DyRef lifts Qwen-Image-Edit-2511 from 4.97 to 8.38 on multi-reference editing

量子位 · wechat · 2026-07-29

• DyRef targets complex multi-reference image generation, where a model must combine identity, background, pose, lighting, and style from multiple reference images without letting them conflict.
• The team built OmniRef-Bench, a 395-sample benchmark covering 5 reference types, 2 to 7 reference images per sample, and up to 4 simultaneous reference categories.
• DyRef uses a two-stage training recipe: supervised fine-tuning first, then reinforcement learning with dynamic reward optimization so harder samples get more weight.
• On OmniRef-Bench, the post says Qwen-Image-Edit-2511 improved from 4.97 to 8.38 in MLLM evaluation average, narrowing the gap to NanoBananaPro and beating Seedream4.5.
• The article also claims the method improves performance on MultiBanana, OmniContext, and even single-image editing benchmarks such as ImgEdit and DreamBench++.

Original post →

More from Multimodal

• Surreal AI-generated landscape hides "another landscape inside the measurements" — piotrbinkowski · 2026-10-16
• Fan animation made in Blender with Claude Code (Opus 5.5) — Icy-Appointment3583 · 2026-10-16
• Making Blender animations with Opus, Astra and Argon: how you prompt still matters — chongdashu · 2026-10-16
• Tessellation-free geometric queries on displaced surfaces cut caustics variance 10x — ssh4net · 2026-10-16
• Complex multi-link CAD mechanism finally animates, dev celebrates the breakthrough — zetalyrae · 2026-10-16
• OpenMOSS releases MOSS-Image-SFT: Apache 2.0 text-to-image model on Hugging Face — bdsqlsz · 2026-10-16

Multimodal channel →