Mikubill / sd-webui-controlnet

WebUI extension for ControlNet
GNU General Public License v3.0
16.43k stars 1.91k forks source link

[ControlNet 1.1] The updating track. #736

Closed lllyasviel closed 1 year ago

lllyasviel commented 1 year ago

We will use this repo to track some discussions for updating to ControlNet 1.1.

catboxanon commented 1 year ago

(2) it seem that the previous "using only mid control for high-res" is broken of deleted?

I believe it was superseded by this PR: https://github.com/Mikubill/sd-webui-controlnet/pull/720

lllyasviel commented 1 year ago

about (1) when global average pooling is True, the controlnet should be only put on the cfg conditional side, otherwise the shuffle wont work very well This is without cfg-only image Rightnow I have to image so that it works image but this will mess up all other controlnets so that we should have: when global average pooling is True, the controlnet should be only put on the cfg conditional side

lllyasviel commented 1 year ago

(2) it seem that the previous "using only mid control for high-res" is broken of deleted?

I believe it was superseded by this PR: #720

cool then rightnow what is the best practice for high res. fix?

lllyasviel commented 1 year ago

(6) oneflow needs dependency "pycocotools"

lllyasviel commented 1 year ago

after 5bad033 it still does not work image

lllyasviel commented 1 year ago

cool thanks for the invite i will setup a development environment today

catboxanon commented 1 year ago

after 5bad033 it still does not work

It works for me, don't need to manually invert now. Usually after any extension update you'll need to hard restart the entire web UI to make sure all the modules are reloaded.

lllyasviel commented 1 year ago

after another commit it begin to work now

lllyasviel commented 1 year ago

right now the lineart works very well image

lllyasviel commented 1 year ago

ok i think the 2 lineart model and 3 lineart preprocessors are perfect now.

lllyasviel commented 1 year ago

cool the 3 seg preprocessors are perfect

lllyasviel commented 1 year ago

ok two remaining problems (1) the visualization of softedge_pidinet is inverted (2) global average pooling should be linked to cfg conditional side ControlNet

pkuliyi2015 commented 1 year ago

Can the Tile model work correctly now?

lllyasviel commented 1 year ago

no. Tile and Inpainting needs more time. Perhaps I can also work on inpaint

catboxanon commented 1 year ago

Point (2) was addressed in a comment in the PR, I can agree with you it should probably be enabled automatically: https://github.com/Mikubill/sd-webui-controlnet/pull/742#issuecomment-1509875513

shuffle requires Enable CFG-Based guidance to be enabled (in settings) for better result.


Can the Tile model work correctly now?

See above PR comment. Inpaint and Tile are not yet supported.

lllyasviel commented 1 year ago

openpose is perfect image

congratulations

continue-revolution commented 1 year ago

no. Tile and Inpainting needs more time. Perhaps I can also work on inpaint

please let me know when inpaint is ready. my sam extension will connect to your inpaint. before that I'll try to connect to your semantic segmentation

lllyasviel commented 1 year ago

normalbae is perfect image congratulations

pkuliyi2015 commented 1 year ago

Yes I also want to the Tile model for upscaling. I got everything ready for draw / redraw images tile by tile, and the only thing need is a model trained on tiles.

Great jobs!

lllyasviel commented 1 year ago

hello where is the codes for preprocessor alias? I cannot find it

Ratinod commented 1 year ago

It would be great if a "force unload all controlNet preprocessors/models from VRAM" button exist.

pkuliyi2015 commented 1 year ago

It would be great if a "force unload all controlNet preprocessors/models from VRAM" button exist.

I don't know, but I provide a Free GPU button in my extension. You may try use that button to see if it can unload controlnet stuffs?

Ratinod commented 1 year ago

I don't know, but I provide a Free GPU button in my extension. You may try use that button to see if it can unload controlnet stuffs?

It unload... but only little amount. (preprocessor) test method:

lllyasviel commented 1 year ago

update: guess mode rewritten, avg pool rewritten these are much more correct than before shuffle seed fixed now shuffle does not need more configs

bropines commented 1 year ago

update: guess mode rewritten, avg pool rewritten these are much more correct than before shuffle seed fixed now shuffle does not need more configs

Now you want me to change the branch to yours?

lllyasviel commented 1 year ago

shuffle + canny iron man, high-quality, extremely detailed Negative prompt: long body, low resolution, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 12345, Size: 512x768, Model hash: e1441589a6, Model: v1-5-pruned, ControlNet-0 Enabled: True, ControlNet-0 Module: none, ControlNet-0 Model: control_v11e_sd15_shuffle [526bfdae], ControlNet-0 Weight: 1, ControlNet-0 Guidance Start: 0, ControlNet-0 Guidance End: 1, ControlNet-1 Enabled: True, ControlNet-1 Module: canny, ControlNet-1 Model: control_v11p_sd15_canny [d14c016b], ControlNet-1 Weight: 1, ControlNet-1 Guidance Start: 0, ControlNet-1 Guidance End: 1

canny image image

shuffle image

result image image

lllyasviel commented 1 year ago

this should be the first really useable style transfer

lllyasviel commented 1 year ago

update: guess mode rewritten, avg pool rewritten these are much more correct than before shuffle seed fixed now shuffle does not need more configs

Now you want me to change the branch to yours?

you have a working branch?

bropines commented 1 year ago

update: guess mode rewritten, avg pool rewritten these are much more correct than before shuffle seed fixed now shuffle does not need more configs

Now you want me to change the branch to yours?

you have a working branch?

Now I'm on the ctrlnet 1.1 branch image

lllyasviel commented 1 year ago

soem other attempts image the style image iron man, high-quality, extremely detailed Negative prompt: long body, low resolution, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality Steps: 50, Sampler: Euler a, CFG scale: 7, Seed: 12345, Size: 512x768, Model hash: e1441589a6, Model: v1-5-pruned, ControlNet-0 Enabled: True, ControlNet-0 Module: shuffle, ControlNet-0 Model: control_v11e_sd15_shuffle [526bfdae], ControlNet-0 Weight: 1, ControlNet-0 Guidance Start: 0, ControlNet-0 Guidance End: 1, ControlNet-1 Enabled: True, ControlNet-1 Module: canny, ControlNet-1 Model: control_v11p_sd15_canny [d14c016b], ControlNet-1 Weight: 1, ControlNet-1 Guidance Start: 0, ControlNet-1 Guidance End: 1

lllyasviel commented 1 year ago

update: guess mode rewritten, avg pool rewritten these are much more correct than before shuffle seed fixed now shuffle does not need more configs

Now you want me to change the branch to yours?

you have a working branch?

Now I'm on the ctrlnet 1.1 branch image

it is ok. u may take a look at my implementation and see if you want to fix any bug

lllyasviel commented 1 year ago

oh my god this style transfer is really killing

lllyasviel commented 1 year ago

@Mikubill I do not know how clip vision works. it seems already broken before my first commit. help needed

pkuliyi2015 commented 1 year ago

Amazing work.

I’m also working for some new ideas in v2v and it seems to be good. Let’s change the game :)

bropines commented 1 year ago

I do not know how clip vision works. it seems already broken before my first commit. help needed

Can you make a manga model? I have a problem with the screenton, when inverted (or high sensitivity of the anotator) it causes inode artifacts

Sorry for the crooked table (I haven't figured out the markdown yet)

00014-2639049337 00013-617035015
00015-4101522603 image (1) tmpel316tni
14 image
lllyasviel commented 1 year ago

I do not know how clip vision works. it seems already broken before my first commit. help needed

Can you make a manga model? I have a problem with the screenton, when inverted (or high sensitivity of the anotator) it causes inode artifacts

what is it?

bropines commented 1 year ago

I do not know how clip vision works. it seems already broken before my first commit. help needed

Can you make a manga model? I have a problem with the screenton, when inverted (or high sensitivity of the anotator) it causes inode artifacts

what is it?

On the 6th photo, the original is taken from the manga. At 4,5,7 the results of the anotator's work. The fact is that he is extremely bad at translating the manga screenton into the line version. 1,2,3 photos show the results of "coloring"

lllyasviel commented 1 year ago

update:

clip vision fixed

0xprincess commented 1 year ago

Hi everyone! I'm getting an unexpected keys error when trying to use fp16 v1.1 models. The full version works fine. How can I fix that?

saint4626 commented 1 year ago

image Error after the update, can anyone help?

lllyasviel commented 1 year ago

update and restart everything completely

saint4626 commented 1 year ago

image I updated and completely rebooted everything, it didn't work, annotator works, generation doesn't work

lllyasviel commented 1 year ago

image can you see "cdcfa706"?

saint4626 commented 1 year ago

image yes

lllyasviel commented 1 year ago

then close the console and browser completely and the restart and it should work

saint4626 commented 1 year ago

Finally, a strange mistake hmm, thank you!

AbyszOne commented 1 year ago

Ironically, the 3rd point is the first to be added, with this shuffle model, while no native batch was implemented whatsoever. Happy and sad face.jpg

throttlekitty commented 1 year ago

soem other attempts ... iron man, high-quality, extremely detailed Negative prompt: long body, low resolution, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality Steps: 50, Sampler: Euler a, CFG scale: 7, Seed: 12345, Size: 512x768, Model hash: e1441589a6, Model: v1-5-pruned, ControlNet-0 Enabled: True, ControlNet-0 Module: shuffle, ControlNet-0 Model: control_v11e_sd15_shuffle [526bfdae], ControlNet-0 Weight: 1, ControlNet-0 Guidance Start: 0, ControlNet-0 Guidance End: 1, ControlNet-1 Enabled: True, ControlNet-1 Module: canny, ControlNet-1 Model: control_v11p_sd15_canny [d14c016b], ControlNet-1 Weight: 1, ControlNet-1 Guidance Start: 0, ControlNet-1 Guidance End: 1

Using your settings, I'm not getting similar images with Shuffle, the background is always a swirly mess. Or when using Shuffle by itself, I see that the output directly matches the flow that the annotator preview show. imtmp

lllyasviel commented 1 year ago

@throttlekitty make sure that your yaml file names and model file names are same

throttlekitty commented 1 year ago

That set me in the right direction, thanks. The extension places yaml files in its own models folder, but reads from webui\models\controlnet. I needed to have the yaml files in the same folder. It could also be that my webui\models\ is a symlink from another drive?