Worse in high guidance scale

[yuhuUSTC]

Dear authors,
Thank you for sharing this great work and open source!
When following this work, I found that DPM-solver++ performs worse than DDIM under high classifier guidance scale, eg, scale=8 or 16. Concretely, under scale=8, the performance of DPM-solver++ approaches DDIM, and under scale=16, the performance of DDIM > DPM-solver++ > UniPC. Same issues alsp appy to UniPC.

Config: DPM-solver++ 2M, 50k samples

This contradicts the reported results in the paper.
Thank you for your explanation and help.

[LuChengTHU]

Generally speaking, higher guidance scales will make the ODE ill-conditioned and higher-order solvers will become more and more unstable. The instability issue will be somehow eased with DPM++ because the property of data-pred model, but it cannot address it. So in my original paper I also applied dynamic thresholding to further reduce the condition numbers.

[zxk72]

Hey fellows, really good work and thanks for sharing the code.
@LuChengTHU Hello, I applied dpm-solver to guided diffusion, but the results were not satisfactory. Is this the same question you answered above? Would it be better to use SDE-based dpm-solver? Or is there a better solution? Looking forward for your reply, thank you!

[LuChengTHU]

Hi @zxk72 , did you use these commands?:

dpm-solver/examples/ddpm_and_guided-diffusion/sample.sh

Lines 38 to 50 in 52bc3fb

	# ImageNet256 with classifier guidance (large guidance scale) example
	data="imagenet256_guided"
	scale="8.0"
	sampleMethod='dpmsolver++'
	type="dpmsolver"
	steps="20"
	DIS="time_uniform"
	order="2"
	method="multistep"
	workdir="experiments/"$data"/"$sampleMethod"_"$method"_order"$order"_"$steps"_"$DIS"_scale"$scale"_type-"$type"_thresholding"
	CUDA_VISIBLE_DEVICES=$DEVICES python main.py --config $data".yml" --exp=$workdir --sample --fid --timesteps=$steps --eta 0 --ni --skip_type=$DIS --sample_type=$sampleMethod --dpm_solver_order=$order --dpm_solver_method=$method --dpm_solver_type=$type --port 12350 --scale=$scale --thresholding

[zxk72]

@LuChengTHU thank you for your reply! Sorry, I'm not using the example you provided. It is a diffusion medical segmentation project called MedSegDiff, which uses guided diffusion. The results after using dpm-solver have some noise, which is worse than the results without using it.