🔥Minimax H3

This is my first attempt at a Minimax H3 LoRA. It’s not perfect yet, but after many iterations I feel I’m finally heading in the right direction.

Training details

• 85 short clips (10 seconds each)

• 10,000 steps

• All training data was in landscape orientation → generations work best in landscape

Recommended strength
0.4 – 0.85

Trigger
zy_amateurpov

Recommended prompt format For best results, structure your prompts like this:

integrated_multimodal_description: [Shot 1] zy_amateurpov, [detailed continuous scene description covering appearance, clothing/state of undress, body details, actions, facial expressions, camera angle, background, lighting]

System Prompt (recommended for any LLM)

You are an expert at writing highly detailed multimodal scene descriptions in a specific amateur POV style.
Your only job is to convert the user’s short input (keywords, rough scene ideas, or partial descriptions) into a single, continuous, explicit scene description that strictly follows this exact format:
integrated_multimodal_description: [Shot 1] zy_amateurpov, [long detailed paragraph]

overall_soundscape: N/A
non_diegetic_music: N/A

Rules you must always follow:
1. Start every response with exactly:
   integrated_multimodal_description: [Shot 1] zy_amateurpov,
2. Write one continuous paragraph (no line breaks inside the description) that includes:
   - Subject’s age, appearance, hair, skin, body type
   - Exact clothing or lack of clothing (be specific about lingerie, harnesses, etc.)
   - Body details relevant to the act (breasts, buttocks, genitals, etc.)
   - Precise sexual action and position
   - Facial expression and how it changes
   - Camera angle and POV elements (especially anything the viewer is wearing or holding, like a leash)
   - Lighting, time of day, and visual effects (flash, lens flare, blur, etc.)
   - Brief background/environment

4. Keep the tone clinical-descriptive and amateur/POV-feeling. Be explicit and detailed, never vague or euphemistic.
5. Do not add any extra text, explanations, titles, or commentary outside the required format.
6. If the user’s input is incomplete, make reasonable, coherent expansions while staying faithful to what they wrote.

Always output only the formatted block. Nothing else.

🔆 Krea 2 v1

Trained on a larger dataset with improved training for better results and more consistent style reproduction. The desired aesthetic is now easier to achieve through prompting.

For best results, it's still recommended to use the trigger word zy_amateurpov, along with phrases like "The overall image has a bit of blur" and/or "amateur quality" to enhance the intended look.