SimT2I Demo

This Space is an interactive research demo for Simple Text-to-Image (SimT2I). It currently serves the released PyTorch/Diffusers checkpoints:

• SimT2I-B/16
• SimT2I-L/16

The demo is intended for quick qualitative inspection of the released models, rather than as a benchmark or a replacement for the full research codebase.

Usage

Enter a prompt, choose either SimT2I-B/16 or SimT2I-L/16, and adjust the CFG scale if you want stronger or weaker prompt guidance. The advanced controls let you sample more candidates per prompt when you want a broader qualitative look.

The negative-prompt field is an experimental steering control. When it is set, the CFG reference branch uses prompt + negative prompt instead of an empty text condition.

Checkpoints

The app loads model weights from the private SimT2I/SimT2I model repository at runtime. The Space must therefore be configured with an HF_TOKEN secret that can read the SimT2I organization repositories.

Notes

For each prompt, the demo can generate several candidates and return the one with the highest score under a CLIP+MLP aesthetic predictor. The selector follows the public christophschuhmann/improved-aesthetic-predictor implementation, using the sac+logos+ava1-l14-linearMSE.pth predictor on top of CLIP ViT-L/14 features.

Output quality can vary with prompt phrasing, CFG scale, random seed, and the number of generated candidates. The selected image is the highest-scoring sample under the aesthetic selector, not a guarantee of factual or compositional correctness.