AYS + Ella = anyone else getting superdark image generations?

[Duemellon]

When I combine the the images came out dark, like barely discernible using EpicRealism 1.5 & some other models. Is this a thing for other people too?

[Duemellon]

I think it's the default step count for AYS that was 10. Changed it 20 & it makes it better.

[Duemellon]

still darker than usual even with more steps.

[JettHu]

I think AYS steps and denoise should be the same as Set ELLA Timesteps node. Do you mind share your workflow?

[Duemellon]

15 step, euler

20 step euler

30 step euler

20 step dpmpp

30 step dpmpp

20 step euler

30 step dpmpp

Maybe it's the model? I tried other models & they trend this way too. No one else seems to be having this complaint so it's probably just something I'm doing.

ELLA-AYS.json

[JettHu]

I noticed that add_noise is turned off in your workflow.

You can try turning it on.

steps	add_noise on	add_noise off
20
10

However, I did find that the quality of the generated images will be higher when the steps are larger.

[JettHu]

I think AYS steps and denoise should be the same as Set ELLA Timesteps node. Do you mind share your workflow?

Sorry, what I said before is wrong. Once the sigmas is given to Set ELLA Timesteps node, it's other parameters (scheduler, steps, and denoise) will not take effect.

[JettHu]

In addition, it may be necessary to lower the cfg

The results in the paper used DPM-Solver++(2M) with 10 steps, so I used the same parameters to generate more images:

KSamplerSelect sampler_name: dpmpp_2m; AlignYourStepsScheduler model_type: SD1, steps: 10, denoise: 1.0

all result with same noise_seed

cfg	3	4	5	6	7	8

[Duemellon]

add noise, CFG, & step count helped