People and realism
MiniMax H3 Realism People LoRA
Photoreal faces, skin that keeps its texture, light that behaves like a film set.
This is the most downloaded MiniMax H3 LoRA for human-centered shots. Open the LoRA generator, select Realism People, and type a prompt to begin.
It runs here at 480P or 768P with MiniMax H3's native synchronized audio. Signing up puts 20 credits on your account, enough for two 5-second 480P clips, no card needed.
What the MiniMax H3 Realism People LoRA Does
It fixes the things that make AI footage of people read as AI.
MiniMax H3 on its own is a capable general video model. Where it struggles is the same place most video models struggle: human faces at close range. Skin smooths into plastic, eyes lose coherence between frames, expressions land in the wrong place, and light falls flat.
This adapter is trained specifically on that problem. Skin keeps its texture instead of averaging out. Eyes and micro-expressions stay consistent across the clip. Lighting behaves the way it does on a film set rather than in a render. Camera motion picks up a subtle handheld quality, which is what makes footage read as documentary rather than generated.
It works on close-up talking faces, arguments, several people speaking at once, weathered skin, children, and crowd scenes. Hands at work hold together better than the base model manages.
The adapter keeps MiniMax H3's native synchronized audio—you still get dialogue, room tone, and effects in the same pass.
It replaces an earlier realism adapter, retrained on a larger dataset focused on people.
Before and after
Before and After, Same Prompt, Same Seed
The only variable is the adapter.
A style claim is easy to make and hard to check. The comparison keeps the prompt and seed aligned so the adapter is the variable to watch. The trigger word is present in both prompts, so the trigger itself is not doing the work.
Watch the skin first. On the base model it tends to even out into a single tone; with the adapter it keeps pores, shadow, and irregularity. Then watch the eyes across the length of the clip.
How to Use the MiniMax H3 Realism People LoRA Here
Three settings decide almost everything.
The trigger word is not optional
Start the prompt with r34l1sm. The adapter is trained against that token and has little effect without it. If you clear the prompt, put the trigger back before you render.
r34l1smStrength: start at 1.0
A strength of 1.0 gives the intended realism effect. Move down to 0.6–0.8 when you want a lighter realism pass over a stylized scene or when the adapter is competing with another strong direction. Going above 1.0 is rarely useful.
Describe the light and the lens
Say what kind of light is in the room and how the camera behaves. Soft window light, handheld follow, shallow depth of field, restrained contrast, and a clear lens choice give the adapter useful visual direction.
Use any of the three input modes
Text to video, image to video, and reference to video all accept this adapter. A familiar photograph is the fastest test because you already know how the original face and skin should look.
What It Was Trained On
The adapter's published information describes its lineage but not its dataset composition. What is documented: it succeeds an earlier adapter and was retrained on a larger dataset focused on people.
Beyond that, the training data has not been disclosed publicly. We would rather say so than describe a dataset we have not seen.
License and Attribution
Who made it, and what you can do with what you generate.
This adapter was created and published by fal. The weights themselves remain the adapter author's work.
Licensing follows the base model: the MiniMax H3 Community License, the same agreement that governs MiniMax H3 itself. That license travels with the adapter rather than being replaced by it.
Videos you generate on a paid pack carry commercial use rights from this service, subject to that license. If you are planning a campaign around output from this adapter, read the license text rather than relying on this summary.
MiniMax H3 Realism People LoRA FAQ
The questions people ask before using the Realism People LoRA.
What is the MiniMax H3 Realism People LoRA?
It is an adapter for the MiniMax H3 video model, published by fal, trained to improve human-centered shots: skin texture, facial coherence in close-up, believable expressions, film-style lighting, and handheld camera motion. It keeps MiniMax H3’s native synchronized audio.
What is the trigger word?
r34l1sm. Start your prompt with it. The adapter is trained against that token and does very little without it—a missing trigger is the most common reason a first attempt looks unchanged. The generator inserts it automatically when this LoRA is selected.
What strength should I use?
Start at 1.0. Use 0.6–0.8 when you want a lighter realism influence and more of the base model’s original look. Values above 1.0 usually compete with the rest of the prompt instead of improving it.
Does it output 2K?
The LoRA generator currently offers 480P and 768P output. If you need 2K, use the base MiniMax H3 generator without a LoRA. Open the base MiniMax H3 generator.
What does a clip cost?
LoRA output costs 2 credits per second at 480P and 3 credits per second at 768P. A 5-second test costs 10 credits at 480P or 15 credits at 768P. New accounts receive 20 credits, and failed generations are refunded.
Can I use the output commercially?
Videos generated on a paid credit pack include commercial use rights from this service, subject to the MiniMax H3 Community License and any other terms that apply to the model and your use case. Review the license before using output in a paid campaign.
Try It on Your Own Shot
The fastest way to judge this adapter is to feed it a photograph you already know and watch what it does to the face. The trigger word is inserted when you select the LoRA; start with a strength of 1.0. Your 20 free credits cover two 5-second 480P clips.