Z-image character lora great success with onetrainer with these settings.

A Stable Diffusion community member has shared a breakthrough method for training character-specific LoRAs (Low-Rank Adaptations) that's producing remarkably consistent and photorealistic results. Using Nerogar's OneTrainer software, the approach centers on a specific configuration file called 'z-image-base-onetrainer.json' and a structured captioning system. The creator trained their model on just 50 images, each captioned with a specific format: character name, expression, pose, angle, clothing, and background described in 2-3 words each. This structured data approach appears to give the model much better understanding of character consistency across different scenarios.

What makes this method particularly notable is its optimization for 1024 resolution images, whereas most existing guides focus on 512 resolution training. The creator also shared their complete workflow via Pastebin, which includes using a distill LoRA for faster 8-step generation at Full HD resolution. For those seeking specific aesthetics, they recommend the euler_cfg_pp sampler with beta33 scheduler (available via ComfyUI_PowerShiftScheduler) to achieve an 'Instagram aesthetic' look. The method reportedly handles diverse poses, angles, expressions, and compositions effectively as long as the training dataset includes appropriate captions for those variations.

The creator has shared all configuration files and settings publicly, encouraging others to experiment with the approach. They specifically mentioned looking for similar optimization techniques for Chroma, another AI image generation model, suggesting this methodology might be transferable across different platforms. The detailed sharing of exact settings, samplers, and workflow represents the collaborative spirit of the open-source AI art community, where incremental improvements in training techniques can lead to significant quality leaps for all users.