Everything I'm finding is about training to your data set--which is not what I'm after. I just want the nature of the subject of the picture. Must run locally (the images do not leave the computer) and automated as there are upwards of 50,000 images to deal with.
-
Features
-
Friends of IIDBFriends Recovering from Religion United Coalition of Reason Infidel Guy
Forums Council of Ex-Muslims Rational Skepticism
Social Networks Internet Infidels Facebook Page IIDB Facebook Group
The Archives IIDB Archive Secular Café Archive
You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
Image classification programs
- Thread starter Loren Pechtel
- Start date
Jarhyn
Wizard
- Joined
- Mar 29, 2010
- Messages
- 15,532
- Gender
- Androgyne; they/them
- Basic Beliefs
- Natural Philosophy, Game Theoretic Ethicist
So... Clip?Everything I'm finding is about training to your data set--which is not what I'm after. I just want the nature of the subject of the picture. Must run locally (the images do not leave the computer) and automated as there are upwards of 50,000 images to deal with.
This is a pre-trained part of most genAI models, the CLIP layer.
Most models have such a layer built in. Flux uses something different (better), I think, but by in large you want to look into local "captioning" models (T5).
You're probably going to want something trained on the image subgenre you are trying to positively/negatively ID on specifically.
I think a lot of people like DeepBoru too.
What is your use case?
Last edited:
Sorting out things that might be relevant from the trash that accumulates on a phone. Even person/animal/thing sorting would be quite useful.So... Clip?Everything I'm finding is about training to your data set--which is not what I'm after. I just want the nature of the subject of the picture. Must run locally (the images do not leave the computer) and automated as there are upwards of 50,000 images to deal with.
This is a pre-trained part of most genAI models, the CLIP layer.
Most models have such a layer built in. Flux uses something different (better), I think, but by in large you want to look into local "captioning" models (T5).
You're probably going to want something trained on the image subgenre you are trying to positively/negatively ID on specifically.
I think a lot of people like DeepBoru too.
What is your use case?
Jarhyn
Wizard
- Joined
- Mar 29, 2010
- Messages
- 15,532
- Gender
- Androgyne; they/them
- Basic Beliefs
- Natural Philosophy, Game Theoretic Ethicist
Yeah, CLIP, DeepBoru, or T5 I think would be your best betSorting out things that might be relevant from the trash that accumulates on a phone. Even person/animal/thing sorting would be quite useful.So... Clip?Everything I'm finding is about training to your data set--which is not what I'm after. I just want the nature of the subject of the picture. Must run locally (the images do not leave the computer) and automated as there are upwards of 50,000 images to deal with.
This is a pre-trained part of most genAI models, the CLIP layer.
Most models have such a layer built in. Flux uses something different (better), I think, but by in large you want to look into local "captioning" models (T5).
You're probably going to want something trained on the image subgenre you are trying to positively/negatively ID on specifically.
I think a lot of people like DeepBoru too.
What is your use case?
Which one should I try first?Yeah, CLIP, DeepBoru, or T5 I think would be your best betSorting out things that might be relevant from the trash that accumulates on a phone. Even person/animal/thing sorting would be quite useful.So... Clip?Everything I'm finding is about training to your data set--which is not what I'm after. I just want the nature of the subject of the picture. Must run locally (the images do not leave the computer) and automated as there are upwards of 50,000 images to deal with.
This is a pre-trained part of most genAI models, the CLIP layer.
Most models have such a layer built in. Flux uses something different (better), I think, but by in large you want to look into local "captioning" models (T5).
You're probably going to want something trained on the image subgenre you are trying to positively/negatively ID on specifically.
I think a lot of people like DeepBoru too.
What is your use case?
Jarhyn
Wizard
- Joined
- Mar 29, 2010
- Messages
- 15,532
- Gender
- Androgyne; they/them
- Basic Beliefs
- Natural Philosophy, Game Theoretic Ethicist
Clip is easiest, T5 would be harder but give the best captions, I think. Maybe Forge would be a good place to start?Which one should I try first?Yeah, CLIP, DeepBoru, or T5 I think would be your best betSorting out things that might be relevant from the trash that accumulates on a phone. Even person/animal/thing sorting would be quite useful.So... Clip?Everything I'm finding is about training to your data set--which is not what I'm after. I just want the nature of the subject of the picture. Must run locally (the images do not leave the computer) and automated as there are upwards of 50,000 images to deal with.
This is a pre-trained part of most genAI models, the CLIP layer.
Most models have such a layer built in. Flux uses something different (better), I think, but by in large you want to look into local "captioning" models (T5).
You're probably going to want something trained on the image subgenre you are trying to positively/negatively ID on specifically.
I think a lot of people like DeepBoru too.
What is your use case?