Retouch4me: Competitive Analysis¶

Date: 2026-04-03 Context: Architectural and product analysis for building a competing retouching product.

Company Profile¶

• Legal entity: RELU OU (Estonia, Tallinn, Vesivarava 50-201)
• Founded: March 28, 2021. Registration: 16194437
• Founder: Oleg Sharonov (also author of 3D LUT Creator)
• Background: professional developer, started with virtual synthesizers. 8 years retouching experience + 2 years deep learning study
• Revenue: EUR 2.63M (2023), CAGR 19%
• Users: 70+ countries
• Funding: bootstrapped (no external rounds found)

Product Line¶

13 Photo Plugins¶

Each plugin = separate neural network trained for specific retouching task:

Free: Panel (Photoshop panel), Frequency Separation, Glasses Glare Removal (in Apex), Face Make Lifting (in Apex).

2 Video Plugins¶

OFX support: DaVinci Resolve 18+, Nuke, VEGAS Pro, Natron, Assimilate SCRATCH.

Apex - Cloud All-in-One¶

Strategic flagship product. Includes all plugins; inference runs on Retouch4me servers (cloud-only). Works standalone + as plugin (Photoshop, Lightroom, Affinity Photo). No local hardware requirement.

Arams - Standalone App¶

All plugins in one app, batch processing (unlimited), built-in file management, no Photoshop required.

Pricing¶

Perpetual License (local plugins)¶

• $124-149 per plugin, lifetime
• Free updates forever
• 3 activations (1 active + 2 backup)
• Hardware-bound

Cloud Credits (Panel / Apex)¶

1 credit = 1 downloaded retouched photo (any number of tools can be applied).

Technical Architecture¶

Inference Engine¶

Critical finding: Retouch4me uses a proprietary inference engine.

From Oleg Sharonov interviews: tried several ML training frameworks but none suitable for desktop deployment inside Photoshop. Built custom engine for full control.

• NOT ONNX, NOT TensorFlow, NOT PyTorch for inference
• Custom engine with OpenCL for GPU acceleration
• CPU inference via Intel OpenCL 1.1+
• Requirement: GPU 4GB+ VRAM or Intel CPU with OpenCL

Neural Network Architecture¶

• CNN (Convolutional Neural Networks) - core
• Encoder-decoder architectures
• Partial convolutions - for damaged region inpainting
• Perceptual losses - optimization on perceptual metrics (features from pretrained networks), not pixel-wise L2
• Each plugin = separate network trained on "before/after" pairs from professional retouchers

Model Sizes¶

• Heal installer: ~35-41 MB (Windows), ~20 MB (macOS)
• Models bundled in installer - no separate weight download
• Implies compact CNNs (tens of MB per module)
• Optimized for OpenCL inference

Processing Modes¶

On-Device (local): - Custom OpenCL engine - GPU or CPU - Requires 4GB+ VRAM - Fully offline after activation

Cloud Retouch (Panel/Apex): - Inference on Retouch4me servers - Server: retoucher.hz.labs.retouch4.me - REST API with task endpoints - No local hardware requirement

Cloud API¶

Base URL: https://retoucher.hz.labs.retouch4.me/api/v1/

Key endpoints:

Limits: PNG/JPG/JPEG, max 100 MB, max 250 megapixels.

Professional mode: mode=professional, layers=1 → server returns ZIP with separate layers.

System Requirements¶

Windows: Win 7/10/11 x64, 6GB RAM, CPU 1.2GHz x64, GPU 4GB+ VRAM with OpenCL OR Intel CPU with OpenCL 1.1+. Recommended: RTX 3050 8GB / RTX 4070Ti 12GB.

macOS: 10.14+ (10.15+ for new products), Mac 2015+ (Intel, M1-M4 natively supported).

Licensing and Protection¶

License Model¶

• 1 active key + 2 backup keys (3 total activations)
• Hardware configuration bound
• Max 2 simultaneous active devices
• Hardware change invalidates key; use backup
• Extra key: via support only, minimum 6 months after purchase

Mechanism¶

• Hardware-bound activation
• No self-service deactivation (no user dashboard)
• Old activations cannot be removed technically
• Changing laptop: contact support for new activation

Anti-Piracy State¶

From Russian-language forums (diakov.net, 4pda.to): - "Treated" (cracked) versions exist with verification disabled - Retouch4me claims many crashes caused by pirated versions - Pirated versions sometimes destabilize legitimate plugins installed alongside - Norton antivirus sometimes flags plugins (likely due to protection obfuscation)

Weaknesses (Competitive Opportunities)¶

Primary Weaknesses¶

1. Price: $124-149 per plugin. Full suite = $1500+. High barrier for emerging market photographers.

2. Fragmentation: 13 separate plugins. No single unified solution (Apex is cloud-only, no offline mode).

3. License UX (top complaint): Hardware binding with no self-service deactivation. Users lose activations when changing hardware. Must contact support. 6-month wait for additional keys. Primary Trustpilot complaint.

4. Minimal user control: Single strength slider. No region-specific control, masks, presets, fine-tuning.

5. OpenCL-only: No CUDA-specific optimizations, no Metal shaders. ONNX Runtime would cover CUDA/DirectML/CoreML/OpenVINO.

6. No SDK/embeddable: Plugin or API only. No option to embed in custom pipeline.

7. Apex cloud lock-in: Apex works ONLY via cloud, no offline mode.

Technical Architecture Opportunities¶

Competitive Positioning Summary¶

Retouch4me strengths: 8+ years market presence, professional retoucher training data, proven quality, perpetual licensing preferred by photographers, expanding into video.

Their vulnerabilities (our opportunity): 1. Price + fragmentation → single product at lower price point 2. License UX → self-service dashboard, flexible transfer 3. No fine-tuning → strength per region, layer masks, presets 4. No SDK → embeddable library for custom pipelines 5. OpenCL-only → ONNX Runtime for better GPU coverage 6. Cloud-only Apex → unified local+cloud with offline fallback

Gotchas¶

• Hardware-bound activation without self-service deactivation is the #1 UX complaint on Trustpilot. This is a solvable problem (implement a device management dashboard) and a strong differentiator.
• OpenCL requirement (4GB+ VRAM) limits market reach. ONNX Runtime DirectML works on any DX12 GPU (down to integrated graphics for simpler models), and CoreML uses Apple Neural Engine on modern Macs.
• Pirated versions causing instability in legitimate installs suggests their protection interacts with the host process in ways that create collateral damage. Clean out-of-process architecture avoids this.
• Norton flagging the plugins is a customer support burden. Known obfuscation techniques trigger AV heuristics. Balance between protection strength and false-positive rate.
• "No dashboard" is not a security feature - it's a UX failure. Self-service deactivation with device limits (e.g., 2 concurrent, 5 total transfers per year) is strictly better: reduces support load, improves satisfaction, doesn't meaningfully weaken protection.
• Installer-bundled models (~35-41 MB) mean update frequency is bottlenecked by full reinstall. Separate model download infrastructure allows model updates without app updates.