Overview
Build a consistent-face generation:
- Complete the InstantID installation layers listed above.
- Start from a working text-to-image graph.
- Add Apply InstantID between the prompt/model stage and the KSampler.
- Load a clear reference-face image.
- Add ControlNet Loader and select the installed InstantID ControlNet model.
- Add InstantID Face Analysis and select CUDA when using the CUDA GPU configuration.
- Add InstantID Model Loader and select the installed IP-Adapter file.
- Connect the modified model, positive conditioning, and negative conditioning back to the KSampler.
- Use a portrait-oriented latent size such as 768 by 1024 for the example.
- Enter a scene prompt and negative terms.
- Queue the workflow and inspect identity fidelity.
- Adjust Weight to change face influence.
- Adjust Start and End to change how long identity conditioning operates during denoising.
- Change the visual style in the positive prompt while retaining the same reference face to test character consistency.
The same node-based reasoning carries into newer checkpoints. AuraFlow uses the same basic text-to-image workflow with the checkpoint changed to an AuraFlow SafeTensors file. Flux uses a similar pipeline but requires changes to the model, CLIP loader, and KSampler configuration. SDXL is used throughout most of this module because its ecosystem is more mature, with a broad selection of models, LoRAs, plugins, ControlNet tools, and established workflows.