSIM_Similarity Size
FP32 100% 100% 35.2 GB
FP16 92.7 % 87.6 % 17.6 GB
BF16 92.6 % 87.5 % 17.6 GB
MXFP8 92.0 % 83.8 % 10.3 GB
FP8_scaled 91.6 % 84.8 % 10.1 GB
FP8 85.8 % 59.1 % 8.8 GB
NVFP4 87.2 % 71.9 % 6.8 GB
Int8_ConvRot_HQ 92.6 % 88.6 % 10.6 GB

Model Quantization: Generation Speed Comparison

Generation_Time Time (sec) Size
FP32 371.2 35.2 GB
FP16 98.6 17.6 GB
BF16 96.0 17.6 GB
MXFP8 112.8 10.3 GB
FP8_scaled 73.6 10.1 GB
FP8 92.0 8.8 GB
NVFP4 90.7 6.8 GB
Int8_ConvRot_HQ 53.9 10.6 GB

Benchmark Workflow

Screenshot of the ComfyUI workflow used in this verification

About MAE and SSIM

Article