Have a Safetensor Version?
FP16? or a model not in gguf format? Or have a link to the original repo? Love to learn how these are made. Thanks!
FP16? or a model not in gguf format? Or have a link to the original repo? Love to learn how these are made. Thanks!
I do have full weights, but have decided against releasing them. I am not satisfied with the model just yet. I have several MLX versions available as well in separate repos.
This is a merged model. I took Qwen3 4b Instruct and ran Heretic (abliteration/uncensoring) with targeted refusals for image generation (I actually modified Heretic a bit as well lol... but that's irrelevant) - next I wrote a python script to construct unique seeds, which are full image prompt enhancements requests (as single turn, the output followed the format-
instruction:
User:
Output: , building a synthetic dataset. I generated with 18k-19k conversation examples and trained the LoRA using MLX. I'm done releasing files for V2. I plan to release the full weights, LoRA, training script, etcetera for V3 (release date TBD but it's in the works).
How it works as an enhanced clip, the idea is to give the model better semantic understanding of visual spatial descriptors, lighting, textures, camera angles, etc...
I may release a V2.5 which will be made with a new more refined dataset before releasing V3 - depends on how much extra time I have.