MagicMakeup Transfer¶

Development line: project:magicmakeup · thread magicmakeup-transfer
Last event: 2026-07-28 · 1 dated since 2026-07-28 · Researched: 2026-09-05 · confidence: high

What it is¶

MagicMakeup Transfer is an image-to-image makeup-transfer system with region control for research users who prepare face crops and masks.

• Reference transfer: transfers full-face, eye, and lip appearance from a reference portrait.
• Model checkpoint: uses a MagicMakeup checkpoint on FLUX.1-Kontext-dev.
• Output resolution: produces 1024 × 1024 crops and supports single-pair or all-to-all batch inference.

Development line¶

• 2026-07-28 — MagicMakeup Transfer public project resources linked. A region-controllable diffusion-transformer pipeline with face, eye, and lip transfer, built on FLUX.1-Kontext-dev.

What changed¶

2026-07-28 — The official code and checkpoint release arrived. MagicMakeup became usable as a region-controllable diffusion-transformer pipeline with face, eye, and lip transfer, built on FLUX.1-Kontext-dev.

How to use this¶

Start with the project page, repository, and hosted model resource as of 2026-07-28. Check those primary materials so exact versions and usage details stay verified.

1. Clone the official repository and create its recommended Python 3.10 environment with PyTorch 2.6.0/CUDA 12.4 and the listed dependencies. — https://github.com/vivoCameraResearch/Magic-Makeup
2. Accept the FLUX.1-Kontext-dev license, download that base model, then download the Anyou/MagicMakeup checkpoint. — https://github.com/vivoCameraResearch/Magic-Makeup
3. Place source and makeup-reference portraits in separate directories. Crop primary faces to centered 1024 × 1024 images and retain the preprocessing log. — https://github.com/vivoCameraResearch/Magic-Makeup
4. Generate matching face, eye, or lip masks for both images. Run test_single.py with the desired --label value: eyes, lip, or eyes,lip,face. — https://github.com/vivoCameraResearch/Magic-Makeup
5. Confirm filename-stem matches for every image and mask before running test_dir.py. The script pairs every source with every reference and skips unmatched masks. — https://github.com/vivoCameraResearch/Magic-Makeup

Best practices¶

• Start with a single pair and inspect the comparison panel before an all-to-all batch. Batch mode skips pairs when crop and mask mismatches occur. — https://github.com/vivoCameraResearch/Magic-Makeup
• Use the default model-offload mode first. Choose sequential CPU offload only when GPU memory is constrained, because it slows inference. — https://github.com/vivoCameraResearch/Magic-Makeup
• Keep use within the stated non-commercial academic cosmetics-research scope. Do not repurpose it for face swapping, identity recognition, impersonation, or deceptive edits. — https://github.com/vivoCameraResearch/Magic-Makeup

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• We found no versioned GitHub release or dated checkpoint publication, so the exact publication time of the code and checkpoint cannot be separated from the 2026-07-28 release event.
• The Hugging Face page exposes a generic Diffusers example that omits MagicMakeup’s mask-conditioned pipeline. The official repository remains the authoritative inference route.

Sources¶

Agent brief¶

• Subject: project:magicmakeup, thread magicmakeup-transfer, 1 dated events 2026-07-28 → 2026-07-28.
• Practical note: Start with the project page, repository, and hosted model resource as of 2026-07-28. Verify exact versions and usage details from those primary materials.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.