CIVITAI / Workflows

Z-Image-i2L (Image to LoRA) Fast LoRA Training + ControlNet Testing + Upscale Workflow

This workflow is an expanded Z-Image-i2L production pipeline that combines fast Image-to-LoRA generation, ControlNet structure testing, and high-resolution tiled upscaling into one complete ComfyUI graph. It is designed for creators who do not only want to generate a quick LoRA from reference images, but also want to immediately test that LoRA under real production conditions and then push the result into a more polished final output. The first stage focuses on fast LoRA creation. Multiple reference images are loaded and combined into a training image batch, then passed into the RunningHub Z-Image-i2L system. This allows the workflow to generate a lightweight Z-Image LoRA from a small group of images without requiring a traditional local training setup, dataset folder preparation, caption files, or command-line configuration. It is especially useful for quickly capturing a character identity, fashion style, product look, object concept, creature design, or consistent visual aesthetic. After the LoRA is generated, the workflow immediately saves it and loads it back into Z-Image Base for testing. This is the key advantage of the pipeline: training and validation happen in the same graph. The user can quickly see whether the generated LoRA actually affects the output, whether it preserves the target identity, whether it introduces artifacts, and whether the strength needs to be adjusted. This makes the workflow much more practical than a training-only setup. The second stage adds ControlNet testing. A structure reference image is processed through DepthAnythingV2Preprocessor to create a depth map, then applied through Z-Image Fun ControlNet Union. This lets the newly generated LoRA be tested under controlled composition, depth, layout, and spatial guidance. A LoRA may look fine in a basic text-to-image test, but fail when the camera angle or scene structure becomes more demanding. This workflow helps reveal that immediately. The generation section uses Z-Image Base with qwen_3_4b text encoding, AE VAE, ControlNet guidance, SplitSigmas, DetailDaemonSamplerNode, CFGGuider, and SamplerCustomAdvanced. This gives the workflow a more controlled two-stage sampling structure, where the early phase builds the main layout and the later phase refines the image. It is useful for evaluating prompt compatibility, LoRA strength, structural stability, and final image coher