Amazing comment. Literally scouring the web to find what I am doing wrong and that is the key question I have, since I am doing what I guess is correct but the lora samples come out completely cooked.
yeah im lost here too. This is my first training and training video im following. i have the safetensor files that were generated by the Lora part, i have no clue what to do there.. am i suppose to put those safetensors files under the Lora folder?
It's hard to believe that I used to delete other Loras without thoroughly checking them. And bitch around about having a poor outcome! Thanks buddy, quite helpful
Hi everyone, there is an error when I run the setup.bat file in the kohya_ss directory ( ModuleNotFoundError: No module named 'pkg_resources' ). I have already installed Python 3.10 and visual studio 2015. Any good solutions for it? thank you very much!
@@disposable3167 THANK YOU!!! I had the same issue and this fixed it. I had to manually add it to the PATH in my system settings since I already had it installed but it worked!
Thank you very much! Very useful. Nvidia 3060 does 10 epochs in 10 hours with Network Rank 32 and Network Alpha 16 and 25 - 42 1024x1024 images in the dataset. It's a long time, but it manages.
Putting this out there in case anyone has the same issue that I had and needs a solution. I had python 3.12 installed prior to downloading kohya, then when I was running the setup.bat and it did the "creating venv" thing, it got mad at me that the version of python was not 3.10. So I installed Python 3.10 and made sure to update the PATH in my environment settings but when I tried to run the setup again, it still got mad that Python was incompatible even when running the setup for multiple versions of python being installed... So then I removed Python 3.12 completely, ran it again, and it still got mad saying that no python was found at the python 312 location... Ultimately out of rage, I deleted the kohya folder completely, recloned the repository, ran the setup.bat and it worked! I don't know the technical reason behind what happened but my guess is that something during the initial time that it tried to run remembered the python 3.12 file location and it took a complete restart of the install process to make it forget it and use the python 3.10 location. Currently still installing the files and dependencies but it seems to be working now :)
Can't even get this to start training... error message "image folder does not exist" even though I've selected the source image directory.... Also the new version doesn't have "No half VAE." Would be nice if you had gone more step by step. Will have to search for another video tutorial on this subject.
Thanks for your response, can you tell me where you select the SDXL model in Kohya and which SDXL file you download from Stability's Git repository? @@allyourtechai
@@anovin82 On the main LoRA tab there is the training tab. You will want to select stable diffusion xl base 1.0 for the source model. It should be installed by default but if it isn't, here is the file from the repo: huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0_0.9vae.safetensors
I do have to disagree on the using a similar name. I've done this a few times. Followed the exact steps you do, and gone to the website, found the famous person who looks most like me, used that in the right place - and then every second image comes out as clearly that person and not me, the others a mashup of me and that person which still don't too much like me. (Matt Damon in my case). When I changed the description to a random set of letters and numbers the results were much better. The "infection" that using that name had on the results was way too high in my case. But other than that everything worked really well, so overall I appreciate the video. Thumbs up :)
For some reason, when choosing the sd xl 1.0 base checkpoint, I can't read from the UI the Lora files (which they're .safetensors file). With the other default checkpoint of webui, I could load them. Is this something regarding incompatibility with sd xl 1.0 base and Lora's files versions or something?
I don't think it is compatible. In fact many stable diffusion systems seem to have compatibility issues with 3.11. I always suggest running 3.10 if you can.
I had this issue and was searching here for answers but I just finally stumbled on it in the Dreambooth training tab then saw your question. I found it in one of the horiztonal drop downs under "parameters"
Great tutorial. For me personally, I just completely skip the blip caption, seem unnecessary. I also skip using celebrity lookalike name and just straight up use whatever name I want. For example, I use the name "Adam Iverson" as myself for the trigger word when I trained my own model, I don't bother use the class in prompt like "Adam Iverson man" or stuff like that. I generally use somewhere between 5-10 repeats and then use 10 epoch, the whole training process takes around 30 minutes to 1 hours with my RTX 3090 using the GPU batch size of 6. You can see my result in my profile pictures that I'm using currently.
Looks like a solid result! The training time is impressive I may have to give that a shot. The interesting thing I have found is that most of these settings have very minor impacts on the overall output quality.
@@Majestic_King_Hunter With a 3090 and 42 images I would expect about 10 hours using the default settings. That's going to give you 10 LoRA files though with which you can test. You can modify the epoc's and number of steps for faster model generation at the expense of some quality,
Yeah, i did it with 4gb RTX 3050Ti... it just took a much more longer time, almost 40hours. But then again you can lower your training settings for achieving faster times
i did everything accordingly and whenever i get to BLIP captioning, select directory and name prefix, after pressing captions it doesn't do anything. can yall help me with this?
3:00 i know this is a bit older video but the only time ive noticed changing the learning rate higher is if i move the batch higher. Kohya recommends only to adjust learning rate for every batch you raise, double the learning rate. Batch 1 LR .0001 batch 2 LR .0002 batch 3 .0004. Its the same as training dfl models. I haven't tested any higher than batch 3 but it does help with getting more details
Definitely one of the better tutorials than most out there for sure, thank you! Would the process be the same for training a SD 1.5 LORA? If not then do you have another video/guide I can follow? Thank you again in advance!
Similar, but the settings need tweaking. I’ll do another guide. I should have a video up in a few hours that shows how to train an SDXL Lora on colab for free as well.
Thank you very much, a great video. Finally a more detailed description. I'm going to try it even though I only have a 3050ti/4GB vram, but it seems that the computer also uses normal memory to complement the lack of vram, although it takes longer that way.
If the X/Y/Z sheet resemble the epochs you used and you always go for the once more to left side (so epoch 3 to 5), why do you even set the epochs that high in the first place? Why didnt you go for 6 epochs for example? Is there any quality loss if you go for lower epochs for the results on the left side? (not sure if i made myself clear)
It really comes down to your specific use case. You might use one of the higher epocs if you need a very high quality result that may not be very flexible. If you don't need that, you can get away with stopping the training after the first few epocs.
You are saying once you use Buckets, the resolutions dont matter and also not if they are not cropped perfectly. Fine, but doesnt it say something about "max bucket resolution" in Kohya? And will i benefit in any way from a highe Res image, that i leave on a higher resolution?
Nothing worked for me and it gave me an error "No data found. Please verify arguments" until I clicked the "Prepare training data" button on the Dataset Preparation tab
@@allyourtechai Thank you. I Have another two questions for you: 1) Is it ok that the pack of images of men that you provide looks auto-generated? Or maybe I did something wrong? If I collect real photos, will this improve the Lora training process? Or it doesn't matter for training? 2) when I try to train LORA based on SDXL under Ubuntu 22.04 with my AMD Radeon 6800XT 16GB using your settings and 50 foto, I got an error: torch.cuda.OutOfMemoryError: HIP out of memory. Tried to allocate 2.20 GiB. GPU 0 has a total capacity of 15.98 GiB of which 1.72 GiB is free. Of the allocated memory 13.43 GiB is allocated by PyTorch, and 228.26 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_HIP_ALLOC_CONF I Finally found a solution to deal with this OutOfMemory issue, going through many settings that affect memory consumption. So if anyone encounters the same problems on a AMD RX 6800XT video card and on the Ubuntu operating system, I can post my JSON settings file or the command line that I used to train Lora
1) it doesn't matter for training, it just guides the model, but doesn't impact quality much. 2) It sounds like the system is trying to use cuda even though you have an AMD card. Might be a startup or configuration option that needs to change.
Could you explain a bit more why you set the network rank (dimension) to 256? I know you said it results in better lighting, but I am interested to hear what this setting really does. The default is 8 and on a 3080TI I was looking at very long training times when I started with 256, so I am now trying to start with 8 and going up from there to see how it affects things.
It’s pretty complex under the hood, but basically network rank is the number of features to be trained and controls how they cross layers in the neural network, and network alpha controls how many of the features are applied to the LoRA. Generally higher numbers are better to a point and result in slower training and larger LoRA files. 128 is also a good starting point for characters from what I have seen. Someone here did an in depth test that you can also check out medium.com/@dreamsarereal/understanding-lora-training-part-1-learning-rate-schedulers-network-dimension-and-alpha-c88a8658beb7
When you configured kohya, it's possible you missed a setting? If you have an RTX card you want bf16, and fp16 for older cards. There are other choices as well to select the gpu in the system.
@@allyourtechai I have a 3060ti. Also, what setting could I miss when configuring it? When I run the software I get the following message in the CMD: 19:53:56-550019 INFO Kohya_ss GUI version: v24.1.2 ... 19:54:06-146443 INFO Torch backend: nVidia CUDA 11.8 cuDNN 8700 19:54:06-152433 INFO Torch detected GPU: NVIDIA GeForce RTX 3060 Ti VRAM 8191 Arch (8, 6) Cores 38 So it seems to recognise my GPU...but when I'm training it doesnt seem as though my GPU has any software appearing to take up its processing power in the task-manager. Any ideas?
Hey buddy, I'm having a problem. I follow the entire tutorial, but Kohya's interface looks very different from what's shown in the tutorial, and it doesn't work well. I have the same issue with all tutorials. Maybe there was an update. Do you know anything about this?
@@allyourtechai Looking forward to this video, seems impossible to get kohya installed at the moment. I've tried everything including docker images. Really frustrating software. Thanks for your videos
What Class Prompt would a pixar like character be called? I am training to train a little 8 year old girl for a project. Do I write Girl as a Class Prompt, or do I write Cartoon Girl? Then for the regularization images, do I choose pictures of a little girl or pictures of a bunch of pixar girls?
You don't want something too specific, but it should be descriptive. I would try pixar girl since there are lots of reference images in the existing models for pixar characters.
I just installed this and I had to go put the Kohya local url in my browser to start it. It's those little things you just don't think of that other bats just auto run like with comfy, auto 1111, Fooocus and you expect them just do. I know your video was 4 months ago so I'm not sure if somethings changed or it's my install but I have to do this. You didn't show you needed to. My install would not let me start kohya in the browser with the #5 option. It kept making me go through the multiple options of the install. I actually was stumped for a bit on why it wasn't starting. I'm actually not sure if it shouldn't just actually be starting without me having to put the url in my browser???
@@allyourtechai I do have one followup question. I'm seeing a lot of the images seem to look closer to the celebrity I used in the class prompt for training the LoRA than they do myself. Things like eye color stand out especially. Do you ever use something aside from the celeb name to get it to more closely resemble yourself?
@@nessdevelopment8779 in some cases where the celeb doesn't have many pictures in the base model, it can get off track a bit. The other thing I try to do is be specific in my text annotations when training the LoRA. tag the images with things like eye color, hair color, articles of clothing, etc. YOu can use those keywords later when creating images to get those finer details.
Also I want to make a lora for a symbol which could be used in multiple jewelry, emblems or logo images, problem is that this symbol does not have many pics, but I do have a 2D black and white picture and I have fusion 360, could I make images of a 3D render in every angle and use that as training data? I think I just answered my own question, but hey you can use the idea for a video of you like
Yes, there are a lot of programs out there.That can do LORAS when it comes.To faced likenesses and body types.But when it comes to detailed superhero costumes.LORAS fall short, what if I want a particular costume on a likeness?Then what do you do
hello everybody. I'm planning to switch from mac to windows. In the meantime can I use Mac to do this? do you know a tutorial for that? thanks a lot in advance
Omg, I've spent an hour and a half following this tutorial, I even paid for the patreon subscription to download the files, and in the end I wasn't able to run the training, it just didn't start and the CMD black window just said it was an error with kohya thing... Is there any other way totrain a lora that doesn't take 10 hours just to make 10 epochs ?I've already tried One Trainer and it spent like 12 hours training 1200 epochs on a RTX 4090, but the results were still being disgusting when trying to make a full body portrait but amazing making close-ups. Has anyone experienced something similar?
@@allyourtechai it says "RuntimeWarning: invalid value encountered in scalar divide ret = ret.dtype.type(ret / rcount) mean ar error (without repeats): nan No data found. Please verify arguments (train_data_dir must be the parent of folders with images) / 画像がありません。引数指定を確認してください（train_data_dirには画像があるフォルダではなく、画像があるフォルダの親フォルダを指定する必要があります）"
setup.bat does nothing. In CMD it just has no effect as if it's an empty file, but it's not empty. Clicking on it in Windows has no effect either. Should I give up? Edit: even though I have 4 other AI tools installed and working fine, I now tried to install Python again, and this time setup.bat worked. Very weird.
im so confused at the installation part. When I push in "1" I just installed. It never asked me anything else. Now I don't even know if its using my GPU for training for not.
If you haven't figured this out already look and see if an option says "Manually configure Accelerate" if so select it and it will ask you all the questions shown in this tutorial.
It is giving me an error that a new release of PIP is available, however I ran the code and it confirms I have the latest version, 24.0. Also, it tells me "ERROR: file:///D:/kohya_ss/sd-scripts (from -r requirements.txt (line 35)) does not appear to be a Python project: neither 'setup.py' nor 'pyproject.toml' found." So I am not sure what to do...
[notice] A new release of pip available: 22.3.1 -> 24.0 [notice] To update, run: python.exe -m pip install --upgrade pip 20:06:05-937510 INFO Requirements from requirements_pytorch_windows.txt installed. 20:06:05-938510 INFO Installing requirements from requirements_windows.txt... Obtaining file:///D:/kohya_ss/sd-scripts (from -r requirements.txt (line 35)) ERROR: file:///D:/kohya_ss/sd-scripts (from -r requirements.txt (line 35)) does not appear to be a Python project: neither 'setup.py' nor 'pyproject.toml' found. [notice] A new release of pip available: 22.3.1 -> 24.0 [notice] To update, run: python.exe -m pip install --upgrade pip 20:06:06-379616 INFO Requirements from requirements_windows.txt installed. 'accelerate' is not recognized as an internal or external command, operable program or batch file.
This is a great tutorial and I got fantastic results on a first try! I have one question though - I've noticed that you are training on base SDXL model. Is there any particular reason for that? There are some decent looking SDXL models like Juggernaut or RealVis so I wonder what motivated your choice.
Just keeping things simple for the tutorial, but you bring up a fantastic point. There are some XL models that are far better than base, and I definitely encourage people to experiment with them. I’ll probably do a follow up with some of my favorites
You can for example train on juggernaut if you really like the model, but keep in mind that you could run into issues if you want to use the lora with other models. That's why I usually train on the base model, the results are more flexible.
@@allyourtechaiJuggernaut is practically the standard now. It's sponsored by some AI company and gave the best face similarity result metrics for IPAdapter tests. I'd say it's the best refined model.
Intel i9 13900k with rtx 3080 , training 15 high res images, used your guide and got going spent little less than 24 hours on first stage of this training. now at second stage in training, second last line in cmd says epoch 1/10, last line steps 0% 4/3000 [10:41:44
When you pull up windows performance manager (control, alt, delete), how does the memory usage on your card look? My hunch is you don’t have enough vram for those settings. These are most important settings to lower VRAM memory for SDXL training: * Choose either Adafactor or AdamW8Bit as optmizer * Train batch size 1 * Memory efficient attention checked * Gradient checkpoint checked * Max resolution: 1024,1024 * Enabled buckets checked, Minimum bucket resolution: 64 , maximum bucket resolution: 1024 * Don't upscale bucket resolution unchecked, Bucket resolution steps 64 * No sample preview or limited sample previews 768x768 max * xformers or sdpa
Did the adjustments, and lowered also number of photos used untill I get the settings right. Still I dont see any spikes in GPU load when I start over. @@allyourtechai
Not much, started over with just 5 photos yesterday night , now its done with first part. I see just low % on both GPU and CPU. I could send you the screenshots of cmd text of running process @@allyourtechai
Warning: Python version 3.10.9 is required. Kohya_ss GUI will most likely fail to run. "Installing packaging python module..." Requirement already satisfied: packaging in d:\ai\lora training\kohya_ss\venv\lib\site-packages (24.0) File "D:\Ai\Lora Training\kohya_ss\setup\setup_windows.py", line 7, in import setup_common File "D:\Ai\Lora Training\kohya_ss\setup\setup_common.py", line 11, in import pkg_resources ModuleNotFoundError: No module named 'pkg_resources' i am getting above error. how do i fix it? I am on windows and i have installed python 3.12, my gpu is rtx 2060 mobile.
The background screen is extremely small, nothing shows on it. You are using an old Kohya distribution, the new one does not have the same tabs and options.
@@allyourtechai Thanks for your answer! I somehow managed to fix the pb, but I have an other.. at 13:00, after I followed all your instructions, when I click on "Start training" I have 2 error messages : "python stopped working" and then "Error. Connection errored out". Would you know how to fix it? I tried putting epoch on 1, resolution on 512,512, network rank on 36, and still the same error
I was able to successfully complete everything in the guide here but the images I’m getting do not really look like the source images I provided. Any tips/insights would be appreciated
You have to really play around with the Lora's sliders of weight/strength, & figure out witch Lora looks the best like he did at 15:59. Even at the end of this tutorial his final image doesn't really look like him. Personally I've found that img to img works the best with Loras when trying to create look a like photos
Reporting back on this, I think you may have missed some more critical steps. I just trained a model for 4 hours and when I put it in the Automatic1111 lora folder, it doesn't show up. When I force it to always show networks in the settings, the lora files show up but literally do nothing when added to the prompt.
Thanks for the video! I followed up with all the steps, now waiting to see the results! Is there anything that has changed since the video was released? should I enable more options on a 3090? I currently at 19Gb Vram, some are spare.
Great question. I should have clarified that further. For me specifically I train the LoRA on images of myself for RU-vid thumbnails, and with 10-20 images epoch 3 or 4 works great. If I were training a style or training with 100 images, I may want to use a different epoch. You can definitely save time by cutting the training in most cases, but depending on what you train you might use a later epoch (hopefully that all made sense)
I got an error after running the GUI Bash FIle: "ImportError: cannot import name 'set_documentation_group' from 'gradio_client.documentation' (C:\StableDiffusion\kohya_ss\venv\lib\site-packages\gradio_client\documentation.py)" Don't know that to do. I have an rtx 4090 GPU
Amazing guide, subscribed! Thank you. I was wondering if you have a related tutorial about inserting a person's LoRA into photorealistic scenes, like on a boat, mountain climbing, at a party, doing other cool stuff, etc. That would be really neat, could be an add-on to this! Cheers
Following your tutorial I was able to generate 10 lora files, but each of them is not working when I try to generate images. Below the generated image there is a message "Networks not found: 1" All the other Loras I use work fine. I searched online but couldn't find this error anywhere, apart from a reddit post with the same issue from 4 months ago with no answers. I would appreciate if you could help. I run an RTX2080ti if that matters and I used fp16
I have not seen that before, but searching around as well. Are you using automatic1111? Have you tried the Lora in something like foocus? Just to narrow it down to the Lora specifically
Unfortunately I don’t think it will be possible. You can train lower resolutions on 8GB and it’s possible to train an SDXL with 12GB, but anything less will be a problem.
@@allyourtechai Thanks a lot for the info! I really needed this reply because I was about to install it tomorrow, but your warning came just in time! Also. I've read online that google colab can train loras for low vrams. Do you recommend I try that, or is it just a waste of time?
is having 64 images for training too much? I'm using that many cause they're not as good quality of the subject or they have glasses on or bangs, hair etc covering the face
I typically find that 10 or so images are enough. Higher quality, higher resolution would be the best choice. Not always possible of course in which case more images may help while also providing a more flexible model. Make sure the elements that are different are in the annotations associated with each image. If they have bangs in some images, make note of that so you can later say “xyz person with bangs standing on the beach”
I followed your instructions and when I clicked on Start Training I am getting the following error right at the start and the training stops kohya_ss\venv\lib\site-packages umpy\core\_methods.py:129: RuntimeWarning: invalid value encountered in scalar divide ret = ret.dtype.type(ret / rcount) INFO mean ar error (without repeats): nan train_util.py:856 ERROR No data found. Please verify arguments (train_data_dir must be the train_network.py:213 parent of folders with images) / any insights?
The error message `ImportError: cannot import name '_imaging' from 'PIL'` typically occurs when there's an issue with the Pillow library in Python, which is a fork of the PIL (Python Imaging Library) used for opening, manipulating, and saving many different image file formats. This error can happen due to several reasons: 1. **Incorrect or Partial Installation**: The Pillow library might not be installed correctly or completely. This can happen if the installation process was interrupted or if the wrong version of Pillow was installed for your Python version. 2. **Environment Path Issues**: There might be a problem with your Python environment paths, where Python is not able to find the installed Pillow library. 3. **Conflicting Libraries**: If you have both PIL and Pillow installed in the same environment, they might conflict with each other. To resolve this issue, you can try the following steps: 1. **Reinstall Pillow**: Uninstall and then reinstall the Pillow library. You can do this via pip: ```python pip uninstall Pillow pip install Pillow ``` 2. **Check for Conflicts**: Make sure you don't have PIL installed in the same environment as Pillow. If you do, remove PIL: ```python pip uninstall PIL ``` 3. **Verify Installation**: After reinstalling, you can check if Pillow is installed correctly by importing it in a Python shell: ```python from PIL import Image ```
Creating venv... Python was not found; run without arguments to install from the Microsoft Store, or disable this shortcut from Settings > Manage App Execution Aliases. The system cannot find the path specified. The system cannot find the path specified. Python was not found; run without arguments to install from the Microsoft Store, or disable this shortcut from Settings > Manage App Execution Aliases. The system cannot find the path specified. I keep getting this everytime I try to run the setup.bat file. why?
Do you have any stable diffusion software running on your pc? It doesn’t sound like it, so one of the items you need to install is python. I would just setup invoke ai, automatic1111, or foocus to generate stable diffusion images, and any of those 3 should also help get python installed. I’m also happy to do a tutorial
When you install Python, there is a checkbox ( I think its somthing like "add pyhon to path" ). You need to check this one so all programs can find your python instalation.
@@allyourtechai Many thanks. I have one PC with Ryzen 2700X and Nvidia RTX2070 and one PC with Ryzen 5600 and AMD RX6800. But somehow I could not run on both, need to check what of the requirements is missing. There is always a Powershell error for not finding something..
@@GiasticeDid you manage to start the training? I have rx 6800 xt and 1 iteration takes about 5 minutes, the program does not use the graphics card at all. Somewhere I'm missing a line of code or checking the option to use AMD. I even turned off xformers and still nothing.
Hey i followed every step but i get the error upon clicking training, 'No data found, please verify arguments (Train_data_dir must be the parent of the folder with images
Check in the video description. I have a couple common errors and solutions toward the bottom. In this case I believe it is the button to prep the training folders and copy them over to the other screen.
Thank you for your proactive response to subscribers... I'll try your method soon... but I'd like to know if there's a way to make a personalized "checkponit" with personal photos for Comfy to use... I have several photos of places in mine city you would like to use as a base; Can you point me to a video? How are "checkpoints" created?
These are most important settings to lower VRAM memory for SDXL training: * Choose either Adafactor or AdamW8Bit as optmizer * Train batch size 1 * Memory efficient attention checked * Gradient checkpoint checked * Max resolution: 1024,1024 * Enabled buckets checked, Minimum bucket resolution: 64 , maximum bucket resolution: 1024 * Don't upscale bucket resolution unchecked, Bucket resolution steps 64 * No sample preview or limited sample previews 768x768 max * xformers or sdpa
Thank you! Wish GUY hadn't moved things around so much since your video. i'm having a hard time finding the settings you are modifying....maybe if it's worth your time you could do an updated video someday? cheers
Great video, but one thing I don't understand: You say to use a celebrity lookalike for "guidance parameter", but in the end, the imagine you created was a blend of yourself and Tom Cruise. What if you want to create images that really look like yourself? Does it work without "guidance parameter" and just use your own name? thx
It would be more like your own image since this is locally trained and your LORA tell SD that the image you provided is how tom cruise looks like. Its other tom cruise images that it knows from is original training will be used automatically as guidance parameters in terms of perspectives, image tones, color grading, photography style, etc.
@@allyourtechaisorry for my late reply, yes only this message but I have another problem... if I try to select the pictures, they are not shown to me in the folder. idk what to do :/
Not with SDXL, but you could do a 512x512 resolution LoRA for stable diffusion 1.5 or 2.1. With up scaling you can still achieve some really great results.
Hello. Your uploaded video lectures have been very helpful. Thank you. I have a question about the video. Is creating a checkpoint model the same as with LoRa? If there are slight differences, what are they? Also, do you have plans to create a tutorial on creating checkpoint models?
A checkpoint is also called a model (SDXL is a model), so you can think of these as the very large base models that we use to generate images. A LoRA modifies the output of a model, so it’s much smaller and easier to train, plus it has the flexibility to run on top of multiple models. What is the outcome you are hoping to achieve? From there I can make some suggestions, but typically a LoRA is what you want.
@@allyourtechai I want to create my own character with flexibility using a checkpoint model. I prefer the extensive flexibility that checkpoint models offer over LoRA, so I'm trying to create one. However, I'm facing various difficulties in the training setup using Koyha.
I am looking for help getting started with training an image-to-image AI. I want something where someone can draw a simple image and have it convert to a map.
You might be able to use Dall-E for that. I have seen a few customGPT's that convert a drawing into an image, so i'm sure a custom prompt could help create a map from a sketch as well.
Brian, thanks a lot for this awesome tutorial, in about 25 hours my GPU will provide me with 10 new 1.7G models. Can pls give us the links for the great photos of yours you show us here. I'd love to create similar photos of myself. Again, thank you and pls keep up the good work.
Thanks for the great video! A lot of the results that i'm getting by following your method resemble the celebrity look alike or resemble the civit prompt more-so than the images of myself. I will continue experimenting of course. But I was wondering if you had any tips to make the images resemble the subject more? Additionally, is there a way to generate multiple images with a different variation or seed for each Lora model? Also I noticed that the "No-Half VAE" option was not present in my UI.
Couple ideas. In automatic1111 when you are doing the X plotting, instead of setting a static seed just ensure -1 is in the seed box. This will get you a random seed for each Lora file/image. You can also open the generated text files from the auto captioning step and find key words to apply in order to get results that resemble the original model more. For mine I use “bald” in my prompt since I have a shaved head. Some of the civitai prompts definitely work better than others though, so definitely experiment. I typically find the my 4th Lora seems to offer a good blend of precision and flexibility.
it still works for sdxl training with a 1060 6gb but it takes 5 hours for 100 steps so really an overnight thing and only 512x512 but its still better quality looking than sd1.5
I heard people doing like up to 100 repeats for a character. Also they said you need the exact amount of reference pictures as the amount of steps you use. Why that is so stays a mystery to me, too.
YOu can increase the repeats, but it also increases the training time assuming all other settings remain the same. I haven't seen a large enough difference to justify the added training time.
im having this issues when clicking start training ERROR The following folders do not match the required pattern _: ERROR Please follow the folder structure documentation found at docs\image_folder_structure.md ... ERROR The following folders do not match the required pattern _: ERROR Please follow the folder structure documentation found at docs\image_folder_structure.md ...
If you click on the TOOLS tabs and fill in the info there (location dataset images, reg images if you have any and then enter a folder in the DESTINATION TRAINING DIRECTORY box and click PREPARE TRAINING DATA it'll create the right structure and move all the files into there for you. Then click COPY INFO TO FOLDERS TAB to make sure you have the right info in the right place.
