@@invokeaiwow nice thanks! I’d be happy to pay for the app as a local run option with premium features. I know it’s hard to do without it just being hacked or spread for free but the cloud option is super expensive and probably not fast and we like local models. If you can figure out a way for $15 a month with maybe some support or $25 with premium support per user, would be good. Appreciate all the free local functions
I would love to see this same video as a do-over for training a concept like hand gestures or a consistent character. And including best practices for great results. My experience with both using pre-made and self-generated LoRAs has been totally hit and miss, and the problem is that there’s just no consensus on how to prepare data set images and captions and what settings to use for training. For example, on my first SDXL LoRA using Invoke Training, I had no idea what settings to use (or really what those setting even meant). I resorted to combing Reddit and TH-cam and using what the most knowledgeable content creators consistently recommended. But those settings mostly come from tools like Kohya_SS (which wouldn’t work on my PC for some reason) and they don’t exactly port over well to Invoke training. So I ended up producing a lot of LoRAs that produced garbage or really inconsistent images without any seeming rhyme or reason (like 1 great image in every 10 that really nailed what the LoRA was created for at totally different and often opposite generation settings).
Really useful. Thank you. Would it be possible to include auto-caption on the free community training scripts and improve the UI a bit? I had to install separate caption generators and they really bloat the size of data I use for AI stuff. Like 6-12gb per add on. And they’re not great, I would prefer to stay within the InvokeAI environment.
What logic do you use when deciding how broad your lora should be? For example, if you wanted to train a set of poses, costumes, and facial expressions to be interchangeable amongst a set of characters: -would you train a separate lora for each character for consistency and prompt for them when generating? Would the logic be the same if multiple characters need to interact in a scene? -Would you train a specific lora for each pose (e.g. "sitting"), or train all conceivable poses in one lora? -Would you train costumes and facial expressions together with poses, or train everything separately?
It really depends on what you want it to be able to do. Perfectly replicate a character with accuracy in an setting you prompt for is going to be a lot more than just generally influencing the style or look and feel of your images.
@invokeai Thanks! That's also afraid of. I want to be able to make my own custom characters and then be able to tell the AI the situations they might be in or use them with art software to help use it to draw my characters and then have AI kind of help adjust the characters to look good. I think I may be a little too ambitious at the moment. I'm sure I'll be able to do what I'm trying to do eventually.
The Model Trainer is only available on the Professional edition. We do have the training scripts available open source if you want to run them yourself: github.com/invoke-ai/invoke-training.
I started the free 10 day trial to test this out and I can't access the "concept training". It only gives me 2 options "upgrade now" and "learn more". Even after upgrading, I can't access the "concept training" tab.
Wow! Thank you so much for this and thank you for educating us!
That's great and all, but please let us do it locally as well. 😊
github.com/invoke-ai/invoke-training
There is no free local option for this invoke trainer?
github.com/invoke-ai/invoke-training
@@invokeaiwow nice thanks! I’d be happy to pay for the app as a local run option with premium features. I know it’s hard to do without it just being hacked or spread for free but the cloud option is super expensive and probably not fast and we like local models. If you can figure out a way for $15 a month with maybe some support or $25 with premium support per user, would be good.
Appreciate all the free local functions
I would love to see this same video as a do-over for training a concept like hand gestures or a consistent character. And including best practices for great results. My experience with both using pre-made and self-generated LoRAs has been totally hit and miss, and the problem is that there’s just no consensus on how to prepare data set images and captions and what settings to use for training. For example, on my first SDXL LoRA using Invoke Training, I had no idea what settings to use (or really what those setting even meant). I resorted to combing Reddit and TH-cam and using what the most knowledgeable content creators consistently recommended. But those settings mostly come from tools like Kohya_SS (which wouldn’t work on my PC for some reason) and they don’t exactly port over well to Invoke training. So I ended up producing a lot of LoRAs that produced garbage or really inconsistent images without any seeming rhyme or reason (like 1 great image in every 10 that really nailed what the LoRA was created for at totally different and often opposite generation settings).
Is it possible to locally train ?
Really useful. Thank you. Would it be possible to include auto-caption on the free community training scripts and improve the UI a bit? I had to install separate caption generators and they really bloat the size of data I use for AI stuff. Like 6-12gb per add on. And they’re not great, I would prefer to stay within the InvokeAI environment.
Any news on Text2Video generation? Or does invoke have no plans to implement that?
What logic do you use when deciding how broad your lora should be?
For example, if you wanted to train a set of poses, costumes, and facial expressions to be interchangeable amongst a set of characters:
-would you train a separate lora for each character for consistency and prompt for them when generating? Would the logic be the same if multiple characters need to interact in a scene?
-Would you train a specific lora for each pose (e.g. "sitting"), or train all conceivable poses in one lora?
-Would you train costumes and facial expressions together with poses, or train everything separately?
This is very confusing… you totally forgot to indicate how to access this!!
Thanks for this!
So when training a lora on a character, how big should the data set be?
It really depends on what you want it to be able to do. Perfectly replicate a character with accuracy in an setting you prompt for is going to be a lot more than just generally influencing the style or look and feel of your images.
@invokeai Thanks! That's also afraid of. I want to be able to make my own custom characters and then be able to tell the AI the situations they might be in or use them with art software to help use it to draw my characters and then have AI kind of help adjust the characters to look good.
I think I may be a little too ambitious at the moment. I'm sure I'll be able to do what I'm trying to do eventually.
Is this also possible locally?
We publish our training scripts open source here github.com/invoke-ai/invoke-training but the UI/wizard is only available in the cloud paid version.
tab doesn't seem to exist on local...
The Model Trainer is only available on the Professional edition. We do have the training scripts available open source if you want to run them yourself: github.com/invoke-ai/invoke-training.
Wait. I just installed Invoke-Training, and it looks nothing like this. Is this the commercial version, and I’m using the free community scripts?
I started the free 10 day trial to test this out and I can't access the "concept training". It only gives me 2 options "upgrade now" and "learn more". Even after upgrading, I can't access the "concept training" tab.
Hi Christopher - The concept training tab is available on the Premier tier. Are you now on a full Indie tier or did you upgrade to Premier?