[NeurIPS24] FlowDCN: Exploring DCN-like Architectures for Fast Image Generation with Arbitrary Resolution

[NEWS] [9.26] 💐💐 Our FlowDCN is accepted by NeurIPS 2024! 💐💐

[NEWS] [11.22] 🍺 Our FlowDCN models and code are now available in the official repo!

Pretrained Models

Our Models consistently achieve state-of-the-art results on the sFID metrics compared to SiT/DiT.

Metrics

Our Models consistently has fewer parameters and GFLOPS compared to Transformer counterparts. Our code also support LogNorm and VAR(Various Aspect Ratio Training)

Model-iters	Resolution	Solver	NFE-CFG	FID	sFID	Params	Link
FlowDCN-S-400k	256x256	EulerSDE-250	250x2	54.6	8.8	30.3M	HF
FlowDCN-B-400k	256x256	EulerSDE-250	250x2	28.5	6.09	120M	HF
VAR-FlowDCN-B-400k	256x256	EulerSDE-250	250x2	23.6	7.72	120M	HF
FlowDCN-L-400k	256x256	EulerSDE-250	250x2	13.8	4.69	421M	HF
FlowDCN-XL-2M	256x256	EulerODE-250	250x2	2.01	4.33	618M	HF
FlowDCN-XL-2M	256x256	EulerSDE-250	250x2	2.00	4.37	618M	HF
FlowDCN-XL-2M	256x256	NeuralSolver-10	10x2	2.35	5.07	618M	HF
FlowDCN-XL-100k	512x512	EulerODE-50	50x2	2.76	5.29	618M	HF
FlowDCN-XL-100k	512x512	EulerSDE-250	250x2	2.44	4.53	618M	HF
FlowDCN-XL-100k	512x512	NeuralSolver-10	10x2	2.77	4.68	618M	HF

Visualizations

CFG1.375 Generation Images:

Models	Resolution	Link
FlowDCN-XL-100k	512x512	HF
FlowDCN-XL-2M	256x256	HF

CFG4.0 selected Generation Images:

Various Resolution Extension

Models	256x256 FID	sFID	IS	320x320 FID	sFID	IS	224x448 FID	sFID	IS	160x480 FID	sFID	IS
DiT-B	44.83	8.49	32.05	95.47	108.68	18.38	109.1	110.71	14.00	143.8	122.81	8.93
with EI	44.83	8.49	32.05	81.48	62.25	20.97	133.2	72.53	11.11	160.4	93.91	7.30
with PI	44.83	8.49	32.05	72.47	54.02	24.15	133.4	70.29	11.73	156.5	93.80	7.80
FiT-B (+VAR)	36.36	11.08	40.69	61.35	30.71	31.01	44.67	24.09	37.1	56.81	22.07	25.25
with VisionYaRN	36.36	11.08	40.69	44.76	38.04	44.70	41.92	42.79	45.87	62.84	44.82	27.84
with VisionNTK	36.36	11.08	40.69	57.31	31.31	33.97	43.84	26.25	39.22	56.76	24.18	26.40
FlowDCN-B	28.5	6.09	51	34.4	27.2	52.2	71.7	62.0	23.7	211	111	5.83
FlowDCN-B (+VAR)	23.6	7.72	62.8	29.1	15.8	69.5	31.4	17.0	62.4	44.7	17.8	35.8

Linear-Multi-step Solvers and NeuralSolvers

We also provide a adams-like linear-multi-step solver for the recitified flow sampling. The related configs are named with adam2 or adam4. The solver code are placed in ./src/diffusion/flow_matching/adam_sampling.py.

Compared to Henu/RK4, the linear-multi-step solver is more stable and faster.

During some experiments, we supringly find that the linear-multi-step solver can achieve comparable results even with FlowTurbo.

As they are distinct methods, so armed with Adams, we believe FlowTurbo can be more powerful.

Also, We provide some magic solvers for the recitified flow sampling. These solvers are highly inspired by linear-multi-steps methods, and consists of just some Magic Numbers These solvers are really powerful and interesting. We place the related code in ./src/diffusion/flow_matching/ns_sampling.py.

SiT-XL-R256	Steps	NFE-CFG	Extra-Paramters	FID	IS	PR	Recall
Heun	8	16x2	0	3.68	/	/	/
Heun	11	22x2	0	2.79	/	/	/
Heun	15	30x2	0	2.42	/	/	/
Adam2	6	6x2	0	6.35	190	0.75	0.55
Adam2	8	8x2	0	4.16	212	0.78	0.56
Adam2	16	16x2	0	2.42	237	0.80	0.60
Adam4	16	16x2	0	2.27	243	0.80	0.60

Citation

@inproceedings{
wang2024exploring,
title={Exploring {DCN}-like architecture for fast image generation with arbitrary resolution},
author={Shuai Wang and Zexian Li and Tianhui Song and Xubin Li and Tiezheng Ge and Bo Zheng and Limin Wang},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=e57B7BfA2B}
}