Model Comparison
Side-by-side comparison of every face swap, face enhancement, and video upscaling model available in Recaster.
Recaster includes a range of AI models for face swapping, face enhancement, and video upscaling. Models are organized by tier: Free tier models are available to all users, while Studio models require a Studio license. All models are downloaded automatically on first use from Hugging Face.
Model Downloads
Face Swappers
Face swappers replace a target face with a source face. Higher-quality models produce more realistic results but require more processing time. Ghost models use ArcFace embeddings for superior identity preservation.
| Name | Tier | Variants | Quality | Description |
|---|---|---|---|---|
| InSwapper 128 | Free | 128 | 8.5/10 | Standard face swap model. Good balance of speed and quality using ONNX Runtime. |
| InSwapper 128-FP16 | Free | 128 (FP16) | 8.3/10 | Half-precision variant. Faster processing with slightly reduced precision. |
| Ghost | Studio | 1, 2, 3 | 9.2/10 | High-quality ArcFace-based swapping. Three versions with increasing quality and processing time. |
| BlendSwap | Studio | 256 | 9.0/10 | Advanced blending technology at 256px resolution. Excellent skin tone matching. |
Choosing a Swapper
Face Enhancers
Face enhancers improve the quality of swapped faces by restoring fine details, reducing artifacts, and improving skin texture. Enhancers run as a post-processing step after the face swap and can be blended with the original result.
| Name | Tier | Variants | Quality | Description |
|---|---|---|---|---|
| GFPGAN 1.2 | Free | 1.2 | 8.0/10 | GAN-based face restoration. The original version with solid general-purpose enhancement. |
| GFPGAN 1.3 | Free | 1.3 | 8.5/10 | Improved version with better detail preservation and fewer artifacts. |
| GFPGAN 1.4 | Free | 1.4 | 8.8/10 | Latest GFPGAN release. Best free-tier enhancer with excellent detail reproduction. |
| CodeFormer | Studio | — | 9.5/10 | Transformer-based enhancement. Exceptional quality with adjustable fidelity-quality balance. |
| GPEN | Studio | 256, 512, 1024, 2048 | 9.0/10 | Progressive enhancement network. Choose resolution variant based on input quality and speed needs. |
| RestoreFormer++ | Studio | ++ | 9.3/10 | Latest restoration architecture. Excellent at recovering fine details like pores and hair. |
Enhancement Controls
When using an enhancer, two controls affect the final output:
- Blend — Controls how much of the enhanced face is blended with the swapped face. A value of 0.8 means 80% enhanced, 20% original swap.
- Weight — Controls the strength of the enhancement effect itself. Lower values produce subtler enhancement.
Video Upscalers
Video upscalers increase the resolution of your output videos. These models process frame-by-frame and can be combined with the processing pipeline for denoising, sharpening, and color correction.
| Name | Tier | Variants | Quality | Description |
|---|---|---|---|---|
| Real-ESRGAN 2x | Free | 2x | 9.2/10 | Fast GAN-based upscaling. Doubles resolution with excellent detail generation. |
| Real-ESRGAN 4x+ | Free | 4x | 9.2/10 | Enhanced quality GAN upscaling. 4x resolution increase for HD to 4K conversions. |
| SwinIR 2x | Free | 2x | 9.7/10 | Transformer-based architecture. Best quality upscaler with superior detail preservation. |
| SwinIR 4x | Free | 4x | 9.5/10 | Transformer direct 4x upscaling. Excellent quality for large resolution jumps. |
| Real-ESRGAN 8x | Studio | 8x | 9.0/10 | Maximum resolution upscaling. Best for extreme upscaling of low-resolution footage. |
Upscaler Notes
- SwinIR vs Real-ESRGAN: SwinIR produces slightly higher quality output with better detail preservation. Real-ESRGAN is faster and produces excellent results for most use cases.
- Multi-pass upscaling: For 4x results, using two 2x passes often produces better quality than a single 4x pass. This option is available in the pipeline settings.
- SwinIR on macOS: SwinIR models fall back to CPU processing on macOS due to CoreML limitations. Use Real-ESRGAN for GPU-accelerated upscaling on Mac, or use remote upscaling for SwinIR quality.
Tier Summary
The table below summarizes model availability by tier.
| Category | Free | Studio |
|---|---|---|
| Face Swappers | InSwapper (2 variants) | + Ghost (3), BlendSwap (1) |
| Face Enhancers | GFPGAN (3 variants) | + CodeFormer, GPEN (4), RestoreFormer++ |
| Video Upscalers | Real-ESRGAN 2x/4x, SwinIR 2x/4x | + Real-ESRGAN 8x |
Was this page helpful?