Dress-ED: Instruction-Guided Editing for Virtual Try-On and Try-Off

Davide Lobba^1,3*  Fulvio Sanguigni^2,3*  Bin Ren⁴  Marcella Cornia²  Rita Cucchiara²  Nicu Sebe¹

¹University of Trento   ²University of Modena and Reggio Emilia   ³University of Pisa   ⁴MBZUAI

^* Equal contribution

ECCV 2026

Dress-ED is the first benchmark for instruction-driven virtual try-on and try-off, with over 146k verified quadruplet samples across seven editing types — covering both appearance and structural garment modifications.

---

📌 Overview

Dress-ED (Dress Editing Dataset) is the first large-scale benchmark that unifies Virtual Try-On (VTON), Virtual Try-Off (VTOFF), and text-guided garment editing within a single framework. Each sample pairs an in-shop garment image and the corresponding person image with their edited counterparts and a natural-language instruction describing the desired modification. The dataset spans three garment categories, seven edit types, and 146,460 verified quadruplets, enabling rigorous training and evaluation of controllable, instruction-driven fashion generation models.

---

✨ Highlights

• First unified benchmark integrating VTON, VTOFF, and instruction-driven garment editing in a single framework.
• 146,460 verified quadruplet samples across three garment categories and seven editing types.
• 49,664 distinct garment identities and 6,073 unique editing instructions.
• Fully automated curation pipeline: MLLM-based attribute extraction → diffusion-based synthesis → LLM-guided quality verification.
• Quadruplet structure (I_garment, I_person, I_garment_edit, I_person_edit, T_inst) supports VTON, VTOFF, garment editing, and person editing tasks simultaneously.
• Evaluation protocol based on real ground-truth images via an inverse-editing strategy, ensuring unbiased metrics.

⚙️ Curation Pipeline

Dress-ED is built on top of Dress Code through a fully automated four-stage multimodal pipeline:

Stage 1 — Structured Attribute Extraction & Instruction Generation Qwen3-VL processes each (I_garment, I_person) pair to extract structured garment semantics (color, pattern, material, neckline, sleeve length, etc.) as a JSON representation. Natural-language editing instructions are then synthesized via rule-based templates, covering both appearance and structural edit categories.

Stage 2 — Edited Garment Generation FLUX.2 Klein applies the editing instruction to the in-shop garment image, producing I_garment_edit. This stage yields ~300k candidate edited garments across all categories.

Stage 3 — Edited Try-On Generation FitDiT renders the person wearing the edited garment, producing I_person_edit for each valid (I_person, I_garment_edit) pair.

Stage 4 — Automated Quality Verification A two-step filtering pipeline ensures semantic correctness and visual fidelity:

1. GPT-5 scores each triplet (I, I_edit, T_inst) on instruction adherence, content preservation, and realism (score ∈ [0, 100]).
2. A fine-tuned InternVL-3.5 verifier — distilled from ~5k GPT-5-annotated samples — is applied at scale. Samples scoring below 80 are discarded.

The final Dress-ED retains 146,460 high-quality verified quadruplets.

---

🧪 Benchmark Tasks

Dress-ED defines two core benchmarks, both evaluated using a real-image inverse-editing protocol (the model reconstructs the original unedited image from its edited counterpart using an inverse instruction, ensuring metrics are computed against real ground-truth images):

Instruction-Driven Virtual Try-On

• Paired setting: given (I_person, I_garment, T_inst), generate the person with the garment modified as instructed.
• Unpaired setting: given (I_person, I'_garment, T_inst) with a different garment, generate the person wearing the edited version of I'_garment.

Instruction-Driven Virtual Try-Off Given (I_person, T_inst), generate the edited in-shop garment I_garment_edit matching the described modification.

---

🧱 Dataset Structure

Each sample in Dress-ED is a verified quadruplet defined as:

S = (I_garment,  I_person,  I_garment_edit,  I_person_edit,  T_inst)

Field	Description
I_garment	Original in-shop garment image
I_person	Person wearing the original garment
I_garment_edit	Edited in-shop garment image
I_person_edit	Person wearing the edited garment (VTON output)
T_inst	Natural-language editing instruction

Edit Types

Edits are grouped into two macro-categories:

Appearance edits — modify the visual properties of the garment:

• Change Color · Change Pattern · Change Material · Fine-Grained

Structural edits — alter the geometry or composition of the garment:

• Add Detail · Remove Element · Modify Structure

---

📊 Dataset Statistics

By Edit Type

Edit Type	# Samples	Proportion
Add Detail	39,776	27%
Change Pattern	39,559	27%
Change Color	29,983	20%
Modify Structure	15,670	11%
Change Material	14,091	10%
Remove Element	5,305	4%
Fine-Grained	2,076	1%
Total	146,460	100%

By Garment Category

Category	# Samples
Dresses	80,865
Upper-body	45,567
Lower-body	20,028

Splits

Following the same train/test protocol as Dress Code, with no overlap of garment identities across partitions:

Split	# Samples
Train	132,201
Test	14,259

All images are standardized to 768 × 1024 resolution.

---

Evaluation Metrics

Dimension	Metrics
Visual Fidelity	FID · KID · SSIM · LPIPS · DISTS
Edit Correctness	DINO-I (content-level similarity to ground truth)

---

📦 Dataset Access

The dataset is available on HuggingFace at davidelobba/Dress-ED.

---

📚 Citation

If you use Dress-ED in your research, please cite:

@article{dressED2026,
  title={Dress-ED: Instruction-Guided Editing for Virtual Try-On and Try-Off},
  author={Lobba, Davide and Sanguigni, Fulvio and Ren, Bin and Cornia, Marcella and Cucchiara, Rita and Sebe, Nicu},
  journal={arXiv preprint arXiv:2603.22607},
  year={2026}
}

---

🙏 Acknowledgements

Dress-ED extends the Dress Code dataset. The curation pipeline builds on Qwen3-VL, FLUX.2 Klein, FitDiT, GPT-5, and InternVL-3.5. We thank the authors of these works for making their models publicly available.

---

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

You are free to share and adapt the material for non-commercial research purposes only, provided appropriate credit is given. Please also refer to the Dress Code license for terms governing the underlying garment and person imagery.