Pruna AI open-sources object-removal LoRAs for Qwen-Image 2.1, 6 steps vs 40
pbaylies · x · 2026-11-13
Pruna AI open-sourced P-Q-Image-2.1-Remove-Object, a set of LoRA adapters for Alibaba's Qwen-Image 2.1 that do one job: remove the object you point at — along with its shadow and reflection — while leaving everything else untouched and aligned with the input.
Highlights:
- Efficiency: few-step variants built on their distillation LoRA run in 6 steps (8-step also available) without CFG, vs 40 steps for the base model
- Quality: removes shadows/reflections even outside the object's outline; the rest of the image stays put — fixing Qwen-Image 2.1's tendency to alter or shift the whole frame
- Usage: loads directly on the base pipeline; point at objects with a rough red outline or name them in the instruction
Weights and a ComfyUI version are both available.
Related event: Pruna AI Open-Sources Object-Removal LoRA for Qwen-Image 2.1(3 posts)→
More from Multimodal
- The final render everyone wants: polished Zelda-style 3D character scene — banteg · 2026-11-13
- One Prompt + Claude Code Produces a 5-Minute 3D Explainer Video — Western-Airline9355 · 2026-11-13
- RTX 5070 Ti Benchmarks: Tuned ComfyUI Flags Cut H3 Video Time by Over Half — AnalElder · 2026-11-13
- Team reveals $102,725, 45-day AI film budget and shares full workflow — NightsRadiant · 2026-11-13
- Midjourney SREF share: curling black lines and dusty lilac, code included — tisch_eins · 2026-11-13
- Anthropic has the most creative image gen despite having no image gen model — Angaisb_ · 2026-11-13