Hi! Nice video, thanks! Can you tell me if there are any ideas? I generate a picture, then through "inpaint + fill + only masked" add a new object, is there any way to quickly and accurately pull the same object in a separate picture on a white background ? The main idea, to adapt SD to generate game art for the hidden objects genre.
this video is incredible!. I am looking to create a diablo-2 2D like game in godot engine, using isometric style tiles and stuff. And I was wondering If stable diffusion could help me to create most of the art for the game and seems like it can be possible!
Wait, there's a method to do character animation too. But I would have to test it to make sure it can maintain the same character and its style. Bro if only these tools existed 2 years ago when I was working on my independent project that flopped. Whelp, never too late.
I did this locally using github.com/bmaltais/kohya_ss and a RTX3090 which is good if you are a software engineer or highly technical. But the easier way is using Colab (which I did a video on th-cam.com/video/aNthC5r7Pso/w-d-xo.html) or Automatic1111 Dreambooth extension.
@@pjgalbraith What asset of training images did you use to get such a great result? Unfortunately, from this your video (th-cam.com/video/aNthC5r7Pso/w-d-xo.html) it is not completely clear the whole learning process, which led to high-quality generated images.
@@oleggergart8260 In the video above I show the source images I used as well as the recommended training settings. When it comes to style training it can take time to find the right dataset. The model I used in the video was trained against the Stable Diffusion 2 768 model. But that is a harder model to train against and needs more vram.
Good question. For fine-tuning I used Midjourney AI v4 and some images from Pinterest to teach it the general style. For a commercial project I would recommend spending more time creating the dataset manually to get a more unique style.
@@pjgalbraith I've had a hard time figuring out how to train it. I have several top-down images but of different objects and scenarios. Dreambooth allows very little input information about the images, e.g. just a specific prompt and a broad one to compare them to. How do I describe the various images individually, or get it to just understand what "top-down" means in general?
That kind of style training can be harder, you may need a larger set of images of people, animals, objects, and buildings. Around 100+. The other option is something like EveryDream which supports larger labelled datasets.
BTW - could you consider checking your audio levels? Music is annoyingly loud, while your speech is really quiet at the same time. Have to adjust volume every time you stop speaking.
Wait im confused.. did you already have a trained dreambooth model of the isometric artstyle and using it in this video? Cause your video title seems to suggest on how to train a dreambooth model using isometric artstyle
So glad someone else remembers Chaos Lords. Such a crazy but interesting idea.
I love this, you just gain a new sub. Thanks for the great content!
Awesome, thank you!
The video is absolutely mind blowing!
Syndicate meets Industries of Titan. Nice!
SD is amazing. But ultimately dedication and better artistic skills Ultima yield the best results and you showed that here. 👍
Thank you so much 😀
Not to mention no way to animate.
Hi! Nice video, thanks! Can you tell me if there are any ideas? I generate a picture, then through "inpaint + fill + only masked" add a new object, is there any way to quickly and accurately pull the same object in a separate picture on a white background ? The main idea, to adapt SD to generate game art for the hidden objects genre.
Amazing!! thanks for sharing :)
Glad you liked it!
What's the song at 3:20?
Good job, TY
Very nice.
Crazy how much faster you can generate assest instead make them manually.
could you please make a tutorial for new dreambooth ui? thanks nice content
Could you please share the link for this model you created? I really like your isometric model
Not sure if I'll get around to releasing it but you can try out MirageML's model which is very similar huggingface.co/MirageML/lowpoly-cyberpunk
@@pjgalbraith thank you
this video is incredible!. I am looking to create a diablo-2 2D like game in godot engine, using isometric style tiles and stuff. And I was wondering If stable diffusion could help me to create most of the art for the game and seems like it can be possible!
Yeah it's a good fit for static assets like that.
Wait, there's a method to do character animation too. But I would have to test it to make sure it can maintain the same character and its style. Bro if only these tools existed 2 years ago when I was working on my independent project that flopped. Whelp, never too late.
Patrick, can you give a link to a video where you show how you trained Dreambooth for this game? Thanks
I did this locally using github.com/bmaltais/kohya_ss and a RTX3090 which is good if you are a software engineer or highly technical. But the easier way is using Colab (which I did a video on th-cam.com/video/aNthC5r7Pso/w-d-xo.html) or Automatic1111 Dreambooth extension.
@@pjgalbraith What asset of training images did you use to get such a great result?
Unfortunately, from this your video (th-cam.com/video/aNthC5r7Pso/w-d-xo.html) it is not completely clear the whole learning process, which led to high-quality generated images.
@@oleggergart8260 In the video above I show the source images I used as well as the recommended training settings. When it comes to style training it can take time to find the right dataset. The model I used in the video was trained against the Stable Diffusion 2 768 model. But that is a harder model to train against and needs more vram.
Do these generators understand tilt shift? Might be cool.
Yeah they do but I'd do it as a post processing effect
Patrick, did you create all those images you used to train the model?
Good question. For fine-tuning I used Midjourney AI v4 and some images from Pinterest to teach it the general style. For a commercial project I would recommend spending more time creating the dataset manually to get a more unique style.
price of chest assets just went down by 90% x)
Yep. Amazing chest ahead! Is what I should have titled it.
"ANY" can't do top-down
Not sure what you mean. You can do top down it just requires fine-tuning a model.
@@pjgalbraith I've had a hard time figuring out how to train it. I have several top-down images but of different objects and scenarios. Dreambooth allows very little input information about the images, e.g. just a specific prompt and a broad one to compare them to. How do I describe the various images individually, or get it to just understand what "top-down" means in general?
That kind of style training can be harder, you may need a larger set of images of people, animals, objects, and buildings. Around 100+. The other option is something like EveryDream which supports larger labelled datasets.
You can also try asking on the Dreambooth discord style channel for advice.
@@pjgalbraith Thank you, I'll look into both of those
BTW - could you consider checking your audio levels? Music is annoyingly loud, while your speech is really quiet at the same time. Have to adjust volume every time you stop speaking.
I saw your last comment and turned the music down a lot this time. I'll look into what might be going on.
Your audio is fine
Wait im confused.. did you already have a trained dreambooth model of the isometric artstyle and using it in this video? Cause your video title seems to suggest on how to train a dreambooth model using isometric artstyle
Yeah I explain in the video how I trained it and the dataset I used.
I did another video specifically about DreamBooth training if you have any questions let me know.