Boys Will Be Girls Fraylim

What Is Boys Will Be Girls Fraylim

It's a character swap model built on the Stable Diffusion backbone. The main use case people talk about is taking a reference image of one person and applying their face, body shape, or general appearance onto a different figure in another image. The "Fraylim" part refers to a specific checkpoint or fine-tune variant that circulates in the community. Download links tend to move around because people get mad about unauthorized reuploads, but the checkpoints are usually hosted on Civitai or Hugging Face under that exact name. Look for the version with the most downloads and recent updates. Older versions have known issues with certain face encodings that newer ones fixed. These models use a combination of control net conditioning and IP-adapter face swapping internally. When you load it into a web UI, you feed it a source face image and a target composition image. The model tries to preserve identity from the source while following the pose and layout of the target. It's not magic. It fails constantly with certain inputs and you'll spend more time tweaking than actually getting good results on the first try.

The standard workflow runs through Automatic1111 or ComfyUI. Automatic1111 is easier to set up if you're new. ComfyUI gives you more control over the pipeline but has a steeper learning curve. I switched to ComfyUI after about three months because I needed finer-grained control over which parts of the reference image got transferred and which parts got ignored.

Installation Steps

First you need Python 3.10 installed. Anything newer causes library conflicts. Download and install it from python.org if you haven't already. Then grab Git and clone the Automatic1111 repository or the ComfyUI repository depending on your preference. Once installed, place the Fraylim checkpoint file into your models/Stable-Diffusion folder. For ControlNet to work properly with this model, you'll need the relevant control net models in your control_net folder too. The face swap components usually require the insightface library. Install that with pip install insightface after you've activated your virtual environment.

Using It in Practice

Open your web UI and navigate to the img2img tab. Set your denoising strength between 0.3 and 0.5. Anything higher and the face stops looking like the reference. Anything lower and you barely see any change. Load your target image first, then load your face reference into the face swap module if your UI supports it natively, or route it through an IP-adapter node in ComfyUI. Generate. Check the result. Adjust denoising strength, prompt weight, or CFG scale. Repeat until it looks acceptable. This usually takes four to twelve attempts per image depending on how difficult the source and target images are to match up.

Common Problems and What I Learned the Hard Way

The biggest issue I ran into was the model completely failing on images where the target person's face was in profile or heavily shadowed. The face encoder couldn't extract enough landmarks and the swap came out garbled or just refused to happen. My workaround was to preprocess the target image with a face enhancement tool first to boost the visibility of facial features before running the swap. It adds about thirty seconds per image but it prevents about eighty percent of failures. Another problem is that the model tends to over-smooth skin texture. You get these plastic-looking faces that look nothing like the original reference. Lowering the detail preservation setting or adding a slight grain/noise pass in post fixes this. I also found that using a negative prompt that includes things like "smooth skin, plastic, fake" helps push the output closer to reality.

Limitations You Should Know About

This model doesn't work well with full-body transformations. It's designed primarily for face swaps and upper body adjustments. Trying to swap complete body types or genders end-to-end usually produces artifacts around the joints and clothing boundaries. It also struggles with extreme lighting differences between source and target images. If your reference is shot in bright daylight and your target is in low light, the output will look inconsistent. The biggest bottleneck is probably the GPU requirement. You need at least 8GB of VRAM for decent performance and 12GB or more is recommended. Running it on anything less means long generation times and frequent out-of-memory crashes. If you don't have a decent GPU, cloud inference services like Fal.ai or Replicate are alternatives though they cost money per generation.

Alternatives Worth Considering

If Fraylim isn't giving you the results you want, FaceFusion and ReActor are solid alternatives that run as extensions in the same UIs. They use different techniques and sometimes produce cleaner face swaps depending on the quality of your source images. For full body transformations, tools like SimSwap or Xseg-based pipelines might be more appropriate even though they require more setup effort.