Qwen Image Edit appears trained at 1MP: resizing inputs to 1024x1024 yields far better results
Civil_Fee_7862 · reddit · 2026-08-26
The author noticed Qwen Image Edit performs substantially better when images are resized to 1024x1024 during encoding, then upscaled back after generation. The speculation is that the model was trained on 1MP inputs, but no docs confirm this. The author also notes 1MP seems optimal for other models as well.
More from Multimodal
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15
- DiffSynth-Studio Adds OPD Workflow: Few-Step Diffusion Training in ~2 GPU-Hours — bdsqlsz · 2026-10-15
- AI film and TV hit the market: Prime Video picks up AI spin-off, pilots screen at Mipcom — lmoroney · 2026-10-15
- Ramayana film taps DNEG's massive VFX and AI-assisted production for global release — CurieuxExplorer · 2026-10-15