Skip to content

Latest commit

 

History

History
 
 

bench

README.md

🤖Benchmarks

🔥Hightlight

Comparisons between different FnBn compute block configurations show that more compute blocks result in higher precision. For example, the F8B0_W8MC0 configuration achieves the best Clip Score (33.007) and ImageReward (1.0333). The meaning of parameter configuration is as follows (such as F8B0_W8M0MC0_T0O1_R0.08): (Device: NVIDIA L20.)

  • F: Fn_compute_blocks, int, Specifies that DBCache uses thefirst nTransformer blocks to fit the information at time step t, enabling the calculation of a more stable L1 difference and delivering more accurate information to subsequent blocks.
  • B: Bn_compute_blocks, int, Further fuses approximate information in thelast nTransformer blocks to enhance prediction accuracy. These blocks act as an auto-scaler for approximate hidden states that use residual cache.
  • W: max_warmup_steps, int, DBCache does not apply the caching strategy when the number of running steps is less than or equal to this value, ensuring the model sufficiently learns basic features during warmup.
  • I: warmup_interval, int, defaults to 1, Skip interval in warmup steps, e.g., when warmup_interval is 2, only 0, 2, 4, ... steps in warmup steps will be computed, others will use dynamic cache.
  • M: max_cached_steps, int, DBCache disables the caching strategy when the previous cached steps exceed this value to prevent precision degradation.
  • MC: max_continuous_cached_steps, int, DBCache disables the caching strategy when the previous continuous cached steps exceed this value to prevent precision degradation. (namely, hybrid dynamic cache and static cache)
  • T: enable talyorseer or not (namely, hybrid taylorseer w/ dynamic cache - DBCache). DBCache acts as the Indicator to decide when to cache, while the Calibrator decides how to cache.
  • O: The taylorseer order, int, e.g., O1 means order 1.
  • R: The residual diff threshold of DBCache, range [0, 1.0)
  • Latency(s): Recorded compute time (eager mode) that w/o other optimizations
  • TFLOPs: Recorded compute FLOPs using calflops's calculate_flops API.

Note

Among all the accuracy indicators, the overall accuracy has slightly improved after using TaylorSeer.

📚Text2Image DrawBench: FLUX.1-dev

Config Clip Score(↑) ImageReward(↑) PSNR(↑) TFLOPs(↓) SpeedUp(↑)
[FLUX.1-dev]: 50 steps 32.9217 1.0412 INF 3726.87 1.00x
F8B0_W4MC0_T0O1_R0.08 32.9871 1.0370 33.8317 2064.81 1.80x
F8B0_W4MC2_T0O1_R0.12 32.9535 1.0185 32.7346 1935.73 1.93x
F8B0_W4MC3_T0O1_R0.12 32.9234 1.0085 32.5385 1816.58 2.05x
F4B0_W4MC3_T0O1_R0.12 32.8981 1.0130 31.8031 1507.83 2.47x
F4B0_W4MC4_T0O1_R0.12 32.8384 1.0065 31.5292 1400.08 2.66x

The comparison between cache-dit: DBCache and algorithms such as Δ-DiT, Chipmunk, FORA, DuCa, TaylorSeer and FoCa is as follows. Now, in the comparison with a speedup ratio less than 3x, cache-dit achieved the best accuracy. Please check 📚How to Reproduce? for more details.

Method TFLOPs(↓) SpeedUp(↑) ImageReward(↑) Clip Score(↑)
[FLUX.1-dev]: 50 steps 3726.87 1.00× 0.9898 32.404
[FLUX.1-dev]: 60% steps 2231.70 1.67× 0.9663 32.312
Δ-DiT(N=2) 2480.01 1.50× 0.9444 32.273
Δ-DiT(N=3) 1686.76 2.21× 0.8721 32.102
[FLUX.1-dev]: 34% steps 1264.63 3.13× 0.9453 32.114
Chipmunk 1505.87 2.47× 0.9936 32.776
FORA(N=3) 1320.07 2.82× 0.9776 32.266
DBCache(S) 1400.08 2.66× 1.0065 32.838
DuCa(N=5) 978.76 3.80× 0.9955 32.241
TaylorSeer(N=4,O=2) 1042.27 3.57× 0.9857 32.413
DBCache(S)+TS 1153.05 3.23× 1.0221 32.819
DBCache(M) 944.75 3.94× 0.9997 32.849
DBCache(M)+TS 944.75 3.94× 1.0107 32.865
FoCa(N=5): arxiv.2508.16211 893.54 4.16× 1.0029 32.948
[FLUX.1-dev]: 22% steps 818.29 4.55× 0.8183 31.772
FORA(N=7) 670.14 5.55× 0.7418 31.519
ToCa(N=12) 644.70 5.77× 0.7155 31.808
DuCa(N=10) 606.91 6.13× 0.8382 31.759
TeaCache(l=1.2) 669.27 5.56× 0.7394 31.704
TaylorSeer(N=7,O=2) 670.44 5.54× 0.9128 32.128
DBCache(F) 651.90 5.72x 0.9271 32.552
FoCa(N=8): arxiv.2508.16211 596.07 6.24× 0.9502 32.706
DBCache(F)+TS 651.90 5.72x 0.9526 32.568
DBCache(U)+TS 505.47 7.37x 0.8645 32.719

NOTE: Except for DBCache, other performance data are referenced from the paper FoCa, arxiv.2508.16211. The configurations of DBCache are listed as belows:

Algo Configuration Algo Configuration
DBCache(S:Slow) F=4,B=0,W=4,I=1,MC=4 DBCache(S:Slow)+TS:TaylorSeer F=1,B=0,W=4,I=1,MC=4,O=1
DBCache(M:Medium) F=1,B=0,W=4,I=1,MC=6 DBCache(M:Medium)+TS:TaylorSeer F=1,B=0,W=4,I=1,MC=6,O=1
DBCache(F:Fast) F=1,B=0,W=8,I=2,MC=8 DBCache(F:Fast)+TS:TaylorSeer F=1,B=0,W=8,I=2,MC=8,O=1
DBCache(U:Ultra) F=1,B=0,W=8,I=4,MC=10, DBCache(U:Ultra)+TS:TaylorSeer F=1,B=0,W=8,I=4,MC=10,O=1

📚Text2Image Distillation DrawBench: Qwen-Image-Lightning

Surprisingly, cache-dit: DBCache still works in the extremely few-step distill model. For example, Qwen-Image-Lightning w/ 4 steps, with the F16B16 configuration, the PSNR is 34.8163, the Clip Score is 35.6109, and the ImageReward is 1.2614. It maintained a relatively high precision.

Config PSNR(↑) Clip Score(↑) ImageReward(↑) TFLOPs(↓) SpeedUp(↑)
[Lightning]: 4 steps INF 35.5797 1.2630 274.33 1.00x
F24B24_W2MC1_T0O1_R0.8 36.3242 35.6224 1.2630 264.74 1.04x
F16B16_W2MC1_T0O1_R0.8 34.8163 35.6109 1.2614 244.25 1.12x
F12B12_W2MC1_T0O1_R0.8 33.8953 35.6535 1.2549 234.63 1.17x
F8B8_W2MC1_T0O1_R0.8 33.1374 35.7284 1.2517 224.29 1.22x
F1B0_W2MC1_T0O1_R0.8 31.8317 35.6651 1.2397 206.90 1.33x

📚How to Reproduce?

⚙️Installation

# install requirements
pip3 install git+https://github.com/openai/CLIP.git
pip3 install git+https://github.com/chengzegang/calculate-flops.pytorch.git
pip3 install image-reward
pip3 install git+https://github.com/vipshop/cache-dit.git

📖Download

git clone https://github.com/vipshop/cache-dit.git
cd cache-dit/bench && mkdir tmp && mkdir log && mkdir hf_models && cd hf_models

# FLUX.1-dev
modelscope download black-forest-labs/FLUX.1-dev --local_dir ./FLUX.1-dev
hf download black-forest-labs/FLUX.1-dev --local-dir ./FLUX.1-dev
export FLUX_DIR=$PWD/FLUX.1-dev

# Qwen-Image-Lightning
modelscope download Qwen/Qwen-Image --local_dir ./Qwen-Image
modelscope download lightx2v/Qwen-Image-Lightning --local_dir ./Qwen-Image-Lightning
hf download Qwen/Qwen-Image --local-dir ./Qwen-Image
hf download lightx2v/Qwen-Image-Lightning --local-dir ./Qwen-Image-Lightning
export QWEN_IMAGE_DIR=$PWD/Qwen-Image
export QWEN_IMAGE_LIGHT_DIR=$PWD/Qwen-Image-Lightning

# Clip Score & Image Reward
modelscope download laion/CLIP-ViT-g-14-laion2B-s12B-b42K --local_dir ./CLIP-ViT-g-14-laion2B-s12B-b42K
modelscope download ZhipuAI/ImageReward --local_dir ./ImageReward
hf download laion/CLIP-ViT-g-14-laion2B-s12B-b42K --local-dir ./CLIP-ViT-g-14-laion2B-s12B-b42K
hf download ZhipuAI/ImageReward --local-dir ./ImageReward
export CLIP_MODEL_DIR=$PWD/CLIP-ViT-g-14-laion2B-s12B-b42K
export IMAGEREWARD_MODEL_DIR=$PWD/ImageReward

cd ..

📖Evaluation

# NOTE: The reported benchmark was run on NVIDIA L20 device.

# FLUX.1-dev DrawBench w/ low speedup ratio
export CUDA_VISIBLE_DEVICES=0
nohup bash bench.sh default > log/cache_dit_bench_default.log 2>&1 &
export CUDA_VISIBLE_DEVICES=1
nohup bash bench.sh taylorseer > log/cache_dit_bench_taylorseer.log 2>&1 &
bash ./metrics.sh

# FLUX.1-dev DrawBench w/ high speedup ratio
export CUDA_VISIBLE_DEVICES=0
nohup bash bench_fast.sh default > log/cache_dit_bench_fast.log 2>&1 &
export CUDA_VISIBLE_DEVICES=1
nohup bash bench_fast.sh taylorseer > log/cache_dit_bench_taylorseer_fast.log 2>&1 &
bash ./metrics_fast.sh

# Qwen-Image-Lightning DrawBench
export CUDA_VISIBLE_DEVICES=0,1
nohup bash bench_distill.sh 8_steps > log/cache_dit_bench_distill_8_steps.log 2>&1 &
export CUDA_VISIBLE_DEVICES=2,3
nohup bash bench_distill.sh 4_steps > log/cache_dit_bench_distill_4_steps.log 2>&1 &
bash ./metrics_distill.sh

📚Detailed Benchmark

📖ClipScore: DBCache(Slow) FnBn

Config CLIP_SCORE Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] 32.9217 42.74 1.00 3726.87 1.00
F1B0_W4M0MC0_R0.08_T0O0 33.0745 23.21 1.84 1729.60 2.15
F8B0_W4M0MC3_R0.32_T0O0 33.0055 22.07 1.94 1548.80 2.41
F8B0_W4M0MC4_R0.24_T0O0 32.9921 20.79 2.06 1421.96 2.62
F8B0_W4M0MC6_R0.32_T0O0 32.9804 19.38 2.21 1262.77 2.95
F4B0_W4M0MC4_R0.24_T0O0 32.9720 18.77 2.28 1231.01 3.03
F4B0_W4M0MC4_R0.32_T0O0 32.9680 18.10 2.36 1163.10 3.20
F4B0_W4M0MC6_R0.32_T0O0 32.8926 16.62 2.57 1024.52 3.64
F1B0_W4M0MC4_R0.32_T0O0 32.8640 16.30 2.62 1017.97 3.66
F1B0_W4M0MC6_R0.24_T0O0 32.8496 15.54 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T0O0 32.8098 15.65 2.73 942.92 3.95

📖ClipScore: DBCache(Slow) FnBn + TaylorSeer O(1)

Config CLIP_SCORE Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] 32.9217 42.92 1.00 3726.87 1.00
F1B0_W4M0MC0_R0.08_T1O1 33.0398 23.29 1.84 1730.70 2.15
F4B0_W4M0MC3_R0.12_T1O1 32.9795 21.36 2.01 1499.51 2.49
F1B0_W4M0MC6_R0.32_T1O1 32.9589 15.70 2.73 944.02 3.95

📖ClipScore: DBCache(Fast) FnBn

Config CLIP_SCORE Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
C0_Q0_NONE 32.9616 44.61 1.00 3726.87 1.00
F1B0_W16I4M0MC10_R0.8_T0O0 32.6306 10.48 4.26 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T0O0 32.5522 11.26 3.96 651.90 5.72
F1B0_W8I4M0MC8_R0.8_T0O0 32.5420 10.46 4.26 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T0O0 32.4479 9.71 4.59 505.47 7.37
F1B0_W8I2M0MC8_R0.8_T0O0 32.4395 11.40 3.91 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T0O0 32.4207 10.62 4.20 578.69 6.44

📖ClipScore: DBCache(Fast) FnBn + TaylorSeer O(1)

Config CLIP_SCORE Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
C0_Q0_NONE 32.9616 44.94 1.00 3726.87 1.00
F1B0_W16I4M0MC8_R0.8_T1O1 33.0988 11.42 3.94 651.90 5.72
F1B0_W16I4M0MC10_R0.8_T1O1 32.9764 10.65 4.22 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T1O1 32.7195 9.83 4.57 505.47 7.37
F1B0_W8I4M0MC8_R0.8_T1O1 32.6448 10.56 4.26 578.69 6.44
F1B0_W8I2M0MC8_R0.8_T1O1 32.5682 11.44 3.93 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T1O1 32.3999 10.66 4.22 578.69 6.44

📖ImageReward: DBCache(Slow) FnBn

Config IMAGE_REWARD Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] 1.0412 42.74 1.00 3726.87 1.00
F1B0_W4M0MC2_R0.16_T0O0 1.0461 21.29 2.01 1530.46 2.44
F1B0_W4M0MC2_R0.24_T0O0 1.0418 20.51 2.08 1457.25 2.56
F1B0_W4M0MC2_R0.2_T0O0 1.0418 20.53 2.08 1457.25 2.56
F1B0_W4M0MC2_R0.32_T0O0 1.0418 20.63 2.07 1457.25 2.56
F1B0_W4M0MC3_R0.16_T0O0 1.0360 19.09 2.24 1310.82 2.84
F1B0_W4M0MC3_R0.2_T0O0 1.0323 18.57 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.16_T0O0 1.0232 17.75 2.41 1164.76 3.20
F1B0_W4M0MC4_R0.2_T0O0 1.0229 17.71 2.41 1153.05 3.23
F1B0_W4M0MC4_R0.24_T0O0 1.0096 16.98 2.52 1091.18 3.42
F1B0_W4M0MC4_R0.32_T0O0 1.0083 16.30 2.62 1017.97 3.66
F1B0_W4M0MC6_R0.24_T0O0 0.9997 15.54 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T0O0 0.9921 15.65 2.73 942.92 3.95

📖ImageReward: DBCache(Slow) FnBn + TaylorSeer O(1)

Config IMAGE_REWARD Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] 1.0412 42.92 1.00 3726.87 1.00
F1B0_W4M0MC0_R0.08_T1O1 1.0591 23.29 1.84 1730.70 2.15
F1B0_W4M0MC2_R0.16_T1O1 1.0389 21.35 2.01 1530.46 2.44
F1B0_W4M0MC2_R0.24_T1O1 1.0379 20.61 2.08 1457.25 2.56
F1B0_W4M0MC2_R0.2_T1O1 1.0379 20.59 2.08 1457.25 2.56
F1B0_W4M0MC2_R0.32_T1O1 1.0379 20.70 2.07 1457.25 2.56
F1B0_W4M0MC3_R0.12_T1O1 1.0292 20.70 2.07 1457.62 2.56
F1B0_W4M0MC3_R0.2_T1O1 1.0257 18.63 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.2_T1O1 1.0221 17.71 2.42 1153.05 3.23
F1B0_W4M0MC4_R0.32_T1O1 1.0149 16.33 2.63 1017.97 3.66
F1B0_W4M0MC6_R0.24_T1O1 1.0107 15.61 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T1O1 1.0025 15.70 2.73 944.02 3.95

📖ImageReward: DBCache(Fast) FnBn

Config IMAGE_REWARD Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
C0_Q0_NONE 1.0445 44.61 1.00 3726.87 1.00
F1B0_W8I2M0MC8_R0.8_T0O0 0.9271 11.40 3.91 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T0O0 0.9234 10.62 4.20 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T0O0 0.9115 11.26 3.96 651.90 5.72
F1B0_W16I4M0MC10_R0.8_T0O0 0.8644 10.48 4.26 578.69 6.44
F1B0_W8I4M0MC8_R0.8_T0O0 0.8301 10.46 4.26 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T0O0 0.8092 9.71 4.59 505.47 7.37

📖ImageReward: DBCache(Fast) FnBn + TaylorSeer O(1)

Config IMAGE_REWARD Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
C0_Q0_NONE 1.0445 44.94 1.00 3726.87 1.00
F1B0_W8I2M0MC8_R0.8_T1O1 0.9526 11.44 3.93 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T1O1 0.9360 10.66 4.22 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T1O1 0.9285 11.42 3.94 651.90 5.72
F1B0_W8I4M0MC8_R0.8_T1O1 0.9088 10.56 4.26 578.69 6.44
F1B0_W16I4M0MC10_R0.8_T1O1 0.8941 10.65 4.22 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T1O1 0.8645 9.83 4.57 505.47 7.37

📖PSNR: DBCache(Slow) FnBn

Config PSNR Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1.dev, 50 steps] - 42.74 1.00 3726.87 1.00
F8B0_W8M0MC0_R0.08_T0O0 35.2008 27.87 1.53 2162.19 1.72
F8B0_W8M0MC2_R0.12_T0O0 34.7449 27.22 1.57 2072.18 1.80
F8B0_W8M0MC2_R0.16_T0O0 34.6659 26.34 1.62 2002.67 1.86
F8B0_W8M0MC2_R0.2_T0O0 34.6579 26.33 1.62 1994.67 1.87
F8B0_W8M0MC2_R0.24_T0O0 34.5399 25.62 1.67 1933.17 1.93
F8B0_W8M0MC2_R0.32_T0O0 34.4860 25.66 1.67 1933.17 1.93
F1B0_W4M0MC0_R0.08_T0O0 33.9639 23.21 1.84 1729.60 2.15
F1B0_W4M0MC2_R0.12_T0O0 33.1898 21.92 1.95 1604.78 2.32
F1B0_W4M0MC3_R0.12_T0O0 33.0037 20.61 2.07 1457.62 2.56
F1B0_W4M0MC4_R0.12_T0O0 32.9462 20.01 2.14 1401.61 2.66
F1B0_W4M0MC3_R0.16_T0O0 31.9664 19.09 2.24 1310.82 2.84
F1B0_W4M0MC3_R0.2_T0O0 31.9315 18.57 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.16_T0O0 31.7174 17.75 2.41 1164.76 3.20
F1B0_W4M0MC4_R0.2_T0O0 31.7109 17.71 2.41 1153.05 3.23
F1B0_W4M0MC4_R0.24_T0O0 31.3081 16.98 2.52 1091.18 3.42
F1B0_W4M0MC6_R0.24_T0O0 31.0524 15.54 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T0O0 30.6637 15.65 2.73 942.92 3.95

📖PSNR: DBCache(Slow) FnBn + TaylorSeer O(1)

Config PSNR Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] - 42.74 1.00 3726.87 1.00
F8B0_W8M0MC0_R0.08_T1O1 36.7296 28.06 1.53 2172.76 1.72
F8B0_W8M0MC2_R0.12_T1O1 36.1337 27.31 1.57 2074.10 1.80
F8B0_W8M0MC2_R0.16_T1O1 36.1062 26.46 1.62 2005.56 1.86
F8B0_W8M0MC2_R0.2_T1O1 36.0838 26.22 1.64 1985.38 1.88
F8B0_W8M0MC2_R0.24_T1O1 35.9808 25.64 1.67 1933.17 1.93
F8B0_W8M0MC2_R0.32_T1O1 35.9008 25.71 1.67 1933.17 1.93
F8B0_W8M0MC4_R0.12_T1O1 35.4352 25.59 1.68 1912.99 1.95
F8B0_W8M0MC3_R0.16_T1O1 34.7351 24.67 1.74 1815.62 2.05
F8B0_W8M0MC3_R0.2_T1O1 34.6739 24.46 1.75 1794.16 2.08
F8B0_W8M0MC3_R0.24_T1O1 34.6091 23.84 1.80 1740.99 2.14
F8B0_W8M0MC3_R0.32_T1O1 34.5803 23.98 1.79 1740.99 2.14
F1B0_W4M0MC0_R0.08_T1O1 34.5305 23.29 1.84 1730.70 2.15
F1B0_W4M0MC2_R0.12_T1O1 34.1546 22.07 1.94 1604.04 2.32
F1B0_W4M0MC4_R0.12_T1O1 33.8600 20.08 2.14 1404.17 2.65
F1B0_W4M0MC3_R0.16_T1O1 31.8628 19.20 2.24 1310.82 2.84
F1B0_W4M0MC3_R0.2_T1O1 31.8328 18.63 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.16_T1O1 31.6688 17.82 2.41 1164.76 3.20
F1B0_W4M0MC4_R0.2_T1O1 31.6624 17.71 2.42 1153.05 3.23
F1B0_W4M0MC4_R0.24_T1O1 30.8172 17.04 2.52 1091.18 3.42
F1B0_W4M0MC6_R0.24_T1O1 30.5131 15.61 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T1O1 29.8860 15.70 2.73 944.02 3.95

📖PSNR: DBCache(Fast) FnBn

Config PSNR Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
F1B0_W8I2M0MC8_R0.8_T0O0 29.3428 11.40 3.91 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T0O0 29.2732 10.62 4.20 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T0O0 28.5724 11.26 3.96 651.90 5.72
F1B0_W16I4M0MC10_R0.8_T0O0 28.5644 10.48 4.26 578.69 6.44
F1B0_W8I4M0MC8_R0.8_T0O0 28.5440 10.46 4.26 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T0O0 28.5201 9.71 4.59 505.47 7.37

📖PSNR: DBCache(Fast) FnBn + TaylorSeer O(1)

Config PSNR Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
F1B0_W8I2M0MC8_R0.8_T1O1 29.2251 11.44 3.93 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T1O1 28.9526 10.66 4.22 578.69 6.44
F1B0_W8I4M0MC8_R0.8_T1O1 28.7110 10.56 4.26 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T1O1 28.6300 11.42 3.94 651.90 5.72
F1B0_W16I4M0MC10_R0.8_T1O1 28.5842 10.65 4.22 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T1O1 28.5654 9.83 4.57 505.47 7.37

📖SSIM: DBCache(Slow) FnBn

Config SSIM Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] - 42.74 1.00 3726.87 1.00
F8B0_W8M0MC0_R0.08_T0O0 0.9131 27.87 1.53 2162.19 1.72
F8B0_W8M0MC2_R0.12_T0O0 0.9017 27.22 1.57 2072.18 1.80
F8B0_W8M0MC2_R0.16_T0O0 0.8999 26.34 1.62 2002.67 1.86
F8B0_W8M0MC2_R0.2_T0O0 0.8996 26.33 1.62 1994.67 1.87
F8B0_W8M0MC2_R0.24_T0O0 0.8953 25.62 1.67 1933.17 1.93
F8B0_W8M0MC2_R0.32_T0O0 0.8936 25.66 1.67 1933.17 1.93
F8B0_W8M0MC4_R0.12_T0O0 0.8858 25.41 1.68 1897.61 1.96
F8B0_W8M0MC3_R0.16_T0O0 0.8774 24.55 1.74 1811.77 2.06
F8B0_W8M0MC3_R0.2_T0O0 0.8761 24.56 1.74 1803.45 2.07
F1B0_W4M0MC0_R0.08_T0O0 0.8727 23.21 1.84 1729.60 2.15
F8B0_W8M0MC4_R0.2_T0O0 0.8571 23.40 1.83 1678.21 2.22
F1B0_W4M0MC2_R0.12_T0O0 0.8536 21.92 1.95 1604.78 2.32
F1B0_W4M0MC3_R0.12_T0O0 0.8521 20.61 2.07 1457.62 2.56
F1B0_W4M0MC4_R0.12_T0O0 0.8461 20.01 2.14 1401.61 2.66
F1B0_W4M0MC3_R0.16_T0O0 0.8092 19.09 2.24 1310.82 2.84
F1B0_W4M0MC3_R0.2_T0O0 0.8053 18.57 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.16_T0O0 0.7964 17.75 2.41 1164.76 3.20
F1B0_W4M0MC4_R0.2_T0O0 0.7952 17.71 2.41 1153.05 3.23
F1B0_W4M0MC6_R0.24_T0O0 0.7620 15.54 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T0O0 0.7438 15.65 2.73 942.92 3.95

📖SSIM: DBCache(Slow) FnBn + TaylorSeer O(1)

Config SSIM Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] - 42.74 1.00 3726.87 1.00
F8B0_W8M0MC0_R0.08_T1O1 0.9340 28.06 1.53 2172.76 1.72
F8B0_W8M0MC2_R0.12_T1O1 0.9201 27.31 1.57 2074.10 1.80
F8B0_W8M0MC2_R0.16_T1O1 0.9194 26.46 1.62 2005.56 1.86
F8B0_W8M0MC2_R0.2_T1O1 0.9189 26.22 1.64 1985.38 1.88
F8B0_W8M0MC2_R0.24_T1O1 0.9169 25.64 1.67 1933.17 1.93
F8B0_W8M0MC2_R0.32_T1O1 0.9157 25.71 1.67 1933.17 1.93
F8B0_W8M0MC3_R0.12_T1O1 0.9116 26.09 1.65 1974.49 1.89
F8B0_W8M0MC4_R0.12_T1O1 0.9098 25.59 1.68 1912.99 1.95
F8B0_W8M0MC3_R0.16_T1O1 0.8953 24.67 1.74 1815.62 2.05
F8B0_W8M0MC3_R0.2_T1O1 0.8926 24.46 1.75 1794.16 2.08
F8B0_W8M0MC3_R0.24_T1O1 0.8907 23.84 1.80 1740.99 2.14
F8B0_W8M0MC3_R0.32_T1O1 0.8905 23.98 1.79 1740.99 2.14
F1B0_W4M0MC0_R0.08_T1O1 0.8845 23.29 1.84 1730.70 2.15
F1B0_W4M0MC2_R0.12_T1O1 0.8802 22.07 1.94 1604.04 2.32
F1B0_W4M0MC3_R0.12_T1O1 0.8752 20.70 2.07 1457.62 2.56
F1B0_W4M0MC4_R0.12_T1O1 0.8745 20.08 2.14 1404.17 2.65
F1B0_W4M0MC3_R0.16_T1O1 0.8222 19.20 2.24 1310.82 2.84
F1B0_W4M0MC3_R0.2_T1O1 0.8198 18.63 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.16_T1O1 0.8125 17.82 2.41 1164.76 3.20
F1B0_W4M0MC4_R0.2_T1O1 0.8119 17.71 2.42 1153.05 3.23
F1B0_W4M0MC4_R0.24_T1O1 0.7837 17.04 2.52 1091.18 3.42
F1B0_W4M0MC6_R0.24_T1O1 0.7718 15.61 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T1O1 0.7423 15.70 2.73 944.02 3.95

📖SSIM: DBCache(Fast) FnBn

Config SSIM Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
F1B0_W8I2M0MC8_R0.8_T0O0 0.6681 11.40 3.91 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T0O0 0.6528 10.62 4.20 578.69 6.44
F1B0_W8I4M0MC8_R0.8_T0O0 0.6161 10.46 4.26 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T0O0 0.6066 11.26 3.96 651.90 5.72
F1B0_W8I4M0MC10_R0.8_T0O0 0.6065 9.71 4.59 505.47 7.37
F1B0_W16I4M0MC10_R0.8_T0O0 0.5977 10.48 4.26 578.69 6.44

📖SSIM: DBCache(Fast) FnBn + TaylorSeer O(1)

Config SSIM Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
F1B0_W8I2M0MC8_R0.8_T1O1 0.6653 11.44 3.93 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T1O1 0.6219 10.66 4.22 578.69 6.44
F1B0_W8I4M0MC8_R0.8_T1O1 0.6194 10.56 4.26 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T1O1 0.6076 11.42 3.94 651.90 5.72
F1B0_W8I4M0MC10_R0.8_T1O1 0.5947 9.83 4.57 505.47 7.37
F1B0_W16I4M0MC10_R0.8_T1O1 0.5944 10.65 4.22 578.69 6.44

📖LPIPS: DBCache(Slow) FnBn

Config LPIPS Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] - 42.74 1.00 3726.87 1.00
F8B0_W8M0MC0_R0.08_T0O0 0.0786 27.87 1.53 2162.19 1.72
F8B0_W8M0MC2_R0.12_T0O0 0.0895 27.22 1.57 2072.18 1.80
F8B0_W8M0MC2_R0.16_T0O0 0.0919 26.34 1.62 2002.67 1.86
F8B0_W8M0MC2_R0.2_T0O0 0.0923 26.33 1.62 1994.67 1.87
F8B0_W8M0MC2_R0.24_T0O0 0.0963 25.62 1.67 1933.17 1.93
F8B0_W8M0MC2_R0.32_T0O0 0.0990 25.66 1.67 1933.17 1.93
F8B0_W8M0MC4_R0.12_T0O0 0.1085 25.41 1.68 1897.61 1.96
F8B0_W8M0MC3_R0.16_T0O0 0.1182 24.55 1.74 1811.77 2.06
F1B0_W4M0MC0_R0.08_T0O0 0.1196 23.21 1.84 1729.60 2.15
F1B0_W4M0MC2_R0.12_T0O0 0.1390 21.92 1.95 1604.78 2.32
F1B0_W4M0MC3_R0.12_T0O0 0.1426 20.61 2.07 1457.62 2.56
F1B0_W4M0MC4_R0.12_T0O0 0.1500 20.01 2.14 1401.61 2.66
F1B0_W4M0MC3_R0.16_T0O0 0.1909 19.09 2.24 1310.82 2.84
F1B0_W4M0MC3_R0.2_T0O0 0.1942 18.57 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.16_T0O0 0.2082 17.75 2.41 1164.76 3.20
F1B0_W4M0MC4_R0.2_T0O0 0.2094 17.71 2.41 1153.05 3.23
F1B0_W4M0MC4_R0.24_T0O0 0.2289 16.98 2.52 1091.18 3.42
F1B0_W4M0MC6_R0.24_T0O0 0.2553 15.54 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T0O0 0.2826 15.65 2.73 942.92 3.95

📖LPIPS: DBCache(Slow) FnBn + TaylorSeer O(1)

Config LPIPS Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
Baseline: [FLUX.1 50 steps] - 42.74 1.00 3726.87 1.00
F8B0_W8M0MC0_R0.08_T1O1 0.0558 28.06 1.53 2172.76 1.72
F8B0_W8M0MC2_R0.12_T1O1 0.0687 27.31 1.57 2074.10 1.80
F8B0_W8M0MC2_R0.16_T1O1 0.0690 26.46 1.62 2005.56 1.86
F8B0_W8M0MC2_R0.2_T1O1 0.0692 26.22 1.64 1985.38 1.88
F8B0_W8M0MC2_R0.24_T1O1 0.0699 25.64 1.67 1933.17 1.93
F8B0_W8M0MC2_R0.32_T1O1 0.0714 25.71 1.67 1933.17 1.93
F8B0_W8M0MC3_R0.12_T1O1 0.0772 26.09 1.65 1974.49 1.89
F8B0_W8M0MC4_R0.12_T1O1 0.0792 25.59 1.68 1912.99 1.95
F8B0_W8M0MC3_R0.16_T1O1 0.0933 24.67 1.74 1815.62 2.05
F8B0_W8M0MC3_R0.2_T1O1 0.0958 24.46 1.75 1794.16 2.08
F8B0_W8M0MC3_R0.24_T1O1 0.0964 23.84 1.80 1740.99 2.14
F8B0_W8M0MC3_R0.32_T1O1 0.0974 23.98 1.79 1740.99 2.14
F1B0_W4M0MC0_R0.08_T1O1 0.1032 23.29 1.84 1730.70 2.15
F1B0_W4M0MC2_R0.12_T1O1 0.1091 22.07 1.94 1604.04 2.32
F1B0_W4M0MC3_R0.12_T1O1 0.1147 20.70 2.07 1457.62 2.56
F1B0_W4M0MC4_R0.12_T1O1 0.1154 20.08 2.14 1404.17 2.65
F1B0_W4M0MC3_R0.2_T1O1 0.1733 18.63 2.30 1237.61 3.01
F1B0_W4M0MC4_R0.16_T1O1 0.1837 17.82 2.41 1164.76 3.20
F1B0_W4M0MC4_R0.2_T1O1 0.1839 17.71 2.42 1153.05 3.23
F1B0_W4M0MC4_R0.24_T1O1 0.2155 17.04 2.52 1091.18 3.42
F1B0_W4M0MC6_R0.24_T1O1 0.2310 15.61 2.75 944.75 3.94
F1B0_W4M0MC6_R0.32_T1O1 0.2677 15.70 2.73 944.02 3.95

📖LPIPS: DBCache(Fast) FnBn

Config LPIPS Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
F1B0_W8I2M0MC8_R0.8_T0O0 0.3899 11.40 3.91 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T0O0 0.4180 10.62 4.20 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T0O0 0.4703 11.26 3.96 651.90 5.72
F1B0_W8I4M0MC8_R0.8_T0O0 0.4829 10.46 4.26 578.69 6.44
F1B0_W16I4M0MC10_R0.8_T0O0 0.4850 10.48 4.26 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T0O0 0.5024 9.71 4.59 505.47 7.37

📖LPIPS: DBCache(Fast) FnBn + TaylorSeer O(1)

Config LPIPS Latency(s) SpeedUp(↑) TFLOPs SpeedUp(↑)
F1B0_W8I2M0MC8_R0.8_T1O1 0.3620 11.44 3.93 651.90 5.72
F1B0_W8I2M0MC10_R0.8_T1O1 0.4248 10.66 4.22 578.69 6.44
F1B0_W8I4M0MC8_R0.8_T1O1 0.4463 10.56 4.26 578.69 6.44
F1B0_W16I4M0MC8_R0.8_T1O1 0.4478 11.42 3.94 651.90 5.72
F1B0_W16I4M0MC10_R0.8_T1O1 0.4662 10.65 4.22 578.69 6.44
F1B0_W8I4M0MC10_R0.8_T1O1 0.4855 9.83 4.57 505.47 7.37