VeraRetouch: A Multi‑Task Inference Framework for Photo Editing (SIGGRAPH 2026)

Background

Mobile photography has become ubiquitous, but professional‑grade post‑processing remains a barrier for ordinary users. Existing one‑click filters either over‑stylize or ignore fine details, while recent multimodal large models still rely on external tools such as Lightroom or Photoshop, preventing end‑to‑end pixel‑level optimization and hindering mobile deployment due to their size.

Key Innovations

VeraRetouch replaces external editing software with a fully differentiable Retouch Renderer . The framework decomposes the retouching space into three relatively independent control dimensions:

Lighting : exposure, shadows, highlights.

Global Color : white‑balance, overall hue, tonal shifts.

Specific Color : fine adjustments for red, orange, blue channels, etc.

This decomposition mirrors professional workflows, making the model’s output more interpretable and stable.

Dataset Construction (AetherRetouch‑1M+)

To train a multi‑task retouching model, the authors built a million‑scale dataset covering Auto‑Retouch, Style‑Retouch, and Param‑Retouch.

For automatic retouching, a “reverse degradation” pipeline starts from high‑quality photos, treats them as the “finished” result, and generates realistic “un‑retouched” inputs by applying expert‑derived color‑lighting changes in reverse.

For style retouching, 5,030 online style presets spanning 11 categories and 193 sub‑categories are matched to images using a visual‑language model, producing diverse textual style instructions.

For parameter retouching, random combinations of lighting, global color, and specific color parameters are sampled to create precisely controllable training pairs.

Structured reasoning traces (image analysis, problem diagnosis, and editing plan) are also embedded in the data.

Technical Details

VeraRetouch is built on FastVLM‑0.5B . An image encoder converts the input image into visual tokens, while a text encoder processes user instructions into prompt tokens. The multimodal language model then generates structured reasoning and control latents.

Retouch tokens corresponding to the three control dimensions are produced, passed through a MLP Retouch Adaptor to align with the Retouch Renderer, which finally outputs the pixel‑level edited image.

Two major advantages arise:

No dependence on external editing software; the entire pipeline runs inside the model, enabling end‑to‑end pixel‑level training.

The model is lightweight (≈0.63 B parameters), far smaller than baselines such as Flux.1 Kontext, Qwen‑Image‑2509, MonetGPT, or JarvisArt, making mobile deployment feasible.

A post‑training strategy called DAPO‑AE adds format, image‑similarity, and aesthetic rewards to improve visual naturalness while preserving instruction fidelity.

Experimental Results

On the FiveK‑Bench automatic retouching benchmark, VeraRetouch‑DAPO‑AE achieves 26.85 dB PSNR, a 1.08 dB gain over Flux.1 Kontext, and excels in SSIM, LPIPS, and histogram consistency.

On the Aether‑Bench style‑retouch benchmark, it attains the best or near‑best scores across PSNR, SSIM, LPIPS, DISTS, GMSD, and Texture Distortion, demonstrating strong style adherence and texture preservation.

For parameter retouching, the model reaches 30.18 dB PSNR, surpassing fine‑tuned diffusion baselines, though it may struggle with out‑of‑distribution parameter samples due to the Gaussian sampling used during data creation.

Inference speed: on an H20 GPU, a 512p image is processed in 6.90 s (vs. 16.78 s for Flux.1 Kontext and 14.31 s for JarvisArt). On a MacBook Air M4, the unquantized model runs in ~7.46 s, and on an iPhone 16 Pro in ~13.56 s, confirming mobile‑side feasibility.

User study with 38 participants shows VeraRetouch receives the highest scores for visual appeal, instruction consistency, and texture preservation; DAPO‑AE further improves human preference to 61.62 % in pairwise comparisons.

Limitations and Future Work

The current model still has room for improvement in localized editing. Incorporating pixel‑level mask mechanisms could enable region‑specific adjustments such as facial brightening, sky recoloring, or background enhancement.

Overall, VeraRetouch moves inference‑based photo editing from a loosely coupled “model → parameters → external tool” pipeline to a unified, lightweight, differentiable framework suitable for on‑device deployment.

Figures