MULTIMODAL AI - AI Dynamics

Stable Diffusion 2: Four New Models for Image Generation

By

–

25 November 2022 18h50

What is Stable Diffusion 2? new models Diffusers supports 4 of them: 768×768 text to image (shown in first tweet) 512×512 base model. Use it for faster results or to use less memory. In-painting specialist model. x4 upscaler (Depth-Estimation coming soon!)

→ View original post on X — @huggingface

25 November 2022

Stable Diffusion 2 Depth Estimation Coming to Diffusers

By

@huggingface

–

25 November 2022 18h50

The last model performs depth estimation to create high quality image masks. Coming soon to diffusers! For in-detail information and examples check the Stable Diffusion 2 doc page: https://
huggingface.co/docs/diffusers
/api/pipelines/stable_diffusion_2
…

→ View original post on X — @huggingface

25 November 2022

Comparing Stable Diffusion 1.5 vs 2.0 Image Generation Quality

By

@karpathy

–

25 November 2022 2h34

easy to compare a lot of images from both models on http://
stableboost.ai , e.g. "cute dog cooking tacos, photorrealistic", grid of boosted images from 1.5 (left) and 2.0 (right). 2.0 looking more distorted, cartoony, simpler, ignores text more. may need more prompt engineering

→ View original post on X — @karpathy

25 November 2022

Stable Diffusion 2.0 Shows Quality Decline Compared to 1.5

By

@karpathy

–

25 November 2022 2h34

plot twist: stable diffusion 2.0 looks quite a bit worse on the few prompts i've tried so far compared to 1.5 (even not including celebrities/artists). Running theory seems to be this is due to an aggressive data sanitization campaign since the original release (?).

→ View original post on X — @karpathy

25 November 2022

Complex Autoencoders Enable Advanced Object Auto-Segmentation in Images

By

@wellingmax

–

24 November 2022 19h48

Fantastic work by @sindy_loewe on complex autoencoders that are able to auto-segment objects in images. https://t.co/RuohaS0WYO
— Max Welling (@wellingmax) 24 novembre 2022