Open-source LoRA turns one RGBA image into any viewing angle with Qwen-Image-2.1
Gradio · x · 2026-09-26
- ML-Intern-lab released a viewpoint-orbit LoRA for Qwen-Image-2.1: give it one transparent RGBA object image plus a relative camera instruction, and it returns the same object from the requested viewpoint, transparency intact
- Supports 7 relative rotations (45-180 degrees) x 3 elevations, 23 instructions total; rotation is relative to the uploaded view
- Trained 2,000 steps at 768px (rank 32/alpha 32, 2.2s/step on one A100-80GB) with ostris/ai-toolkit on synthetic orbit renders of Google Scanned Objects; dataset (gso-orbit-rgba) also open-sourced
- Ships verified end-to-end diffusers loading code with prompt template, plus a hosted Hugging Face Space to try it
More from Multimodal
- Sentinel splits monolithic Three.js into scene-based workflows with HD PNG sequence export — PurzBeats · 2026-09-26
- One-shot product demo video made with Opus 5.5 and browser_use video-use — Scobleizer · 2026-09-26
- Single-prompt motion graphics looks pro: 'it's officially solved', say testers — claud_fuen · 2026-09-26
- Concatenating an empty conditioning fixes Qwen Image Turbo LoRA artifacts at 14 steps — GTManiK · 2026-09-26
- Google makes AI video generation free for all accounts via Vids, no subscription needed — alexcovo_eth · 2026-09-26
- Claude Opus 5.5 generates a full launch video — animation, music, voiceover — in 20 minutes — cedric_chee · 2026-09-26