Train your own custom AI model directly from Telegram
LoRA (Low-Rank Adaptation) is a technique that teaches the AI to recognize and generate specific characters, art styles, or concepts. You provide reference images, the AI learns from them, and creates a reusable model that can be activated with a simple trigger word.
Cost: 50 credits per training | Requires: Premium membership (Basic+) | Duration: ~20-40 minutes
Once your LoRA is trained, it becomes available to all users in the /lora menu under the Community category. Your contribution helps grow the collection for everyone!
Type /trainlora in the bot. You'll see a template selection menu with different training types.
Select the template that matches what you want to train:
Enter a unique trigger word for your LoRA. This is the keyword you'll use in prompts to activate it. It will automatically be prefixed with amb_.
Example: You type mydalmatian → trigger becomes amb_mydalmatian
Tip: Choose something unique! amb_jennydog is better than amb_dog (too generic).
Enter a readable name that appears in the /lora menu. Example: Jenny The Dalmatian
Send your reference images to the bot. You can send them one by one or as albums. A counter shows your progress.
Click Finish when all images are uploaded. 50 credits are deducted and training begins. You'll receive live progress updates every 30 seconds showing the current epoch and a progress bar.
After ~20-40 minutes you'll receive a notification. Your LoRA is automatically registered and ready to use. Find it in /lora under the Community category.
Usage: /gen amb_yourtrigger, your prompt here
Quality matters more than quantity! 15 great images beat 50 bad ones.
Use images that are at least 1024x1024 pixels. Higher resolution means the AI can learn finer details like fur texture, eye patterns, and accessories.
Include front view, side view, 3/4 view, and back view. The AI needs to understand your character from all perspectives to generate it correctly.
Standing, sitting, running, happy, angry, relaxed. The more variety, the more flexible the LoRA becomes for different prompts.
Mix plain backgrounds, outdoor scenes, indoor settings. This teaches the AI to separate your character from any specific environment.
The subject should be clearly visible, well-lit, and take up most of the image. Avoid dark, cluttered, or heavily filtered images.
All images should be in a similar art style. Mixing realistic photos with cartoon drawings confuses the training. Pick one style and stick with it.
If you have character reference sheets (turnaround views, color palette, detail shots), include them! They're the single best image type for character training because they show the character consistently from multiple angles in one clean format.
Include a mix of close-up face/head shots and full-body images. This teaches the AI both the fine facial details and the overall body proportions and design.
After your LoRA is ready, test it with a simple prompt first: /gen amb_yourtrigger, solo, simple background. If the character looks right, try more complex prompts.
In /lora settings, you can adjust the weight (default: 1.0). If the LoRA effect is too strong (overrides the art style), lower it to 0.6-0.8. If too weak, raise to 1.2-1.5.
Always include your trigger word at the beginning of the prompt for best results. The trigger word tells the AI to activate the LoRA features.
Example: /gen amb_mychar, anthro wolf, standing in forest, detailed background, sunset lighting