Qwen-Image-Edit preserves ~90% of reference object identity in inpainting, but takes 4 min per inference
Super_Mission_3130 · reddit · 2026-09-07
A developer details experiments on reference-based object replacement (source image + mask + reference image → same scene with the object swapped), where identity preservation of the reference object is the hard requirement.
- SDXL Inpainting with ControlNet and IP-Adapter only produced visually similar objects, losing structural details.
- Qwen-Image-Edit-2511 with ReCoEdit-RL preserved roughly 90% of the reference object's identity, including details SDXL struggled with.
- The trade-off: high VRAM usage and 4 minutes per inference on their setup, too slow for repeated production use.
The author is asking for open-source pipelines that balance reference fidelity, VRAM and speed — ideally adapting the reference object's geometry/viewpoint first, then transferring appearance with correct scale, perspective, lighting and shadows.
More from Multimodal
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- Full animation short made with Google Gemini tools, 10,564 frames — AI_Andrew · 2026-10-26
- Same prompt, two styles: Claude Opus 5.5 vs GPT-6 Astra on a C-section 3D animation — Sourcecode12 · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26
- Community checkpoint 'Narrative Drift' brings art-directed visuals to Krea 2 Turbo — MoonbearAIArt · 2026-10-26
- Same prompt, split results: Claude Opus 5.5 vs GPT-6 Astra on a 3D C-section explainer — Sourcecode12 · 2026-10-26