Pixeljourney XL Model:
Pixeljourney, isn't an all-encompassing model that excels at everything.
It struggles somewhat with accurately following prompts and tends to be somewhat inflexible.
However, it uniquely directs outputs into a stylistic direction inspired by various AI artists and the Midjourney style.
Capabilities of Pixeljourney:
Excels at: Semi-realistic motifs, portraits, aliens, creatures, cyborgs, artistic themes, weird fashion, fusions between humans & different creatures, creative motifs.
Adequate at: Photorealism.
Struggles with: Anime, low-resolution monochrome sketches.
Technical Aspects:
Sampler for Examples: I used dpmpp_3m_sde/karras and a (cfg) range of 5-7.
How to Use Pixeljourney:
Start with SDXL resolution Sampler at 30+ steps.
Apply HiRes-Fix or NNLatentUpscale.
Proceed with UpscaleSampler for 20 steps and Denoise around 0.04.
Optionally, use SDXL Refiner for around 8 steps with Denoise at 0.2.
Optionally, apply a Face Detailer for enhanced facial details (not needed for Portrait shots).
Dataset Composition:
The dataset includes 1500 images from the LinkedIn AI/CC Community, mainly created using Midjourney.
Additionally, 500 high-resolution photos were added to enhance details like eyes and skin.
Training Details:
The model was trained with 38,000 steps and a batch size of 4 using a 48GB RTX 6000 Ada for approximately 25 hours.
NSFW:
Pixeljourney is an uncensored model, capable of creating NSFW content.
However, it has not been specifically trained for such content and don't preform verry well at it.
Finetuned with ashleykza's kohya_ss template @ runpod
Description
Beta_XL
Trained on 2k Images, 38k Steps
FAQ
Comments (8)
It rocks! New rev?
I'm currently working on a dataset for the second training. But first, I will create a merge and then train this merge. The dataset is about 70% complete.
I've changed the plan, working on the dataset, trying to balance it better and using a new capturing system based on the chatgpt_api.
the trend i've been noticing with people using AI models as their base, is that while its little more creative, up close they look terrible. I always zoom in because I enlarge these, and so far everything i've tossed at this look quite bad up close. Its fine if you look at it from a distance, but it looks like AI up close. And I want it to look like a photo or a painting or anything other than AI. Some images are almost there, but others not really at all.
Yes, the graphics of reality are indeed very good.
I think it will require some more research and quite a few improvements to the AI models before they have better graphics than reality.
It's too wild by itself. The key is to use Mohawk v18 as the hires fix model. That tames it. I did all sorts of stuff with it that way. Now I look at them and maybe they could be cleaned up a little, if you zoom in. I'm trying to upscale some with Flux right now without killing the creativity. Sure makes me wish I had a Pixeljourney Flux finetune. ;> Nontrivial, I know.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.