- This topic has 104 replies, 7 voices, and was last updated 2 years, 1 month ago by
.
-
Posts
-
Had some time the past 2 days to really get to know this software, it has a lot of different uses, I think i've definitely become a fan of using it.
Had lots of output that looked like something that crawled out of 1980s Chernobyl: people with too many limbs and other bodyparts, and other deformities.
But after getting the hang of it, got it under control. Just some questions for some of you all that have mastered it:
1) any good places for Lora's and models besides civitai?
2) good source of learning this inside and out? Thinking of paying for a membership at that one site that seems to be the major place for info, but wondering if there's something else available? Maybe that already is the best option? Theres lots of sites with vids and info but when you are starting at ground-level all their info flies right over my head.
3) can a number of various DAZ renders be used to create good "pose" Lora's? I know how create them with kohya, just wondering of the pro's and cons of doing this with DAZ renders vs using various images from the net. Any any tips on doing it with DAZ, such as camera angles and number of images to use, etc to make a good Lora.1. Civitai is the go-to place for resouces. There are other sites but bear in mind tensor files (python modules) present a security issue. Arbitrary/malicious code can be executed when run. Civitai is probably safer than random sites. They label loras as either Safetensors or pickle, the latter being inherently unsafe and should be avoided.
2. Personally I'm a visual learner so Youtube is the place I use for learning. Look for the most recent tutes because a lot are out of date as AI software changes continually. Many tutes include a direct link to the recourses.
3. There's no reason why DS images can't be used to generate SD poses with ControlNet. I've made OpenPoses with photographs in ControlNet but not too successfully. Civitai has a lot of free poses you can use if you have ControlNet installed.
ps - I only started to use the Adetailer extension a few weeks back and I'm kicking myself for being late to the party. It's probably the best extension out of all of them and saves a ton of time. SD images browser is another must-have.
Thanks for the answers. Ive seen some vids on YT, of course, but generally they are concerning a particular subject/task. And some require knowledge of things I havent learned yet. Guess im just looking for a learning "series", kinda like the DAZ masterclass vids, where each chapter builds upon what you learned previous. Something to work through. Not sure it exists, only thing ive seen resembling that is that membership site, somewhat.
I've tried some openPose in controlNet, its been hit or miss(mostly miss), seems like it ignores my pose completely a lot. Figured having a Lora would help, maybe even use together with an openpose.
Some excellent SFW examples of how good it can be, can be found here...
https://www.tumblr.com/renderman1406
Go quick as he's taking it down soon as it's run it's course. We chatted quite a bit and learned a lot both ways. He's also a long time Studio user.I've found that Lora poses tend to change your character's look much more than openposes do.
These Youtube people may not have the learning sequence you're after but they're tutes are clear and easy to follow.
OlivioSarikas
Sebastian Kamphthank you both. Ill def check those resources out.
And yes, I see the Lora's do change the character look. It "appears" that parts of the results of a Lora are remnants of the images used to train it. It seems like the result is a mish-mash of the trained images. Ive seen remnants of watermarks and logos in final results. Thats why Im "assuming" a good DAZ render would be superior, it could be very high quality training images without noise/grain/watermarks and can be exactly what I require without deviations. Im just wondering if there is not enough variance to get a good train, thats why i'm wondering if anyone else had tried that and had input and/or tips.
I've made Loras that worked really well despite not having used a lot of high quality images. I think as long as there a few good ones in the batch SD can fill in the gaps. Also note that some checkpoints, particularly CyberRealistic, already contain loads of data on well-known people.
These were were made using no Loras just the names. (King Charles wouldn't initially work because I forgot he wasn't a king when CyberRealistic was made, had to call him prince)
I was thinking more of for creating poses with the lora's rather than for using for characters. I cannot seem to get openpose to work, like at all. When I first installed it on day one, it seemed to work when I ran a test prompt, and then not at all ever since. I cannot even get it to do a simple pose like a hand on a hip, dozens of times rendering in batches of 8, still wont budge. And then when I finally do get one, im pretty sure its just a fluke, even a broken clock is correct twice a day.
Tried re-downloading the controlnet files, still not happening. The only relief I get is using a lora or img2img. Are openpose only an option for specific checkpoints? I've tried multiple checkpoints, still no luck.So in the meantime, until I get that figured out, was thinking I could make Loras. My thinking was to just quick gen8 character, create about 25 camera angles. Pose it. Render all cameras.
Then: Load a full-body morph, change expression, and use bone doctor to slightly relax or contract all bones by like 1-2%, then render all cameras again. Rinse and repeat with new body morphs etc, should get a few hundred good quality images. Wondering if anyone has tried such a thing before I waste day of work.Make sure you have enabled it and set the Preprocessor to "None". Select OpenPose as the control type and "Control_v11p_sd15_openpose" as the model.
Yes, I've been doing that exactly. Picked up that same knowledge from this place: https://www.nextdiffusion.ai/tutorials/how-to-use-open-pose-controlnet-in-stable-diffusion
But no matter what, it just doesnt work at all, the only other thing I could try is to uninstall and re-install, but I really hate to do that.
Ok, I got it working. I first un-installed ControlNet and Openpose, then re-installed. That didnt fix it. (and it was a bit of a problem doing it, but different story for a different day). But anyway, it wasnt resolved so I uninstalled a bunch of other extensions I had on there (a few other openpose editors, and a few animation extensions).
After removing those other extensions, then re-installing just the openpose and controlnet, Its now working again perfectly. So I guess one of the other extensions was conflicting with something, not sure which one it was. But its resolved.
I may still experiment with the Lora's using DAZ renders though. It seems like it has some potential for something useful, just not sure what that is yet.
You don't have to uninstall problem extensions or delete them. You can just disable them in the GUI. I've had quite a few that have caused problems in the past.
As for SD installation issues, a few months back I was having so many problems I was about to give up on SD altogether.I wanted to delete them anyway because they were redundant (for the most part) of similar extensions, so I kept only the ones I use. If I ever need the other stuff I can always put them back on. I noticed it seems a bit speedier to load the software now, though, so thats a plus.
But good to have openposes working now.
Another question: I know I can add multiple characters poses in one openpose file. But is there a way to specify which person in a scene gets assigned which of the poses?
Not at present unfortunately. I know with assigning faces using Roop you can select character 1 and 2 to separate them by left and right. I've had more success with ControlNet depth maps than Openpose in getting poses to apply to specific characters but it remains very hit and miss. You might have more success with Lora poses than Openpose although that tends to change faces a lot at higher strengths.
Yes, the Loras for poses change the character a LOT, I could be incorrect, but it appears because it bases the results on the physical attributes of the people in the images it was trained with.
Thats why im wondering if DAZ is the way to go to make training images. Make them very neutral with little distinguishing attributes other than the pose. Im wondering if that will force SD to fill in the blank slate of physical attributes based on the prompt moreso than the Lora, since it will be very neutral, and just use the lora just as the base for poses. Or it could have the opposite effect and make everything look weird, or all the same, or some other unknown. Who knows.
- You must be logged in to reply to this topic.
