search

DarthAffe / StableDiffusion.NET

C# Wrapper for StableDiffusion.cpp
MIT License
60 stars 10 forks source link

when change to 1.2.0，got full noised image output #8

Closed taotaow closed 5 months ago

taotaow commented 6 months ago

when changed to StableDiffusion.NET.Backend.Cpu 1.2.0 got full noised image output stabledifffusion.net.backend.cpu.1.1.0 version is good

DarthAffe commented 6 months ago

i can't reproduce that. are you using any loras? in that case this might be related: https://github.com/leejet/stable-diffusion.cpp/issues/242

if not, does the same issue occur when you run the tool provided by stable-diffusion.cpp (sd.exe) with the exact same parameters?

taotaow commented 6 months ago

yes,I used lcm lora.I check the log,and see large amout skip lora errors like below: model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_proj_in.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_proj_out.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_k.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_out_0.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_q.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_v.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn2_to_k.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn2_to_out_0.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn2_to_q.alpha' with n_dims 0

model.cpp:893 - skip tensor 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn2_to_v.alpha' with n_dims 0

taotaow commented 6 months ago

should apply this commit to fix this problem: https://github.com/leejet/stable-diffusion.cpp/commit/760cfaa61896a02dab8b1ac84a59b17ddf951090

DarthAffe commented 6 months ago

this commit is included in the 1.2 backends (they are based on the currently latest release master-ce1bcc7) but it seems like this fix somehow caused the issue, but i don't know any more about this than what's written in the issue over there,

taotaow commented 6 months ago

master-6d16f68 is ok master-760cfaa commit https://github.com/leejet/stable-diffusion.cpp/commit/760cfaa61896a02dab8b1ac84a59b17ddf951090 cause the bug master-ce1bcc7 keep that bug and add skip tensor log.

DarthAffe commented 6 months ago

there are new backend-packages available (1.2.1) that contain the latest release of stable-diffusion.cpp which should fix the issue.

taotaow commented 5 months ago

yes,1.2.1fix the issue thank you