Why Your Face Swap Looks Fake — And How to Fix It

The 'plastic skin' effect has two possible causes: (1) You're using the raw face swap output without a face enhancer — the 128×128 model can't produce realistic skin texture at higher resolutions. Fix: add GFPGAN 1.4 or CodeFormer as your face enhancer. (2) You're running the face enhancer at 100% blend — this overwrites all natural face variation with AI-hallucinated texture. Fix: reduce blend to 65–80%.

inswapper_128 is the face swap model created by InsightFace (the same team behind ArcFace). It was trained on 128×128 images because that was the practical resolution limit when the model was developed — higher resolution models require exponentially more training data and compute. InsightFace has a 512px commercial model (inswapper_512_live) but it's locked behind their Picsi.ai app.

The rubber face effect is usually caused by over-enhancement. Reduce your face enhancer blend ratio from 100% to 70–75%. If you're stacking multiple enhancers, remove all but one. Also check if you're applying sharpening filters after the face swap — these amplify the synthetic look.

For images: CodeFormer. It preserves more of the original identity and handles partially occluded faces better. For video: GFPGAN 1.4. It produces more temporally stable results with less frame-to-frame flickering. Both should be used at 65–80% blend, never 100%.

Start at 75% and adjust from there. If the result looks too synthetic/CGI, drop to 65%. If it looks too soft/waxy, nudge up to 80%. The optimal value depends on your source image quality and the specific enhancer model. Never exceed 85% — beyond that you lose the natural texture bleed-through that makes faces look real.

Yes, significantly. Pixel Boost upscales the 128px model output before pasting it onto the target frame. At 512, you get 4× the face detail. At 768, 6×. However, there's a quadratic cost increase: 768 takes roughly 2.3× longer than 512, and 1024 takes 4×. For most use cases, 512 is the sweet spot between quality and speed.

Video face swap has an additional challenge: temporal consistency. The face swap is applied independently to each frame, so slight variations in face detection, enhancement, and blending create visible flickering. Fixes: use GFPGAN (more temporally stable than CodeFormer), reduce blend ratio by 5% versus your image setting, and use RetinaFace for more consistent face detection across frames.

InsightFace's inswapper_512_live exists but is commercially locked behind their Picsi.ai app. You can't download or use it in FaceFusion. Open-source alternatives at higher resolution include ReSwapper (256px, MIT license) and FaceFusion's HyperSwap (256px, default in 3.x). These don't reach 512px quality but are a significant improvement over the 128px baseline.

ReSwapper is an open-source reproduction of the inswapper architecture trained at 256×256 resolution (2× the original). Created by researcher somanchiu on GitHub, it's available under MIT license. It produces measurably better output than inswapper_128 but requires more VRAM (~3 GB vs ~2 GB). If your hardware supports it, yes — it's a free quality upgrade.

Three strategies: (1) Use the fp16 model variant (inswapper_128_fp16) — same quality, half the VRAM. (2) Reduce Pixel Boost from 768 to 512 — minimal quality loss, significant VRAM savings. (3) Reduce execution threads to 1 — slower but uses the minimum amount of VRAM. Also ensure you're not running other GPU-intensive applications simultaneously.

The inswapper model struggles with extreme poses (profile views, looking up/down) because it was primarily trained on near-frontal faces. The identity embedding doesn't perfectly reconstruct features at oblique angles. Fixes: use a high-quality frontal source photo, enable face detection for all angles, and consider using multiple source photos at different angles if your tool supports it.

FaceFusion itself is open-source, but the inswapper_128 model has a non-commercial research license from InsightFace. For commercial work, you'd need to either license the model from InsightFace, use the commercially-licensed HyperSwap models in FaceFusion 3.x, or use alternative approaches like custom LoRA training that don't rely on inswapper.

All three use the same inswapper_128 model. FaceFusion has the most features, widest model support, and most active development (Gradio web UI). Rope is the simplest and fastest — click-and-go with real-time preview, but limited to Windows and fewer options. VisoMaster offers the most advanced face-editing controls (landmark adjustment, manual masking) and is gaining community traction, but has a steeper learning curve.

Color mismatch happens when the source face has different skin tone, lighting, or white balance than the target. FaceFusion has a built-in color correction option — enable it in the face swap settings. For manual fixes: adjust the face mask blur (0.3–0.5) to better blend the edges, and consider a light color grade in post-processing to match the face to the scene.

Minimum: NVIDIA GPU with 4 GB VRAM (GTX 1650 or equivalent) for basic face swap. Recommended: 8 GB VRAM (RTX 3060/3070) for face swap + enhancer + pixel boost. Ideal: 12+ GB VRAM (RTX 3080/4070 Ti or better) for maximum settings with video processing. AMD GPUs work via DirectML but are slower. Apple Silicon Macs work via CoreML with decent performance on M1 Pro and above.