4 ms·

I am amazed, though not entirely surprised, that these models keep getting smaller while the quality and effectiveness increases. z image turbo is wild, I'm loo

I am amazed, though not entirely surprised, that these models keep getting smaller while the quality and effectiveness increases. z image turbo is wild, I'm looking forward to trying this one out. An older thread on this has a lot of comments: https://news.ycombinator.com/item?id=46046916 https://news.ycombinator.com/item?id=46046916

roenxi 9mo ago
There are probably some more subtle tipping points that small models hit too. One of the challenges of a 100GB model is that there is non-trivial difficulty in downloading and running the thing that a 4GB model doesn't face. At 4GB I think it might be reasonable to assume that most devs can just try it and see what it does.
Auracle 9mo ago
Quality is increasing, but these small models have very little knowledge compared to their big brothers (Qwen Image/Full size Flux 2). As in characters, artists, specific items, etc.
- efskap 9mo ago
  I smell the bias-variance tradeoff. By underfitting more, they get closer to the degenerate case of a model that only knows one perfect photo.
- vunderba 9mo ago
  Agreed - given what Tongyi-MAI Lab was able to accomplish with a 6b model - I would love to see what they could do with something larger. Somewhere in the range of 15-20b, between these smaller models (ZiT, Klein) and the significantly larger models (Flux.2 dev).
- littlestymaar 9mo ago
  That's what LoRAs are for. And small models are also much easier to fine tune than large ones.
  - Auracle 9mo ago
    I hate that excuse. I want the model to know who the Paw Patrol is without either finding a lora (which probably won't exist because they're mostly porn) or needing to make a dataset, tag it, and then train it myself.
aitchnyu 9mo ago
Is there a theoritical minimum for params for a given output? I saw news about GPT 3.5, then Deepseek training models at a fraction of that cost, then laptops running a model that beats 3.5. When does it stop?

roenxi 9mo ago

There are probably some more subtle tipping points that small models hit too. One of the challenges of a 100GB model is that there is non-trivial difficulty in downloading and running the thing that a 4GB model doesn't face. At 4GB I think it might be reasonable to assume that most devs can just try it and see what it does.

Auracle 9mo ago

Quality is increasing, but these small models have very little knowledge compared to their big brothers (Qwen Image/Full size Flux 2). As in characters, artists, specific items, etc.

efskap 9mo ago
I smell the bias-variance tradeoff. By underfitting more, they get closer to the degenerate case of a model that only knows one perfect photo.
vunderba 9mo ago
Agreed - given what Tongyi-MAI Lab was able to accomplish with a 6b model - I would love to see what they could do with something larger. Somewhere in the range of 15-20b, between these smaller models (ZiT, Klein) and the significantly larger models (Flux.2 dev).
littlestymaar 9mo ago
That's what LoRAs are for. And small models are also much easier to fine tune than large ones.
- Auracle 9mo ago
  I hate that excuse. I want the model to know who the Paw Patrol is without either finding a lora (which probably won't exist because they're mostly porn) or needing to make a dataset, tag it, and then train it myself.

aitchnyu 9mo ago

Is there a theoritical minimum for params for a given output? I saw news about GPT 3.5, then Deepseek training models at a fraction of that cost, then laptops running a model that beats 3.5. When does it stop?