Loading showcase data…
Interactive Demo
AI surgical outcome previews, straight from the real pipeline
A face photo and one line of surgeon shorthand go in, a realistic surgical preview comes out. Pick a patient, a procedure, and an instruction below. Every prompt expansion, image, and metric on this page came out of the real pipeline. The rhinoplasty LoRA was trained on 134 real before and after pairs, and across the 18 cases here it lifts mean identity similarity from 0.65 to 0.90 ArcFace over the base model, at about 10 seconds per edit.
All outputs precomputed on a single RTX 5090 · editor Qwen-Image-Edit-2511, NF4 + Lightning
· expander Qwen3.5-9B in 4 bit · faces are StyleGAN synthetic, no real patients
07
The pipeline
photo + shorthand→ Qwen3.5-9B expander (structure and guardrails)→ Qwen-Image-Edit-2511 (NF4, Lightning 8 step)→ procedure LoRA (QLoRA on real pairs)→ eval canary (DINO · ArcFace · LPIPS · CLIP)
The whole system runs on one 32 GB consumer GPU, training included. The first build died from silent identity collapse, where the model learned to hand back its input untouched. The rebuild scores every output with an edit magnitude canary and you can see it reporting under each result above. During LoRA training the same canary quarantined three checkpoints that drifted toward under editing, so a bad checkpoint can never be promoted quietly.
08
Credits and citations
Created by Munish Persaud, Logan Flickinger, Md Sahif Hossain, and Micah Patrick.
The procedure LoRA was trained and evaluated on the HDA Plastic Surgery Face Database, used under its research only license. None of its images appear on this page.
- C. Rathgeb, D. Dogan, F. Stockhardt, M. De Marsico, C. Busch, "Plastic Surgery: An Obstacle for Deep Face Recognition?", 15th IEEE Computer Society Workshop on Biometrics (CVPRW), pp. 3510–3517, 2020.
- Qwen Team, "Qwen-Image Technical Report", arXiv:2508.02324, 2025. Editor weights: Qwen/Qwen-Image-Edit-2511 (Apache 2.0).
- Qwen Team, Qwen3.5, 2026. Expander weights: Qwen/Qwen3.5-9B (Apache 2.0).
- LightX2V, Qwen-Image-Edit-2511-Lightning step distillation LoRA: lightx2v/Qwen-Image-Edit-2511-Lightning.
- T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, T. Aila, "Analyzing and Improving the Image Quality of StyleGAN", CVPR 2020. Source of the synthetic faces shown here.
- Metrics: DINOv2 (M. Oquab et al., 2023) for edit magnitude, ArcFace (J. Deng et al., CVPR 2019) for identity, LPIPS (R. Zhang et al., CVPR 2018), and CLIP (A. Radford et al., ICML 2021) for instruction alignment.