|
--- |
|
license: openrail |
|
--- |
|
|
|
# MJv4 Hallucinations |
|
|
|
These are 3 models trained on a small (<2000) dataset of Midjourney v4 images with no particular style. <b> These models are nowhere near as good as Midjourney v4 </b>, and they all suffer from a lot of "language drift" but they do have an interesting style. They are the best of something like 60 different models I trained as part of a set of experiments aimed at replicating Midjourney v4's style with only a few, uncaptioned images. |
|
|
|
The models are: |
|
|
|
- <b>mjg-4000-model.ckpt</b>: trained on 250 MJv4 images with no regularization for 4000 steps, prompt: "mjg style" |
|
- <b>mjg-12000-model.ckpt</b>: trained on 250 MJv4 images with no regularization for 12000 steps, prompt: "mjg style" |
|
- <b>mjv-1200-model.ckpt</b>: trained on 7 MJv4 images with 1000 regularization images for 1200 steps, prompt: "mjv style" |
|
|
|
Models you can download are <b>bolded</b> |
|
|
|
<img src="https://github.com/Lewington-pitsos/mj4-hallucinations/blob/main/compare.png?raw=true" width="100%"/> |
|
|
|
In my subjective opinion, only <b>mjv-1200-model.ckpt<\b> is actually worth downloading. |
|
|
|
## Credits: |
|
|
|
- [NitroSock](https://github.com/nitrosocke/dreambooth-training-guide) for the regularization images |
|
- [prompthero](https://huggingface.co/prompthero/openjourney) whose idea I copied |
|
|
|
## Take Down |
|
|
|
As far as I can tell, uploading these models does not cause any person or corporate entity any harm, but if you think I am wrong about this please reach out. |
