wtf is headhunter?

HeadHunter helps you get better images out of AI image generators like Stable Diffusion 3 and FLUX. When these models generate an image, they use internal components called "attention heads" to decide what the picture should look like. The project's key insight is that individual attention heads actually specialize in different visual qualities, one head might control lighting, another might shape structure, and another might handle texture or style. Instead of applying a uniform guiding force across the whole model, HeadHunter identifies which specific heads to tweak and combines them to achieve a desired visual effect. The framework works by running an iterative search. It systematically perturbs individual attention heads, generates test images, and then scores those images using automated quality metrics like PickScore or CLIPScore. Over several rounds, it narrows down the best combination of heads to adjust for your specific goal. The project also introduces "SoftPAG," a method that lets you control how strongly these heads are perturbed, preventing the final image from looking oversaturated or unnatural. This tool is designed for researchers and builders working with diffusion models who want fine-grained control over generated images. For example, if you are building a creative app and want to consistently push your model to produce images with a specific artistic style or better structural coherence, HeadHunter automates the process of finding the right internal settings. It is essentially an optimization pipeline: you provide the text prompts and a scoring preference, and it finds the exact internal heads to adjust for the best possible output. The project is built as a multi-GPU distributed pipeline, meaning it can spread the heavy computational work of testing different head combinations across multiple graphics cards. This is necessary because testing every combination of attention heads across a model's many layers would be extremely slow on a single machine. The codebase supports a variety of specific perturbation strategies, like zeroing out certain attention blocks or replacing attention patterns with identity matrices, giving advanced users deep flexibility in how they steer the model.