AstroWitch - Cinematic Comic Style LoRA for MiniMax H3 by AstroBurner AI
AstroWitch is a custom style LoRA created for MiniMax H3, designed to bring a dramatic cinematic comic-book and graphic-novel aesthetic into AI-generated video.
The LoRA focuses on the visual language rather than a specific character.
Expect expressive faces, refined cel-shaded rendering, strong inked contours, dramatic contrast, rich shadows, cinematic lighting and a darker, elegant graphic-novel atmosphere.
It works especially well for:
Cinematic dialogue scenes
Action sequences
Romance and emotional scenes
Comedy and everyday interactions
Noir and thriller environments
Character close-ups
Dramatic urban scenes
Multi-character conversations
Stylized cinematic storytelling
The LoRA can be used for completely normal scenes. Your characters do not need armor, supernatural abilities or fantasy elements.
A woman drinking coffee, two people arguing in an office, a romantic conversation at a train station or three friends driving in a car can all be rendered in the AstroWitch style.
Trigger
Use:
ASTROWITCHV01H3
I recommend placing the trigger directly at the beginning of your MiniMax H3 prompt.
Example:
ASTROWITCHV01H3. Comic-book dark fantasy illustration with refined cel-shaded rendering, expressive faces, elegant inked contours, rich contrast, dramatic cinematic lighting, and a moody graphic-novel atmosphere.
After the style definition, describe the actual video normally.
MiniMax H3 Prompting
For MiniMax H3 I recommend treating the prompt more like video direction than a traditional image prompt.
A useful structure is:
Scene → Subjects → Initial State → Timeline → Visible Changes → Camera → Performance → Lighting → Sound
For example:
0-2s: Establish the characters and their initial expressions.
2-4s: Character A begins speaking while maintaining eye contact.
4-6s: Character B reacts, briefly looks away and answers.
6-8s: Their posture or emotional relationship changes.
8-10s: Finish with a reaction, camera move or final emotional beat.
Pay particular attention to:
Eye direction
Facial expressions
Micro-expressions
Body language
Speaker identity
Character reactions
Camera movement
Temporal progression
MiniMax H3 responds very well when the scene clearly explains who looks where, who speaks, how the other person reacts, and how the situation develops over time.
Simple Prompt Example
ASTROWITCHV01H3. Comic-book dark fantasy illustration with refined cel-shaded rendering, expressive faces, elegant inked contours, rich contrast, dramatic cinematic lighting, and a moody graphic-novel atmosphere.
Scene:
A quiet city café at night with wet pavement reflecting warm lights.
Subjects:
A long-haired woman wearing a fitted black coat stands opposite a man in a charcoal jacket.
Initial state:
She holds a coffee cup and looks directly into his eyes with a guarded expression. He watches her carefully.
Action timeline:
0-2s:
She briefly lowers her gaze toward the coffee cup before looking back at him.
2-5s:
She raises one eyebrow and says, "You always get quiet when something matters."
5-7s:
He looks toward the street for a moment, then returns his gaze to her and replies, "Maybe I just care what you think."
7-10s:
Her expression softens into a small genuine smile as she steps slightly closer. The camera slowly arcs around them while they maintain eye contact.
Camera:
Medium two-shot transitioning into closer reaction shots and finishing with a subtle cinematic arc.
Lighting:
Warm café lighting mixed with cool blue nighttime ambience and reflections from the wet street.
Soundscape:
Soft city traffic, distant café ambience and light wind.
Important
The LoRA was trained to reproduce a visual style, so you are free to build your own characters, environments and stories.
Fantasy elements are optional.
If you want a normal contemporary scene, simply describe a normal contemporary scene and let the LoRA provide the visual treatment.
If you want to push it further into dark fantasy, supernatural action or more extreme comic-book compositions, you can do that through the prompt.
Example Videos
The videos attached to this model page were generated with MiniMax H3 using this LoRA.
Where possible, I will include the prompts used for the examples so you can see how the LoRA and MiniMax H3 prompting work together.
Use them as a starting point, modify the characters, dialogue, camera direction and setting, and build your own scenes.
Have fun experimenting.
AstroBurner AI
Description
FAQ
Comments (15)
That's cool! looks like a telltale game, i'll try to render some shots in my setup.
That was exactly what i thought. But i used Witchblade comic images . Funny how a video chnages a style.
@Astroburner I'm DYING to know! How did you caption your training images??? 😢 Any advice on how you achieved this is so important! I'm giving you extra buzz for your time and ANY advice you have for us! ❤️
Start always with ASTROWITCHV01H3. Comic-book dark fantasy illustration with refined cel-shaded rendering, expressive faces, elegant inked contours, rich contrast, dramatic cinematic lighting, and a moody graphic-novel atmosphere.
After this Prompt text start to typ the standard MiniMax H3 Prompt structure.
Do you want a specific prompt maybe I can create it for you
@Astroburner
So just to be clear....your training images were captioned like this? Each and every one of them?
---
ASTROWITCHV01H3. Comic-book dark fantasy illustration with refined cel-shaded rendering, expressive faces, dramatic lighting, elegant inked contours, rich contrast, and a moody cinematic atmosphere. The style influence should appear through the visual treatment, color palette, facial design, shading, and environment mood, while the characters wear normal everyday clothing. Scene: A quiet city street café at night, just outside the entrance under warm hanging lights. The pavement is slightly wet from recent rain. Soft reflections shimmer on the ground. Through the windows, a cozy amber interior glow is visible. The atmosphere feels intimate, urban, and slightly noir. Subjects: A woman with long dark hair wearing a fitted black coat over a dark red top, simple earrings, and dark jeans. A man with short dark hair wearing a charcoal jacket over a dark button shirt. Initial state: The woman stands near the café entrance facing the man at a slight angle. She keeps one hand lightly around a paper coffee cup while the other rests near her coat pocket. Her posture is calm but attentive. The man stands opposite her, relaxed but slightly uncertain. At first, the woman looks at his face with a measured, thoughtful expression, while he looks at her, waiting for her response. Action timeline: 0-2s: A medium two-shot begins with a gentle lateral camera drift from left to right. The woman lowers her gaze briefly to the coffee cup, then lifts her eyes back to the man. Her expression softens from guarded to quietly amused. The man watches her carefully, searching her face. 2-4s: The camera slowly pushes in. The woman tilts her head slightly, raises one eyebrow, and gives a faint knowing smile. She looks directly into his eyes and says, "You always get quiet when something matters." 4-6s: Cut to a closer angle on the man. He glances away for a moment toward the wet street, then looks back at her with a restrained, slightly nervous smile. His shoulders loosen. He says, "Maybe I just care what you think." 6-8s: Cut back to the woman in a tighter medium close-up. She studies his face for a beat, then her smile grows warmer and more genuine. Her eyes move from his eyes briefly to his mouth and then back to his eyes, suggesting tenderness and curiosity. She steps half a pace closer and says softly, "That's a dangerous thing to admit." 8-10s: The camera makes a subtle arc around them into a more intimate angled two-shot. The man lets out a short quiet laugh and drops his gaze for a second before meeting her eyes again. She keeps her gaze on him, calm and slightly playful, with a small closed-mouth smile. A soft breeze moves a few strands of her hair. They hold the moment in a quiet charged pause as the café lights glow behind them. Visible state changes: The emotional tone shifts from cautious and reserved to warm, intimate, and lightly flirtatious. The woman’s facial expression evolves from composed and observant to amused and gently affectionate. The man’s expression evolves from uncertain and guarded to sincere and vulnerable. Their body language becomes subtly closer and more open over time. Camera: Begin with a medium two-shot. Use a gentle lateral drift in the opening. Add a slow push-in during the first line of dialogue. Cut to a close reaction shot of the man. Return to a tighter medium close-up of the woman for her response. Finish with a subtle arc into an intimate angled two-shot. Keep framing elegant, stable, and cinematic, with emphasis on eye contact, facial micro-expressions, and small body-language changes. Performance and expression: The woman should play the scene with subtle confidence, emotional intelligence, and a controlled but warm presence. Her facial acting is important: slight eyebrow movement, softening eyes, a gradual smile, measured pauses, and precise eye contact. The man should feel sincere and a little nervous, with small hesitations, brief downward glances, and a careful return to eye contact. Both characters should feel natural and grounded, with nuanced expressions rather than exaggerated acting. Lighting: Warm café light from behind and to the side, mixed with cool ambient street light and faint reflections from the wet ground. The lighting should create strong facial definition, attractive catchlights in the eyes, and a moody comic-noir elegance. Soundscape: Soft city ambience at night, distant traffic, a faint hiss from wet tires on the street, quiet café room tone from inside, and a subtle rustle of clothing in the breeze.
---
May I see one example caption from an image n your dataset so I can have a clear idea of how to caption my dataset images?
@wktra no way thats way to much xD. Everything what the lora learned is saved in the sentence. there is no secret how to write prompts for a lora training. Maybe i can link a video or 2 xD
@Astroburner I only need ONE example from one of your training images from your dataset please! 🥺🙏
@wktra I generally use a consistent activation tag ("ASTROWITCHV01H3") and a short recurring description of the style ("dark fantasy comic style"), while the rest of each caption describes the actual visible content of the individual training image — subject ("young woman"), clothing ("black dress"), pose ("standing"), composition ("medium shot"), lighting ("warm side light"), environment ("old stone hallway"), etc.
@Astroburner I appreciate the response. 😌
Now how much buzz would it cost me for you to give me an actual example from your dataset? 😅
How does this fare with REF2VA? I've tried it but its not getting the same precise look of your showcase.
The LoRA was trained only on Images for txt2video. I did Not test how and If it can be used for ref2video.
Yeah I figured that was the case, might be worth considering doing it for ref2video.. but I'll try the t2v out, looks great and deserves more recognition :)
@CypriotWolf yeah sounds like a good challange. I have to ready a bit how it would work :-)
Fantastic work. Any clue when Flat Vector / Game-Art Style LoRA for MiniMax H3 will be free? The exclusive access timer is missing and I can't comment there. Love this one, gonna keep making storylines so thank you!
The Lora will be always cost something because i set a license fee on it. Sorry to disappoint you. But these support helps me to develop further versions.