Best Image Models to Train Loras On (mesmer.tools)

In a recent evaluation of image models for training Low-Rank Adapters (LoRAs), Ideogram 4 emerged as the top performer, particularly for enhancing likeness in both easy and challenging faces. The model outperformed others by effectively utilizing a native JSON caption format, significantly improving image quality and resemblance. Following Ideogram 4, Krea 2 climbed to second place after transitioning its sampling method to Krea 2 Turbo, which notably enhanced output quality. This ranking shift underscores the impact of training methods and input formatting on AI model performance. The findings are particularly significant for the AI/ML community focused on character and headshot representations. The comparison highlighted that while Flux.1 Dev remains a solid option, it is reaching its limits, indicating an opportunity for model advancements. The experiment revealed critical insights into training methodologies, including the importance of using novel prompts to assess model overfitting and ensuring proper caption structuring to maintain identity resemblance. Such benchmarks are invaluable for developers aiming to refine their pipelines and enhance image generation technologies effectively.