v2
Add a couple images to the dataset and modified a few captions then fine-tuned LoRA by training for 10k steps total. Try training with this base model - it's fast and does a great job. You can also give the image a pass through Z Image Turbo if you prefer its looks better given the difficulty many have reported creating Z Image Base loras. (btw I think my Z Image base LoRAs are just as good as my Qwen ones, sometimes come out better even)
v1
I'm surprised at how well this model did. Without any LoRA it generates the least monsterous dick I've seen, and with training and the right sampler/scheduler you can get some good D out of this. And it doesn't put the head on backwards nearly as often as some of the other models, but I am always tweaking my dataset + captions, so maybe that helps as well.
Not sure if trigger word is always needed, but I changed from using bwc to p3n15 for this one since Klein 9B often thought I wanted a black and white image not a big white cock image.
Description
I added a couple images to the dataset and modified a few captions then fine-tuned LoRA by training for 10k steps total
FAQ
Comments (6)
Ernie looks worse (fried) than Z-Image and Klein from what I'm seeing. I'm wondering why anyone is bothering with Ernie and why a company would put out yet another small image generation model. What we need is a competitor with Wan 2.2 that focuses on image generation.
who knows. i like to see how the various models learn dick. this one learned faster than others the same size.
@coachbate But the results are worse.
@ss9999 so use the other one you prefer. wtf
@coachbate wtf is my reaction to Ernie.
@ss9999 It's not as bad as Flux.2 4B model but Qwen Image 2512 and Z Image Base are my best loras