Quality degradation after multi-inferences

[973398769]

This is my current setup: helper.set_params(cache_interval=3, cache_branch_id=0). When processing a large amount of images, I've noticed some quality degradation, but I'm not sure if it's caused by DeepCache.

[973398769]

I would like to understand when I should enable and when I should disable it. Is it necessary to disable after every inference, and should I enable it before starting a new inference? Super Thanks!

[elismasilva]

I would like to understand when I should enable and when I should disable it. Is it necessary to disable after every inference, and should I enable it before starting a new inference? Super Thanks!

you need disable and reapply in next inference only if you change params. I found this bug too and solved doing this.

[973398769]

Hi Super Thanks! If I only change the seed or prompt, do I also need to reapply it? Thanks 获取 Outlook for iOS<https://aka.ms/o0ukef>

…

________________________________ 发件人: Eliseu Silva ***@***.***> 发送时间: Wednesday, November 20, 2024 7:32:25 AM 收件人: horseee/DeepCache ***@***.***> 抄送: nini_good ***@***.***>; Author ***@***.***> 主题: Re: [horseee/DeepCache] Quality degradation after multi-inferences (Issue #54) I would like to understand when I should enable and when I should disable it. Is it necessary to disable after every inference, and should I enable it before starting a new inference? Super Thanks! you need disable and reapply in next inference only if you change params. I found this bug too and solved doing this. ― Reply to this email directly, view it on GitHub<#54 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AL72YMYZN4O2YAHM25QTRYT2BPDATAVCNFSM6AAAAABR67PXJSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOBWHE3TOMRRGQ>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[elismasilva]

I would like to understand when I should enable and when I should disable it. Is it necessary to disable after every inference, and should I enable it before starting a new inference? Super Thanks!

Sorry, i did a test now, you always need disable and enable again in each inference even if params is not changed.

[973398769]

Could you kindly assist me in understanding how to reproduce this issue? Would it be possible for you to share some sample code that could help in reproducing it?

I would like to understand when I should enable and when I should disable it. Is it necessary to disable after every inference, and should I enable it before starting a new inference? Super Thanks!

Sorry, i did a test now, you always need disable and enable again in each inference even if params is not changed.

[973398769]

I may only encounter the problem when testing with thousands of images. Thank you very much for your help

[elismasilva]

@973398769
just put helper.disable() before other calls.

helper.disable()
helper.set_params(cache_interval=3, cache_branch_id=0)
helper.enable()

this need be before you call pipe()

[973398769]

I would like to reproduce the issue of quality degradation, but I only detect this problem locally when testing with thousands of images. Could you please tell me how to make this issue appear more quickly? Thank you.

[973398769]

This is my test pipeline:
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
use_safetensors=True,
).to("cuda:0")
helper = DeepCacheSDHelper(pipe=pipe)
helper.set_params(
cache_interval=3,
cache_branch_id=0,
)
deepcache_image = pipe(
prompt,
generator=generator,
output_type='pil'
).images[0]

[elismasilva]

This is my test pipeline: pipe = StableDiffusionPipeline.from_pretrained( "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True, ).to("cuda:0") helper = DeepCacheSDHelper(pipe=pipe) helper.set_params( cache_interval=3, cache_branch_id=0, ) deepcache_image = pipe( prompt, generator=generator, output_type='pil' ).images[0]

I don't know if we're talking about the same problem, I'm working with SDXL, but I reuse the pipeline without deleting the variable or restarting the script, because I'm in a Gradio interface. I don't know how you're doing it, if you're running this code inside a loop and reusing the same pipeline variable or if you're destroying the variable. For me, if I order the first inference of an image, it generates it normally, then if I try to generate it again without disabling it first, it already generates the deteriorated image. Like this:

[973398769]

Are you using generate_pipe.enable_model_cpu_offload() in your inference? Enabling this could cause the issue as detailed here: https://www.kaggle.com/code/ledrose/deepcache-cpu-offload-bug

[973398769]

We seem to be discussing the same issue. Although I haven't used enable_model_cpu_offload(), I still experience a decrease in quality when inferring a large number of images, but it's hard to reproduce consistently

[elismasilva]

We seem to be discussing the same issue. Although I haven't used enable_model_cpu_offload(), I still experience a decrease in quality when inferring a large number of images, but it's hard to reproduce consistently

yes i use cpu offload then we need do it disable and enable again inside loop

[elismasilva]

We seem to be discussing the same issue. Although I haven't used enable_model_cpu_offload(), I still experience a decrease in quality when inferring a large number of images, but it's hard to reproduce consistently

how are you generating images, are you using loop or setting number of images in pipeline?