I see that now confused it with the the SS image count.
I think the hash info can get wacky sometimes when your trying to parse it. (9c4943fbae5fbc56c2178c84e5f0af3f) for the checkpoint.
I think the hash info can get wacky sometimes when your trying to parse it. (9c4943fbae5fbc56c2178c84e5f0af3f) for the checkpoint.
daddyCzapo
Member
- Mar 26, 2019
- 245
- 1,507
First real try with ComfyUI
I must say it's fast as fuck boi. 3072x4608 px in 90 seconds. The same specs of image wouldn't even generate on A1111 let alone in that time. I'm veery impressed. All it took from me was 1 sleepless night XD
I must say it's fast as fuck boi. 3072x4608 px in 90 seconds. The same specs of image wouldn't even generate on A1111 let alone in that time. I'm veery impressed. All it took from me was 1 sleepless night XD
Sharinel
Active Member
- Dec 23, 2018
- 613
- 2,586
Sleep is for the weak. How easy was it to install? I might give it a go, although I have so much invested in SD it might not be worth moving.First real try with ComfyUI
I must say it's fast as fuck boi. 3072x4608 px in 90 seconds. The same specs of image wouldn't even generate on A1111 let alone in that time. I'm veery impressed. All it took from me was 1 sleepless night XD
Sharinel
Active Member
- Dec 23, 2018
- 613
- 2,586
Stylejourney is just so different from any other checkpoint I've used, not sure I like the 'painting' effect on the skin, does it seem distracting?
daddyCzapo
Member
- Mar 26, 2019
- 245
- 1,507
You basically just unzip the archive that weighs about 1.5GB or git clone the repo. Then you just edit one file to point to your a1111 installation so comfy can just use the checkpoints and you're set. Everything is described
You must be registered to see the links
Glorified_ignorance
Well-Known Member
- Mar 28, 2019
- 1,110
- 4,181
Just a thought, sorry for misplacement of post(probably), but been wondering:
Does Ai capable of cleaning "censored" Art ? The ones that mostly come up from Japanese games. Or basically any.
In case this was talked over somewhere, can anyone redirect me there?
Also this:
Does Ai capable of cleaning "censored" Art ? The ones that mostly come up from Japanese games. Or basically any.
In case this was talked over somewhere, can anyone redirect me there?
Also this:
You don't have permission to view the spoiler content.
Log in or register now.
daddyCzapo
Member
- Mar 26, 2019
- 245
- 1,507
I think it depends on how heavily the art is censored.. if it's just blurred you could try inpainting in img2img, or try to upscale using esrgan upscaler, or just interrogate the original and regenerate the whole scene based on a prompt. The only limit is your imagination
Artiour
Member
- Sep 24, 2017
- 314
- 1,319
just a note, my LoRAs work better when generating with 2:3 (512X786 or 683X1024) than with 1:1 (786X786 or 1024X1024), I don't know yet if it's a model thing or a LoRA thing, or even AI thing
Artiour
Member
- Sep 24, 2017
- 314
- 1,319
They have tools that where trained to UN-censor depending on the type of censoring. (Mosaic vs black bar etc.)
The issue isn't so much with the lack of tool's it is the complexity of getting the images back into a game and the time it takes to do so.
Testing out Img to img on Breeders of the Nephelym images (That game could make some interesting loras)
devilkkw
Member
- Mar 17, 2021
- 338
- 1,145
Don't tested the model, but those result seem a merged model with incorrect layer, i say that because i had same result in some of my merging test. This don't mean model is not good, i like your result, but those crap effect on jewel and eyes for me ruined image. If you like model, you can try to fix with merging, or by resetting clip layer.
Some image for today:
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
I had the same experience but watch out for the automatic 1111 fanboys when you speak to how much faster comfy is....they usually are like you just installed automatic 1111 wrong
I have both tools still installed on my machine and automatic is still the first tool I used and would recommend for first timers though.
It uses SD checkpoints, LORA's etc, making sure to install it to a virtual environment is important as it doesn't have the .bat files like Auto1111
(My Comfy is installed to base machine which is a mistake I am not going to fix)
If you edit the extra_model_paths.yaml to point to your Automatic 1111 install it will save you copying all the checkpoints and lora's and allow you to use both programs easily.
Its comparable to installing old Automatic 1111 WebUI when it was all command line (Before the bat files)
Last edited:
Nyah! nyah....bad cat girl, swallow cream....Butt punishment later
Converted to Anime Style
Original Realistic Style
Converted to Anime Style
Original Realistic Style
Last edited:
- ComfyUI is the fastest, and arguably the most powerful. Just can be difficult to adapt to a full node workflow.
- InvokeAI is the new kid on the block almost as fast as ComfyUI and forms a halfway house between Comfy A1111. Features can lag behind both comfy and A1111, but it's got a faster and easier interface than either, nodes are also under rapid development. It's also had a major UI and backend overhaul recently which is why it's behind.
- A1111 is the grand daddy of them all, interface is a PITA to use but it's real power lies in it's extensibility, it is slower than either of the above and lacks the node workflow.
Probably somewhat system dependent, having reinstalled both many times I have a slightly different opinion on ease of install.
I'm not sure if you have used the latest Auto 1111 WebUI (Not just Auto 1111) the .bat file does everything including making the VENV
Have you ever tried the full Onyx conversion...cracks me up NVIDIA and Microsoft partnered to do something that arguably benefits AMD cards more.
My Big Fat Greek Wedding
Last edited:
Most of my work is on Linux, mainly because I run a headless box but when I tested Windows I ran 'conda to manage dependencies and installation of libraries. Onyx is a mixed bag in my experience, AMD still has not ported the full ROCm suite. MIOpen and a few other core libraries are missing before we can truly assess AMD performance at least on Windows. On Linux AMD is just as good as nVidia in many cases, arguably better due to them shoving more RAM on their boards.
I have not installed A1111 for a while, I prefer to not do multiple installs if I can help it. Once it's up, barring something going FUBAR, I tend to leave well alone. This is something Invoke gets right, it's installer is also a repair and management tool making system management a tad bit easier. I only use A1111 when I need some oddball extensions that's not available on the other two UI. In most cases I can use git pull to update, but Invoke has an inbuilt updater to it's startup script. Really comes down to picking your poison as there is no clear winner between them as they all have their strengths and weaknesses.
You might want to look at
You must be registered to see the links
they claim a 2x boost over SD checkpoints and Microsoft made a conversion tool.Funny that the NVIDIA uses 768x768 and AMD uses 512x512 but they both claim around 2x speed for IT/S
Having CUDA I haven't bothered. (Also the
You must be registered to see the links
being command line doesn't make me want to use it but I did download it.)In theory any existing model on huggingface could be converted as the store all the generation data
I think the full unpruned XL checkpoint is 14GB so I would hate to see the requirement to covert it
With a background in Linux you probably found a lot of things easier then some of us coming in with no GIT or python or conda, or maybe even windows power shell experience.
Im with you on the each have their own faults and benefits. I still don't think their is a 3D object generator for comfy like their is for Auto1111. (Not that it ever generated anything useful for me.)
Last edited:
devilkkw
Member
- Mar 17, 2021
- 338
- 1,145
For censorship i use inpaint, but in different way.
Many of user just impaint a part of body and try to decensor it, there's one way.
I prefer impaint the part i want to keep, set "impaint not masked" and "original" in masked content. also check "impaint whole image.
This is an example( also called deepfake)
I used image from shirogane_sama cosplayer.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
Prompt was simple "nude girl".
Settings is really different for model's you are using. in my case there are mask blur:45, cfg:40, denoise:0,45.
Made some test and find value in different model require time, you need to test and find correct for your.
Also value change for anime, toon model, just try.