A 1-Bit Image Model Just Launched And It’s Great!
Earlier this week, I did a video lamenting why I really dont mess with local AI image generation - it still remains something I do exclusively on the cloud for the most part. It turns out, that version of the model was *broken* and it has been fixed.
So in this video, I go deep into the new PrismML Bonsai Image 4B (base Flux.2 Klein 4B) and rerun my tests and do some real benchmarking - and it changed my entire view on local image gen. Turns out Ternary Bonsai Image is 78% less memory when under load - which is nothing to scoff at considering the images are still quite good!
*Links* :
AnythingLLM: https://anythingllm.com/
Flux.2 Klein: https://huggingface.co/black-forest-l...
PrismML Blog: https://prismml.com/news/bonsai-image-4b
Bonsai Image White paper: https://github.com/PrismML-Eng/Bonsai...
PrismML Ternary MLX: https://huggingface.co/prism-ml/bonsa...
*Chapters* :
0:00 My Previous View On Image Gen Locally Is Incorrect
3:11 Who is PrismML?
4:53 Bonsai Image 4B is a better Flux.2 Klein 4B
8:07 What are are going to benchmark and compare today
9:00 Running Bonsai Image 4B locally on your PC
12:54 What *Actually* got improved in Bonsai Image?
16:27 Comparing Bonsai Image Outputs to other engines
25:25 Performance & Memory Benchmarks are impressive
29:21 Is this the future for image gen models?
30:52 But I have some questions for the PrismML team...