Alvaro Somoza (@OzzyGT) profile picture

ML Engineer @HuggingFace

Chile
Joined March 2009
Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)Photo by Alvaro Somoza (@OzzyGT)
II wrote a guide on making your own model integrations with Modular Diffusers using Claude. Why would you need this? So you can run all your models in one environment with no dependency conflicts, and use the diffusers optimizations. huggingface.co/blog/OzzyGT/u…
🤖 Made with AI
1
1
5
225
MiniMax-H3-Image has one of the coolest prediction previews I've seen, they transition from a drawing to a real photo almost for all the generations. At step 4 you can get a cool artistic image of the final image.
1
112
I did the SDNQ quantizations of MiniMax-H3-Image so you can use it with diffusers. This model is pretty good, I found that pretty much you have to use it with 16MP to get a good result. And as always, 8bit is practically lossless.
1
101
4bit is pretty good too
1
28
After doing a custom modular diffusers integration of MotionBricks, which I want to use for doing a smooth transition between character motions, I couldn't resist to do a PoC game to show what this model can do in realtime inside a game.
1
2
94
I'm making a series of tutorials and guides for using modular diffusers with different models to make PoC games, since most of the ones I see are just low effort copies, I want for people that want to learn how to use different tools to build their game and not just let the AI do it all. Also I highly recommend to use human taste for the art direction because it makes a big difference. One of this models that it's really underrated is kimodo soma from @nvidia which can let you do all the motions you need for your game.
1
212
for Moss-Image by @Open_MOSS I did a quick test using a reference image so I didn't have to write the prompt, same as krea-2, and it works pretty good with diffusers, now I don't have to write most of the prompt if I want to use another image as a base
1
179
Alvaro Somoza retweeted
FLUX 3 Image Klein is here. Open Weights with the quality of a large frontier model. Customizable and fast. 5.6B parameters and runs on a single consumer GPU. Native 4K, bounding boxes, and pixel-perfect editing.
178
397
224
3,308
345,780
Another pretty good use for Minimax-H3 I found is generating SFX for games. You can feed it a clip of whatever you want the SFX for, and let the model generate the sound for it
2
192
So I needed a model that could position each part I want in a specific position, like the engines and mounted guns because for a game you need this, there is currently only one model that can do this. I still don't quite understand why people don't like bounding boxes...
1
118
I posted that you can extend videos with Minimax-H3, but did you know you can do the same with an image? It doesn't need to have the same aspect ratio. You can use any image, mask it, and let the model outpaint the rest. And this model is damn good at it.
2
421
In case you didn't know, Minimax-H3 can spatially outpaint videos without any LoRA, you just need to use masks. If this isn't something everyone knows and you want to see how to do it with Diffusers, let me know and I can make an example.
1
2
228
But this isn't limited to spatial outpainting. You can also do temporal outpainting in the same pass, extending the video both spatially and in time, while also generating the soundtrack for the clip.
1
36
Following the prediction previews I did for Krea-2, I did the same for Minimax-H3 as a video example. This can be useful if you have a clear composition in mind, since you can cancel the generation early if it doesn't match what you want. huggingface.co/OzzyGT/minima…
1
93
another example
1
22