The Black Box Hidden in Your Photo Editing App — vivo Dissects It into a Scalpel with 0.5B Parameters

Open your phone's photo editing app, tap "Auto Enhance," and get the result in 3 seconds. Are you satisfied? About 70% of the time, it’s okay.

What about the remaining 30%? Skin tones are too yellow, shadows are muddy, highlights are blown out — you don’t know what it adjusted, and you don’t know how to fix it. One-click filters feel like a one-shot deal: either accept everything or start over.

Where’s the problem? Photo editing tools are black boxes. They don’t tell you, “This image is underexposed by 0.3 stops and the color temperature is 200K too warm.” They just change the pixels directly. You see only the result, not the process.

VeraRetouch is here to dismantle that black box. A 0.5B-parameter vision-language model acts as the “retouching brain,” and a fully differentiable RetouchRenderer serves as the “retouching scalpel” — every adjustment has a name, a direction, and a numerical value. The paper has been accepted by SIGGRAPH 2026, and deployment on iOS and macOS has been released.

Architecture: Brain + Scalpel, Not Brain + Black Box

Existing reasoning-based retouching methods follow this pipeline: multimodal large model analyzes the photo → generates retouching instructions → calls Lightroom/Photoshop to execute. It sounds reasonable, but there’s a fatal problem — external software is not differentiable.

What does non-differentiability mean? The model says, “Warm the color temperature by 200K,” Lightroom executes it, but whether the result looks good or not — the gradient cannot be backpropagated. The model cannot learn from pixel-level outcomes whether “this adjustment really looks good.” It can only be trained on manually annotated instruction-parameter pairs. There is a disconnect between training and inference.

VeraRetouch’s core innovation: embedding the retouching executor into a differentiable framework.

Input Image + Instruction (optional)
        │
        ▼
0.5B VLM (Retouching Brain)
   Analyze image defects → Generate retouching reasoning chain → Output control latent
        │
        ▼
RetouchRenderer (Retouching Scalpel)
  Lighting latent  → Exposure/Shadows/Highlights
  Global Color latent → Color Temperature/Tint/Overall Color
  Specific Color latent → Red/Orange/Blue and other specific channels
        │
        ▼
Retouching Result (pixel-level, gradient-backpropagatable)

The three control dimensions correspond to the workflow of a professional retoucher: first adjust light (Lighting), then set the tone (Global Color), and finally fine-tune (Specific Color). Each step has explicit physical meaning, not magic inside a black box.

Three-Dimensional Vector.

Three Modes: From "Retouch for Me" to "Retouch as I Say"

Auto mode solves "I don’t know what to adjust"; Style mode solves "I know the feeling I want but don’t know how to achieve it"; Param mode solves "I know the exact parameters but don’t want to manually apply them."

All three modes share the same VLM and Renderer — not three models glued together, but one framework with three input types. This means the model’s parameter count does not grow linearly with the number of tasks.

Million-Level Dataset: Inverse Degradation, Not Forward Degradation

The biggest bottleneck for training retouching models is data. You need “before-retouching” and “after-retouching” paired data, but professional retouchers won’t record their process for you.

VeraRetouch’s solution: inverse degradation. Instead of going from a good image to a bad one (forward degradation), it starts from a good image and reverse-engineers “what was done to this good image.”

Specific pipeline: Collect millions of professional retouching works → Use an inverse degradation workflow to infer the original image and retouching parameters → Construct the AetherRetouch-1M+ dataset. Style-Retouch uses 5,000+ style presets, and Param-Retouch directly maps parameters to pixels.

This dataset is currently not publicly released (Issue #1 remains open).

Competitive Comparison: Divergent Paths

The core divide: Does retouching need generation?

PerTouch uses a diffusion model for retouching — essentially “repainting” a stylized image. The advantage is strong stylistic expressiveness; the disadvantage is the potential to alter the original content (a common problem with diffusion models: creative but not faithful).

VeraRetouch follows the “adjustment” path — it does not generate new pixels, only adjusts the tonal and color properties of existing pixels. The advantage is strict preservation of original structure and details; the disadvantage is limited stylistic expressiveness within the color adjustment space.

For mobile phone photography retouching, the “adjustment” path is more reasonable — users want “my photo looks better,” not “AI drew me a new picture.”

Deployment: 3D LUT Acceleration, Live Photo Support

On June 16, 2026, VeraRetouch released deployment versions for macOS and iOS, also supporting Live Photo retouching — select a reference frame, and the retouching effect automatically applies to the entire temporal sequence.

Key acceleration technique: 3D LUT (Look-Up Table). Precompute the color mapping of RetouchRenderer into a 3D lookup table. During inference, directly query the table instead of forward computation, significantly improving inference speed for high-resolution images with minimal quality loss.

This is a critical step from “paper” to “product.” 0.5B parameters + 3D LUT acceleration make on-device deployment feasible.

Reality Check

The source article is a technical blog post by the project team. vivo BlueImage Lab is the core development team for VeraRetouch, with Zhejiang University and the Chinese Academy of Sciences as collaborators. The experimental data in the source article are self-measured by the project team, not third-party independent evaluations.

Training code is not open-sourced. Issue #3 still asks “when will release the training code.” Currently only inference code and pre-trained weights are available. The training process cannot be reproduced, and fine-tuning on your own data is impossible.

Dataset not released. AetherRetouch-1M+ is one of the paper’s core contributions, but Issue #1 shows that HuggingFace release is still pending. The construction method (inverse degradation workflow) of a million-level professional retouching dataset deserves independent validation, but currently it cannot be done.

Color grading still has deviations. Issue #4 is titled “The color grading result isn't perfect, mainly in terms of brightness” — brightness adjustments are not precise enough. This is an inherent limitation of the differentiable Renderer: the three-dimensional decomposition of the color space (Lighting/Global Color/Specific Color) cannot cover all retouching operations; local adjustments like dodge & burn are outside the current framework.

Limited reasoning capability of the 0.5B VLM. The source article mentions using DAPO-AE (reinforcement learning post-training strategy) to enhance autonomous aesthetic cognition, but a 0.5B-parameter VLM still has a ceiling in reasoning quality for complex scenes — it may not understand “the composition issue in this image causes a shift in visual focus” as well as a 7B model could.

Bus factor = 1. Apollo-Yi contributed 39 commits, being the sole core developer.

Verdict

VeraRetouch’s core contribution is not “yet another retouching model,” but proving that retouching can be fully differentiable end-to-end. The architecture of a 0.5B VLM + differentiable Renderer closes the loop of “reason → adjust → evaluate,” no longer relying on external black boxes.

For phone manufacturers and photo editing app developers, this path is worth following: small parameter count, deployable, interpretable. But the prerequisite is that training code and dataset must be open-sourced — otherwise it remains merely an internal tool for vivo, not reproducible community research.

For ordinary users, VeraRetouch means that future photo editing apps may no longer be black boxes with one-click filters, but transparent tools that can tell you, “This is what I adjusted and why.”

A scalpel is better than a black box, but only if you can get hold of the scalpel.

GitHub: https://github.com/OpenVeraTeam/VeraRetouch Paper: https://arxiv.org/abs/2604.27375