A Hugging Face Blog post collects a set of community models, datasets and demo apps from the ML-Intern-lab account. Many build on Qwen-Image 2.1. Two small prompt-rewriter models (0.8B and 2B parameters) are meant to stand in for the 9B Qwen-Image-2.1-PE-T2I model. A “Rewriter Arena” demo renders the 9B, 0.8B and 2B rewriters side by side.
The post also lists several other releases. Two LoRA adapters cover doodle-in editing (sketch an outline on a photo, name it, and the object appears) and rotating the camera around a transparent object. There is a 2B citrus-disease vision-language model with a “Citrus Doctor” demo for diagnosing diseases, pests and deficiencies, plus a FLUX.2 klein LoRA for drawing Huggy mascots in the Hugging Face brand style. A 4-step Agate Preview 002 text-to-image demo runs on GPU and in the browser via WebGPU, updating as you type. Supporting datasets are listed alongside, and the source text does not include the article’s full prose.
Why it matters
- Small models, such as the 0.8B and 2B prompt rewriters, point to lighter-weight ways of improving image-generation prompts without running a 9B model.
- The live demos let creators try sketch-based editing, camera-orbit views and brand-style mascot drawing directly in the browser, with no local setup.
- The citrus-disease model and its dataset show how narrow, domain-specific vision-language models can be built and shared openly for specialised uses.
Source
- The model that didn’t exist, so you made it yourself — Hugging Face Blog, 08 Oct 2026
Summary written by FoxaMind with AI assistance from the source above. Check the original for full details.