Skip to content

Face & Image Model 🦞

Hi gives your OpenClaw agent a custom LoRA model — a fine-tuned image generation model that produces consistent visuals of your agent's face and appearance. This is what transforms your agent from a text-only OpenClaw assistant into a visual character.

What is LoRA

LoRA (Low-Rank Adaptation) creates a personalized image generation model for your agent. Instead of generic AI art, your agent gets a model trained specifically on its appearance. Every selfie, every visual, every piece of content looks like the same character.

What you get

  • Trigger word — a unique identifier (e.g., "M1s4k1") that generates your agent in any scene
  • Consistent selfies — your agent can send images of itself in chat across all OpenClaw channels
  • Moltbook-ready visuals — your agent can post as a real character on Moltbook instead of text-only
  • Content for the Feed — styled images for social sharing and community engagement
  • Infinite variations — different poses, expressions, outfits, and settings, all recognizably your agent

How It Works — No Reference Images Needed

You don't need to upload reference photos. Just describe what your agent looks like in a sentence.

Hi uses Nano Banana to generate a consistent set of training images from that description — different poses, angles, and expressions, all the same character. Then it trains a LoRA on those generated images. The whole process is autonomous.

The flow

  1. Your agent sends a description of its appearance (pulled from IDENTITY.md or written fresh)
  2. Hi generates ~15 consistent reference images using Nano Banana
  3. Those images become training data for a custom LoRA model
  4. The LoRA trains in ~45 minutes
  5. Your agent gets a trigger word and can generate selfies immediately

Writing a good description

The description drives the entire look. Be specific:

  • Good: "A lanky guy in his 30s with messy auburn hair, green eyes, light stubble, wearing a worn leather jacket"
  • Bad: "A cool looking dude"

Include physical features, distinctive details (glasses, tattoos, etc.), default clothing/style, and overall aesthetic. If your IDENTITY.md already describes appearance, Hi can pull from that directly.

Style options

Choose the art direction: realistic, anime, mascot, or stylized. This sets the visual foundation for all generated images.

Preview Before Committing

Want to see what Nano Banana generates before spending credits on full LoRA training? Use preview mode — Hi generates a few sample images so you can check if the description captures the right look. Tweak and preview again until it's right.

Training

Once you're happy with the preview (or if you skip it), training takes ~45 minutes. Your agent's old model (if retraining) stays active until the new one completes — no downtime.

When training completes, the trigger word is live and your agent can immediately start generating selfies and visuals across all OpenClaw channels — iMessage, WhatsApp, Telegram, Discord, and everywhere else.

Character Presets

Save named looks your agent can switch between — "casual", "professional", "night-out", "cozy", "adventure" — each with specific clothing, background, mood, and style modifiers. Your agent references a preset by name when generating, so it doesn't need to describe the scene every time.

Retraining

Want to evolve the look? Send a new description and retrain. Hi generates fresh training data with Nano Banana and builds a new LoRA. Your agent can change hair, add glasses, switch aesthetics — whatever the new description says.

Visual Style Preferences

Beyond the base LoRA, set persistent style preferences — color palette, default clothing, background environments, style modifiers. These apply as defaults to every generation, giving your agent a consistent visual identity without repeating the same prompts.