CharTurner

I would love to do a true orthographic. Each one of the source images in the dataset IS a true orthographic. I will keep training, but it's not something that's easy to teach the AI, at this point. I'm working on it. :D

@WAS yeah, that's the eventual goal, just not there yet. XD

not perfect, but the way I've done is is to make a turnaround of a similar character (in build), replace one of the poses with your character, and then run it through Img2Img with an inpainting model, with the 'target' character masked. Regular models (vs inpainting models) don't look at the character enough to make it work.

@kasukanra Use control net, and other than that, it's only an embedding, not an actual program, so it's something you'll just have to keep trying at. This is a helper, not a generator. :D

YUP! It's AMAZING. It will pull the pose from a good turnaround and you can use it to make the perfect poses, it's AWESOME.

@orwelian84 https://github.com/Mikubill/sd-webui-controlnet

If you have the 1.5 version loaded, and a 2.1 model, it will show it's not valid, or the other way around (both versions are available). Other than that, no idea.

Hm, not sure what's going on. As long as it's in the embedding folder, and you're using the 1.5 version with a 1.5 model, it should work.

Need more info to help. What version, what model, what prompt, or is it not loading?

Hm, weird, i'll look into it.

@Chandle It's not an hypernetwork, it's an embedding. I'm not sure what happened, it working fine on my side, I'll reupload it.

Just to confirm, @bad_leaf and @Chandle there's a version for 1.5 and 2.1 models, and it will skip loading the one that doesn't work. IE, if you're using 21CharturnerV2 in a 1.5 model, you'll get the error. Make sure you download the one for 1.5.

@Chandle There's a 1.5 version as well, if you want to use it in those. Good luck! :D

The embedding was trained on my system, from a small dataset I put together of my own character turnarounds, plus a few of character turnarounds I found online (mostly photo reference, to make sure it didn't only do illustrated turnarounds). I trained it using Auto1111's training tab, which outputs a pickletensor embed (pt file), which I uploaded here.

Make sure you restart. They don't show up under extra networks?

@JohnnyDoe303 The lora needs to be set low, such as .4 or so. Try out one of the prompts on either my example images, or some of the other ones in the reviews (check the little i with a circle around it in the lower corner of each image). It does work, even if it's a little fiddly.

well, it's not a lora. Make sure you have the one loaded for the models you are useing (ie, the 1.5 version won't show up if you' have a 2.1 checkpoint loaded, and vise versa)

@May2xx Download the other version, just in case, and put both in the embedding folder, restart, see if that fixes it.

use control net (extension) find the exact turn you want (pose libraries, photos, other reference) and then use charturner to make sure everybody is wearing the same outfit.

The Lora version might be better to lock it in for your needs.

You can! just do a wide version, and if you want very set positions, use the extension ControlNet to control the poses, and CharTurner to make sure it's all the same character. Good luck!

Should be in the metadata of the images, other than the grids. Wider images get more turns. Square tends to front and back. Sampler used is up to you, I tend toward DDIM at 20 steps with highrez fix, set to 1.3-1.5 x with latent nearest exact. However, it should work with most settings, it's not locked into any particular size or sampler. (

I made this before ControlNet was a thing. OpenPose will 100% be more accurate on poses, but it also won't make up poses (ie, if you want a solid T or A pose, controlNet is the way to go.

Using CharTurner WITH controlNet is my favorite: controlNet keeps the poses exactly where you want, CharTurner makes sure you get multiple images of the SAME character, not several characters. Both is better.

I also trained it pretty carefully to be a POSE embed, not a character embed. As you can see from the wide variety of characters people have made with it, none of which are the characters I trained it on.

There IS a lora version, if you want it. Check my profile. :D

I'm not sure what you're asking for, to be honest. You can use img2img with it, yes. If you're wanting to inpaint poses, make sure you're using an inpainting model.

@hq901112603 Yes. Here's how i do it: https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/035cbefd-7770-424a-4a63-0ffb18dae200/width=1772/5.jpeg

make sure you downloaded the right one? There's a v2 for 1.5 and a v2 for 2.0 models, and a lot of people download the 2.0 one (it's the first one up.). Click on the one that says for 1.5

@mechamuffin The works, but i made it before lora or controlnet were a thing. both of those are a little "easier". However, it SHOULD be working without them. Try using a prompt from one of the example images. using the words "character turnaround" and making your image wider than it is tall both help.

@magikal stable-diffusion-webui\embeddings :)

There are two versions, one that loads for 1.5 models, and one that loads for 2.1 models. The 21CharturnerV2 is for the 2.1 models. If you're working with 1.5 models, make sure you grab the other version. (embeds don't even show up if they are for the wrong model.)

Same way most stuff is trained. Collected a bunch of turnaround (some I made, some I found), captioned by hand, trained, used that training to make some more, trained, repeat until it's useful. I think I trained around 25 versions of this before I was happy with it. Lora is made with the same dataset.

@wsjshaoyeahnet697 Most images have the prompt included, look for the little i in the circle on the example image. Make sure you're a little wider than square, and if you really really really need it in a very particular pose, use this AND control net with an image of the poses you want. While I tried to train in a key word specifically for 3 view, 4 view, 5 view it doesn't consistently enough to recommend it. In general,

"Character turnaround, front back side view," and charturner should be enough to make it work.

100% this is the answer to almost every "i can't find it" issue. It sucks that you have to have both the 1.5 and 2.1 versions if you use both types of models. I tried to make it clear in the text that there's a file for each 'type' of model (1.5/2.1) but it is still unclear.

if you are using 1.5 models, you need CharTurner V2 for 1.5.

If you are using 2.1 models, you need CharTurner v2 for 2.1

@FoxDude that is very weird! if you have both, and they aren't showing, I have no idea what the issue could be. Could try the CharTurner Lora in my profile.

The embed and lora versions won't make the system be able to produce a character the system can't do without it, it only helps you make a character turn around (same character, different pose.) So, if you can make the character without charturner (lora, OR embed) than you should be able to make a turnaround with this.

If you can't make the character, you need a lora to make the character, and then you can use this to make a turnaround.

ControlNet helps make sure the pose is 100% what you want (since the embed and lora sometimes pick their own poses), but doesn't help you keep consistent characters.

Good luck! :D

There's a lora. I don't know how and can't find a way to change an Textual Inversion Embed into a safetensor; checkpoints and LORA have safetensor options, but Embeds do not by default. If you can point me at a resource to do change it into a safetensor, I will make one.

@TheRoamingSwamp I have Auto1111 installed, so i'll check that out. I had thought that only worked for checkpoint models, and not the smaller PT embeds. Thanks for the info.

posting this here for reference

Can I convert other types of models that aren't Checkpoints to .safetensors?

Yes, you can also convert any type of model to .safetensors using this method. This means you can also convert LoRA, embeddings, hypernetworks, VAE and also special models like Pix2Pix, inpainting, ControlNet models and presumably other types of new models that might come out and that aren't already in .safetensors for whatever reason.

Just make sure their file extension is .ckpt. For example if your embedding is called embedding.pt then rename it to embedding.ckpt. They need to be renamed .ckpt so the Model Converter extension can detect them.

Then simply place the embedding, VAE, etc. you want to convert in the root of stable-diffusion-webui\models\Stable-diffusion and then the extension should detect them. Use the blue Refresh list button in the extension if needed to make them show up if they aren't, even after renaming them to .ckpt. Alternatively, restart the UI.

@TheRoamingSwamp Safetensor version uploaded (for 1.5 models, 2.1 version safetensor coming soon)

@TheRoamingSwamp ah, darn, i hadn't checked them. Thank you for the feedback, I will see what i need to do to fix it. Thank you for letting me know.

Sorry about that! The .safetensor version is broken, (the method I was using to change the .pt file to safetensor didn't work.). I thought I had deleted all of my attempts, i missed this one. Sorry about that.

I will eventually get a working safetensor version of the embed working.

If you use control net to control the pose, you can get as many 'gradual turns' out of this as you want, and charTurner will make sure they're all in the same outfit/character. It won't work as separate images, for it to be all the same outfit, it needs to be all the same gen.

If you're talking about giving it the front and side and back and having it fill in the 'missing' poses, you can do that with this, plus control net, plus inpainting. There's an image in the gallery of how to do that.

With controlNet openPose now available, I see no reason for me to spend another 100-150 hours to create either of the requested models. :D

Nope, none yet. I made the PlanIt embed (check my profile) trying to do exactly that. You need a large dataset that shows all the views you want (ie, a well tagged orthographic library dataset) . However, with Planit, you can get part of the way there. Everybody is chasing the AI to 3D pipeline. XD

@latent077 https://civitai.com/models/4775/planit-a-documentation-embedding

Not very well. I have a second embed, called "PlanIt" in my profile, it does better with products, though it doesn't do as constant a turn around.

1) The first tiny dataset was character turnarounds that I made over the last 20 years, or a few I found, and then made a early version, which i used to make a dataset of 35 by the 25-26 version of the trainer.

@langhr241192 You're fine. Yes, the final dataset was 35 or so images. :D

Hm, front and back should be doable. I've noticed that the highres.fix pass sometimes seems to 'flip' somebody around, in some models. Try it without that, as a test? The other thing is you can try the char-turner Lora (check my profile), it might be better at 'forcing' it.

The other thing to try is take one that has the front and 'not back', mask the front, and ask it for the back view, but put "eyes, nose, mouth" in the negative.

Hi there! This was created before we had control net, so originally it was mostly used with text2img to try to generate the whole turntable all at once. if you have one pose, and want to try to get the rest, the easiest way is to find a image (or generate one) that has the rough character shapes you want (very rough), replace your source image over one of the figures in something like photoshop, and then in Img2img, mask the source image and tell it to inpaint not masked. Still can be hit or miss, but i've had pretty good luck with that method. One of the images in the gallery is a little tutorial on that. :D

Yes. One of the images in the gallery is a tutorial on how to do this. if you need more info, check the other comments; it's by far the most often asked question (and why i made a tutorial page. :D)

I don't know. What model are you using? I didn't train it in NAI so i've heard it doesn't work as well in that. Are you making the image wider than it is tall?

@cipher_wh Orange Mix has some issues with it, for sure. AOM probably has the most problems with it of all. However! Try 512x1024, high res fix helps (i usually go 1.3 size with 'latent Nearest Exact')

1) I did train it in the webui, with the included training tab.
2) Yes, it was very difficult to find settings that worked well with my dataset
3) I'm not surprised. This embed was VERY difficult to create. I trained more than 50 copies (went thru the alphabet twice on testing versions). I'd highly suggest you use ControlNet OpenPose to help force the character into the poses. ControlNet hadn't come out when I started down this path to make the embed. XD

Generally, using controlnet will override the poses in the embed, but you can try pushing up the weight to 1.2

@Rikuthin Right. I had 0 characters with wings or tail in the dataset, it was all humans. So the more you ask charturner to be 'strong' its going to pull away from the wings and tail. You probably will have to do it as a 2 step process: charturner with control net for the poses, take them into inpaint for the wings and tail.

I hadn't thought to do one, specifically. I'll think about it, but mostly i've been using the other ones.

are you trying to train a lora of the character? Or are you including it in a larger data set but are training a style or something?

@lovelyzoey You can 100% include these images in a dataset to train a lora. Depending on WHAT you are trying to train your lora to do, it may or may not help. If you put several of the same person into the training set, the lora will learn that person, but just one image of a turnaround probably isn't enough to make that person show up over and over again, unless your dataset is very very small.

I want to help more, but without knowing WHAT you are trying to do I can't give more advice.

I don't make videos. If you search for charturner on youtube, somebody else made one in Spanish.

you can try moving the trigger to the start of the prompt, but the embed doesn't work perfectly with all base models. You can also find a turn around you like and put it into ControlNet, and use and charTurner together. That way you should get the same outfit on all the turns. Also make sure you try making "wide" vs "square" images to increase the number of turns. Good luck

This is an embedding, it will work on any system that can run a textual embedding. It is not a stand alone bit of software; it has to be running inside a stable diffusion front end. I have no idea if there are different types of embeddings for mac. If you are using a front end like Auto1111, on mac or PC, this embedding will work.

I never trained it, because with openpose it's so easy to get exactly the turn around you want. If you still would like it, i will add it to the To Do list.

This is a textual embedding. I've never used Fooocus but this has instructions https://www.reddit.com/r/fooocus/comments/18hth1h/using_embeddings/

which pony version? SDXL pony or 1.5 pony? (i've only used pony a little)

I never got around to making an SDXL version, because controlNet OpenPose became a thing, and most people swapped to using that with sdxl and Pony.

Also, the concept of 'character turnaround' is at least partially existent in the SDXL base model, so if you use openpose and good prompt you can probably do as good as the original dataset would get you if I trained it now. You could run the 1.5 version to get a set of poses you like, and use openPose with pony to guide your generation.

... and i just remembered that people can use the onsite generator. Hm. Okay, maybe I SHOULD train a pony/IL version. Nobody's got one going yet? I sorta stopped working on this because other people had it covered.

@nroonarij129 You have to have the SDXL openpose model, it's different than the 1.5 version. Let me know what kind of troubles you're having. :3

@jellycap I could but if you use openpose you might be able to get the same effect without it. But that doesn't help with the online thing, so i'll work on it.

Hi there! Yeah, this was made long before iP adapter, and the way that control net works messes with this for sure. I'd try doing it without the adapter, and then inpainting all the faces with it as a second step.

The 360 turn is easier with a wide vs square image, but yeah, adapter is messing with that too.

Really i meant this works well with controlnet, if you need very specific poses but the same character and outfit.