- Nov 29, 2022
- 2,477
- 284
Affiliate marketers working with verticals such as crypto, gambling, or nutra have likely come across the use of deepfake technology. The idea of featuring a recognizable public figure to endorse an offer—provided this does not violate regulations or platform policies—can sound highly attractive. However, outsourcing this work is often risky. Prices can be high, and the outcome unpredictable: videos may look unprofessional, with faces that lack natural movement or eye tracking. In some cases, unreliable providers disappear after taking advance payments.
Finding skilled freelancers is difficult, expensive, and comes with uncertainty. The alternative is to master the process yourself. By investing in a capable workstation and dedicating time to learning, you gain full control and the ability to generate creatives at scale. This article explains how to move away from outsourcing and produce professional-level deepfake assets independently.
From entry-level tools to professional software
The market offers a wide range of solutions, which can be grouped into categories depending on complexity and use case.Web services and mobile applications (entry level)
This includes platforms such as HeyGen, D-ID, Hedra, Akool, Reface, and similar mobile-first tools. The workflow is straightforward: upload a photo, provide a text script, and receive a talking-head video.
Requirements: any internet-connected computer
Learning curve: approximately 10 minutes
Quality: from average to good, but largely template-based
Verdict: for affiliate marketing, these are basic tools. Quality is inconsistent, control is limited, and subscription restrictions are common. Suitable for casual use but insufficient for high-performance creatives designed to deliver strong CTR and conversion rates.
Professional desktop solutions
This is the preferred option for serious work. Dedicated desktop software provides full control and maximum quality. The two leading open-source solutions are:DeepFaceLab
The most advanced open-source tool available for Windows and Linux. It requires Python, TensorFlow, and NVIDIA CUDA. Capabilities include face replacement, de-aging, head replacement, and more. Complexity is high, but quality can reach professional standards when properly executed. Work is script-driven, which provides granular control but demands time and technical understanding.
FaceSwap
Another open-source solution for Windows, macOS, and Linux, built on TensorFlow and Keras. Slightly easier to use than DeepFaceLab due to its graphical interface, while still offering high quality results. An excellent entry point for those uncomfortable with command-line operations.
Hardware requirements
Deepfake generation is highly resource-intensive and relies on powerful NVIDIA GPUs. Lightweight laptops are insufficient. Hardware needs depend on your intended workflow.Mid-level setup (sufficient for learning)
CPU: Intel i5 / AMD Ryzen 5 or higherRAM: 16 GB+
GPU: NVIDIA GTX 1060+ (RTX series recommended). The optimal entry point today is the RTX 3060 with 12 GB VRAM.
Storage: 50 GB+ free space on SSD
This setup allows production of deepfakes, but model training will be slow, often taking days.
Professional setup (for high-volume production)
CPU: Intel i7/i9 or AMD Ryzen 7/9RAM: 32 GB+
GPU: NVIDIA RTX 3080/4080+ with 16–24 GB VRAM (RTX 3090/4090 ideal)
Storage: 500 GB+ NVMe SSD
With this configuration, models can be trained in hours instead of days, enabling continuous production at maximum quality. Proficiency in video editing software (After Effects, DaVinci Resolve) and a basic understanding of machine learning is strongly recommended.
Importance of source quality and training time
Even with high-end hardware, poor source material results in poor outputs. Two factors are critical:Source quality: the final result depends heavily on input data. Videos should be high-resolution (4K+), stable, well-lit, and free of visual artifacts.
Dataset size: at least 500 clean frames are required for basic results; 2,000+ for good quality; professionals use datasets of 10,000+ frames.
Training time also plays a central role. Quick training of 1–4 hours yields recognizable but artificial-looking faces. Standard training of 8–24 hours produces solid results. Professional-level realism requires 48+ hours of uninterrupted GPU work.
Workflow: from source video to final creative
Taking DeepFaceLab as an example, the workflow consists of three stages:Data Preparation
Place two videos — data_src.mp4 (source) and data_dst.mp4 (target actor) — into the workspace folder.
Run extract images from video data_src.bat and extract images from video data_dst.bat to split the videos into frames.
Next, run data_src faceset extract S3FD.bat and data_dst faceset extract S3FD.bat to extract the faces.
The most important step: run data_src sort.bat and then manually clean the dataset through data_src view aligned result.bat. Remove all unnecessary frames such as blurred, dark, obstructed, or poorly angled faces.
Model Training
Run train SAEHD.bat. In the console, set parameters. For beginners, most default settings are sufficient except for batch_size, which may require adjustment.
Monitor the preview window and the “loss” values in the console. Your goal is to wait until the preview face becomes sharp and the loss values stabilize. This process constitutes training and takes the majority of time.
Merging and Post-Processing
Run merge SAEHD.bat. An interactive window will appear, allowing you to overlay the trained model onto the target video. Use hotkeys (W-S, E-D, etc.) to fine-tune the mask, color, and smoothing.
Once the optimal settings are found, apply them across the entire video sequence.
Finally, run merged to mp4.bat. This script compiles all frames into the final result.mp4.
An essential but often skipped step is post-processing. Import the output into a video editor. Apply color correction to balance skin tones, add a slight film grain to conceal minor artifacts, and if the creative concept allows, apply subtle background blur. Without this stage, the deepfake almost always looks unnatural and artificial.
Common mistakes leading to detection
Poor source material or insufficient dataLighting mismatches between replacement and original scenes
Skipping post-processing steps such as color correction
Stopping training too early, producing blurred or distorted results
Incorrect model configuration or poor parameter choices
Unnatural edges and lifeless eyes caused by weak datasets or inadequate training