WE NEED VERSION 5, PLEASE!
krigeta opened this issue · comments
its been years but the version 5 is still not released, why?
V5 is essentially already released... controlnet, the lineart model, automatic1111, you have everything you need to color stuff.
Here's an example.
V5 is essentially already released... controlnet, the lineart model, automatic1111, you have everything you need to color stuff. Here's an example.
What do you mean? Where is this hosted?
This github contains only a preview for v5 as far as I know.
Is there an executable for v5 anywhere?
Install this: https://github.com/lllyasviel/stable-diffusion-webui-forge, download the the lineart model here: https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main and download some anime model off of civitAI.
V5 is essentially already released... controlnet, the lineart model, automatic1111, you have everything you need to color stuff.
Here's an example.
No, it is not as good because we have the option to choose desired colors in Style2Paints V4, but using Automatic1111 and ControlNet won't help; it changes the lineart due to the involvement of stable diffusion models. However, in Style2Paints, the lineart remains good as before.
V5 is essentially already released... controlnet, the lineart model, automatic1111, you have everything you need to color stuff.
No, it is not as good because we have the option to choose desired colors in Style2Paints V4, but using Automatic1111 and ControlNet won't help; it changes the lineart due to the involvement of stable diffusion models. However, in Style2Paints, the lineart remains good as before.
By upping the value of controlNet, stable-diffusion will leave the lineart virtually untouched. Color pointers are also perfectly possible either via precoloring in minipaint/krita or automatic via style loras. The only part of a potential V5 that hasn't released yet are the SD models specifically made for coloring (Dorothy/Alice) but I wouldn't count on those ever being released. Again, the coloring in my first post was done in automatic1111 with a controlnet value of 1.3 and the simple prompt: pink hair, blue eyes, blue jacket, red jacket, blue swimsuit, red swimsuit, red thigh highs, blue thigh highs . If you precolor, up the controlnet value and give a better prompt you can get results that are comparable or better to those shown in the V5 preview.
Here's another example with some more effort put into it
Virtually no loss of the original lineart.
Here's another example with some more effort put into it
Virtually no loss of the original lineart.
wow this result is amazing and something I want to do, may you guide me? any social media where we can chat?
Yeah, sure, you can add me on discord: stephen__daedalus
Here's another example with some more effort put into it
Virtually no loss of the original lineart.
the issue is that not all of the panel is coloured, the lighting looks very offputting and there are several other flaws. You're better of using v4(not v4.5)
Here's another example with some more effort put into it
the issue is that not all of the panel is coloured, the lighting looks very offputting and there are several other flaws. You're better of using v4(not v4.5)
Only the part I masked is colored, of course, but there is nothing keeping you from coloring the entire page. Idk what you mean about the lighting being off since shadow's and reflections are part of the base line art... maybe you think that it's too colorful but that can easily be fixed by switching to a model with more matte colors. Any other "flaws" can also be fixed with inpainting.
You guys seriously overestimate the abilities of a potential v5 release. V5 would just be controlnet+ a year old custom SD1.5 model that is likely already outclassed by newer 1.5 models and definitely outclassed by sdxl models.
Here's another example with some more effort put into it
the issue is that not all of the panel is coloured, the lighting looks very offputting and there are several other flaws. You're better of using v4(not v4.5)
Only the part I masked is colored, of course, but there is nothing keeping you from coloring the entire page. Idk what you mean about the lighting being off since shadow's and reflections are part of the base line art... maybe you think that it's too colorful but that can easily be fixed by switching to a model with more matte colors. Any other "flaws" can also be fixed with inpainting.
You guys seriously overestimate the abilities of a potential v5 release. V5 would just be controlnet+ a year old custom SD1.5 model that is likely already outclassed by newer 1.5 models and definitely outclassed by sdxl models.
Its just that v5 will be a more streamlined process that will be different to SD. If you look at the original drafts for v5, this SD incarnation looks nothing alike the draft. Also, V4 seems better imo
Here's another example with some more effort put into it
the issue is that not all of the panel is coloured, the lighting looks very offputting and there are several other flaws. You're better of using v4(not v4.5)
Only the part I masked is colored, of course, but there is nothing keeping you from coloring the entire page. Idk what you mean about the lighting being off since shadow's and reflections are part of the base line art... maybe you think that it's too colorful but that can easily be fixed by switching to a model with more matte colors. Any other "flaws" can also be fixed with inpainting.
You guys seriously overestimate the abilities of a potential v5 release. V5 would just be controlnet+ a year old custom SD1.5 model that is likely already outclassed by newer 1.5 models and definitely outclassed by sdxl models.Its just that v5 will be a more streamlined process that will be different to SD. If you look at the original drafts for v5, this SD incarnation looks nothing alike the draft. Also, V4 seems better imo
You can achieve the same results as the v5 preview in SD. My examples look worse because they are not technically lineart, but if your base image is similar to the ones in the v5 preview (sparse lines, no shading, no solid black or grey colors) then you'll get better results. SD right now beats v4 easily, you just have to take the time to fine tune the settings and prompt.
Here's another example with some more effort put into it
the issue is that not all of the panel is coloured, the lighting looks very offputting and there are several other flaws. You're better of using v4(not v4.5)
Only the part I masked is colored, of course, but there is nothing keeping you from coloring the entire page. Idk what you mean about the lighting being off since shadow's and reflections are part of the base line art... maybe you think that it's too colorful but that can easily be fixed by switching to a model with more matte colors. Any other "flaws" can also be fixed with inpainting.
You guys seriously overestimate the abilities of a potential v5 release. V5 would just be controlnet+ a year old custom SD1.5 model that is likely already outclassed by newer 1.5 models and definitely outclassed by sdxl models.Its just that v5 will be a more streamlined process that will be different to SD. If you look at the original drafts for v5, this SD incarnation looks nothing alike the draft. Also, V4 seems better imo
You can achieve the same results as the v5 preview in SD. My examples look worse because they are not technically lineart, but if your base image is similar to the ones in the v5 preview (sparse lines, no shading, no solid black or grey colors) then you'll get better results. SD right now beats v4 easily, you just have to take the time to fine tune the settings and prompt.
You save a lot more time using v4, as all it is is just lines on a canvas. Also, several of the examples in the demo of v5, look slightly worse than the same examples in V4. I will just wait 5 more years until v5 is released before I can change my opinion.
Yeah, sure, you can add me on discord: stephen__daedalus
I made a discord for people interested in this topic, check it out here!
https://discord.gg/xF4ZMnKsAG
we need v5 sd is too random and not very good at colorization.
In terms of precise control of color at each point, Style2Paint is still more effective than ControlNet.
I've tried many times, SD can sometimes do an exceptional job and sometime it can even due so extremely fast if your lucky or your line art is very simple. But try coloring an entire manga with the same color style, properly color all the objects is extremely error prone, and cumbersome. trying to find the perfect prompt to color each object in a scene is a complete hassle, sure masking can help but it slows things down drastically, Plus when describing colors you have way less control. pink has 1000's of shades, we don't have words to properly describe all that.
with style2paints you simple pick a color say #F953A3 and click the area that should be that color, quick easy, repeatable across many pages,
I've tried many times, SD can sometimes do an exceptional job and sometime it can even due so extremely fast if your lucky or your line art is very simple. But try coloring an entire manga with the same color style, properly color all the objects is extremely error prone, and cumbersome. trying to find the perfect prompt to color each object in a scene is a complete hassle, sure masking can help but it slows things down drastically, Plus when describing colors you have way less control. pink has 1000's of shades, we don't have words to properly describe all that.
with style2paints you simple pick a color say #F953A3 and click the area that should be that color, quick easy, repeatable across many pages,
exactly, I think s2p works better as a GAN than diffusion IMO. I would have preferred the original project SEPA which was a complex node based system to colour line art
controlnet, the lineart model, automatic1111,
@Postmoderncaliban do you have any advice or links? I've been looking around and trying to get any kind of colorizer to run and couldn't find any that would do a good job, what do you use and how?
You say:
Install this: https://github.com/lllyasviel/stable-diffusion-webui-forge, download the the lineart model here: https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main and download some anime model off of civitAI.
I've installed the webui, but then I have no idea what I do with the lineart model or what I do with something off civitAI.
I've downloadedthe webui last week, been searching for instructions around, and it seems like nobody is actually explaining how to do anything anywhere...