📱 GalaxyAce LoRA

🎬 MiniMax H3

Model: MiniMax H3 (open weights, H3-Base)

Recommended LoRA Strength: 1.00

Trigger Word: Not required

This LoRA brings the Galaxy Ace look to MiniMax H3 — the color reproduction, exposure

behavior and lens softness of an early 2010s low-end Android phone, now on a model with

native audio.

⭐ How to Use

Base Model: MiniMax H3-Base — 768p, 24 fps

guidance_scale: 1 — H3 is guidance-distilled. Side effect: the negative prompt does

nothing, so anything unwanted has to be handled with positive wording instead

Frame count must be 17n+5 (5, 22, 39 ... 107, 149, 209, 345). Resolution in multiples of 32

Best with natural scenes: street clips, car interiors at night, kitchens, courtyards,

casual portraits, action from a phone in someone's hand

Drop to 0.7 only if speech goes out of sync — softness blurs the mouth. Don't rewrite the

prompt to fix sync

🧪 Prompting H3 (different from LTX and Krea)

H3's text encoder is a language model (Qwen3-VL), not CLIP. That changes how you write:

– (text:1.3) weighting does nothing — the encoder reads brackets as characters

– tag tails (24yo, car selfie, night, candid) don't work either — write connected prose

– structure that works: scene → beats with timecodes → how the camera behaves → audio,

spelled out → no text or logos

– write the audio explicitly or the model will put music under the clip

– dialogue works, in Russian and the other languages H3 supports. Keep the prompt body in

English and swap only the spoken line

– say nothing about image quality. No "low quality", no "grain", no "compression" — that's

the LoRA's job. Describe light as a fact instead (one bare bulb overhead, a street lamp

passing across a face, sun straight into the lens) and let the sensor react

🧪 Notes

– Trained with Ostris AI-Toolkit, on RTX Pro 6000 Blackwell (96GB)

— about an hour of training

– Trigger-less — the style applies on its own

– Picture only. H3's audio branch is untouched by the LoRA, so voice artefacts belong to the

base model

📝 Example prompt

Close-up selfie video of a beautiful blonde woman in her early twenties in the passenger

seat of a car at night, her face filling most of the frame. Long light blonde hair, blue

eyes, soft natural makeup, a thin gold chain at her throat. The only light on her face is

the phone screen from below and orange street lamps sliding past the window behind her as

the car moves.

0:00-0:02 - She looks out through the side window, then turns her eyes into the phone and

says, "We're almost there."

0:02-0:04 - A street lamp passes and washes orange across her face and the seat behind her.

She says, "Twenty minutes, maybe."

0:04-0:06 - She smiles slightly, looks down, then back up into the phone.

She is holding the phone herself, close to her face, and her hand moves with the car.

Audio: her voice close to the microphone, quiet and unperformed. Tyre noise on wet asphalt,

the low hum of the engine, a turn signal ticking twice. No music.

No text, no captions, no logos, no watermarks anywhere in the frame.

149 frames, 768p, strength 1.00, guidance_scale 1, empty negative.