nebulette/moss-image-2b-encoder

Moqs 2B

The model relies on Qwen3.5 2B, which provides text embeddings for the MOSS Image model.

The previous models were based on the 1M+ corpus, but the dataset here is much smaller for now.

The data samples covered half of the English vocabulary from the tokenizer, including most lowercase words.

Based on the work done in previous years, the LoRAs are expected to work out-of-box.

Here's an obligatory preview image with Moqs&Spectrum

Reason for being

Now that the fast inference in diffusers is complete, the dataset must

• include both comma seperated labels and natural language texts
• include labels, concepts that are not available on danbooru and other censored image boards
• include frequent pixiv, xitter usernames
• support lightweight markdown formats

There are no plans to include text and infographic generation, nor to support the layout format.

Expanding the knowledge of the model is not an objective due to the frozen text encoder, and other modules, which restrict its ability to learn more.

Source data

• common colors, verbs
• danbooru_tag_taxonomy
• danbooru2026_moe
• danbooru_wiki_2606
• gelbooru 40k (each sample has a new, unexplored booru tag)
• openart_painterly_foundations
• saber (2610.22602)
• sdxl_artists
• spright_coco
• pexels_portrait
• z_image_ethnicity

Common phrases were removed from the text entries, then they were refined further.

The model was designed to cover the most common character features, visual and spatial tasks.

Scaling

Multiply the output tensor by a single scalar value. These are three images at different scales