Just posting a copy here of some news that could be relevant for the community. Originally shared by DataPulseEngineering.
Disclaimer: I am not affiliated with Corcel.
Hello Reddit Lemmy!
We’re excited to announce that Mobius, our groundbreaking debiased diffusion model, will be released later this week! Mobius pushes the boundaries of domain-agnostic debiasing and representation realignment, achieving unrivaled generalization across a vast array of styles and domains without the need for costly pretraining.
Key features of Mobius:
- Virtually bias-free image generation across all domains
- Exceptional generalization capabilities for top-quality results
- Efficient fine-tuning for specialized models with minimal resources
Alongside the launch, we’ll release a research paper, “Constructive Deconstruction: Domain-Agnostic Debiasing of Diffusion Models,” detailing our innovative approach.
Stay tuned for the official release date and licensing details. We can’t wait to see the amazing creations you’ll generate with Mobius!
Got questions?
Thanks for your support, The Corcel Team
Examples
For a nuanced approach, here follows some critisism. Link to original post: “Mobius” is just an ad for Corcel
See post by Confident_Appeal_603
Update: the discord server members / friends of Mobius are brigading the comments.
See the model card: https://huggingface.co/Corcelio/mobius
It’s a non-commercial model they want people to pay to use through their API, and won’t allow anyone else to publish the weights, even though they tout the ability to finetune it in the hype post.
Looking deeper into things and it’s using Bittensor to “decentralise AI production”, and it’s using blockchain. Another crypto scam.
It’s quite odd. as a researcher, the claims to cut down on training cost by 2/3rds really stuck out to me, as I would also like to benefit from this advancement. but when you look at how they supposedly achieved this, it’s just another SDXL finetune with 25 million images.
A fun gem from the model card:
- highly suggested to preappenmed watermark to all negatives and keep negatives simple such as “watermark” or “worst, watermark”
A model without any bias shouldn’t really need “watermark” in the negative prompt.
Here’s the license text from the model card:
Mobius is released under a custom license that governs its usage and distribution rights:
Non-commercial use: The model is fully open and available for any non-commercial use. Researchers, students, and enthusiasts are encouraged to explore, modify, and build upon the model freely, as long as they do not use it for commercial purposes.