RealGen is a LoRA trained to improve photorealism. It focuses on reducing common synthetic artifacts (such as overly smooth skin, plastic textures, and unnatural lighting) and producing images with more natural detail, texture variation, and camera-like imperfections. The underlying training approach uses detector-guided rewards and adversarial feedback to bias the model toward outputs that resemble real photographs.
It works best for general photorealistic scenes, portraits, and everyday environments where subtle realism matters. It is not designed for stylized, illustrative, or heavily artistic outputs, and may reduce stylistic expression in favor of realism.
Notice: I did not train this model. The original researchers released it in a Diffusers-only format, so I simply converted the weights for ComfyUI. All credit for the training goes to the original authors.