BeyondImaginationXL_V0_005_Fix
Fixed the issue with model metadata, preventing it from loading in some UIs. Thahks to S1LV3RC01N for help!
BeyondImaginationXL_V0_005
Update: I changed the terms of license. It is now OK to use the model on other generation services. Please, don't forget about crediting me.
Here is the new version of the model. Major update is still coming, but I decided to publish this checkpoint, which has two improvements:
Higher details in images,
Fixed the issue with metadata(actually, no, will have to do some further research)
Recommended settings are the same as before.
BeyondImaginationXL_V0_003
New version is coming soon! Stay tuned!
BeyondImaginationXL is my first attempt to train an SDXL checkpoint. Sure, it's a rookie, but hey, even rockstars had to start somewhere! This checkpoint is like the rebellious offspring of a commercial model I've been cooking up for a few months. It's got attitude, focusing on keeping it real, rocking various visual vibes, finding that sweet spot between work-appropriate and, well, more adventurous content.
Recommended settings are:
Sampler: DPM 2m or 3m
Scheduler: karras
Steps: 30 - 60
CFG: 3-7
Refiner: Not needed
VAE: Baked in
This is a starting point, so feel free to try it, comment, make suggestions, etc. I will be developing it further. The more you try it and comment, the better new versions get!
Further plans:
Looking at the results of my tests here is what I am planning to do:
improve people (both SFW and NSFW), hands and feet
add interior and landscape design styles. That's a part of a commercial project I'm a founder of and future public releases of the model will get some (but not all😉) cool elements from the commercial model
enhance the level of control over the composition of an image
fix the issue with metadata - I've got reports that the model is not properly recognized in some UIs (I use ComfyUI and my own pipeline based on comfy - it works like a charm)
I will not be working on SDXL Turbo models due to inferior quality and licensing policy
what else? - please write in the comments section
About me:
I am a data scientist and I am a founder of startups related to Computer Vision and Generative AI. Please feel free to contact me if you have any commercial ideas to discuss.
Description
Slightly improved detailes.
Hopefully, fixed the bug with lacking metadata (please, let me know if the model is recognised in webui 1.6, 1.7dev, and webui-forge, where there was an issue in v003)
FAQ
Comments (3)
I fixed the model and uploaded it to my Google Drive. So you can update the model page if you want to. Model Link
Oh, thank you! So the problem was in the metadata field. I used some nodes from ComfyUI in my training pipeline to save time and one of the nodes I used without looking inside the code was CheckpointSave. Tuns out it saves the model with incorrect metadata.
@Askel72Â Yeah it saves model with both incorrect metadata and no embeddings.position_ids key in the small CLIP. But comfyanonymous should have at least fixed the last one a couple of weebs back
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.