---
license: apache-2.0
pipeline_tag: image-to-text
tags:
- genlip
- tipsv2
- vision
---

ViT-B model

<span style="color:red; border: 1px solid black;">Please do not download the model. The repo was kept for archival purposes.</span>

The vision encoder can be pretrained with autoregressive language modeling objective - no contrastive loss, no dual-tower architecture, and no extra text decoder. 

The output includes both visual and textual data. The model encodes [text token](https://huggingface.co/nebulette/booru-character-aware-tokenizer) positions using ALiBi bias in attention.

![](images/attn_mask.png)

Causal attention was applied to text tokens, while fewer vision tokens were visited in the middle blocks. This resulted in a faster training cycle.

![](images/efficient_vit.png)

SV(O) booru tags are applied with increased weighting in the loss calculation (2605.00809, figure 4) for both solo and multi-character images.

Source data
- danbooru 2025-26
- gelbooru for a certain locked tag
- Kimi K2 style visual descriptions from multiple thinking models
- cell1/tagutl was used for short tags until v0.5; a newer model was used later on

References
- 2108.12409
- 2501.04765
- 2604.12012
- 2605.00809