MiniMax-H3 gets a 4-step LoRA: 1344x768 video with stereo audio from a 50-step model

Sponsored by Klovera

Two bad moods lead the issue today. The clips both put a sulk next to someone who has not noticed it: a cartoon girl fuming beside a cheerful dad in a shark shirt, and a woman bathing three baby dragons while a far bigger one stomps up the riverbank. The news opens with a 4-step version of MiniMax-H3 that keeps the stereo audio. The Secret is about lighting and location, and why a room that looks lived-in makes a product look owned, not placed. Trending is where we begin.

📈 Trending

Both of today's clips put a mood in the frame and a second character who ignores it. A cartoon girl fumes while her dad beams; three baby dragons get scrubbed while a much larger one wades up behind the tub. Neither needs a line of dialogue, because the contrast is the joke.

1. Four thank-yous from a cartoon girl who means none

▶ Watch on Instagram · @rudenko.aivideo

A 3D-animated girl with a long blonde fringe stands in a doorway, glaring, a curl of steam rising off her head. Beside her a huge bearded man in a shark t-shirt grins straight ahead, completely unbothered. The caption is 'Thank you' four times, each with a different face, so the clip is built on four versions of one line. It is from @rudenko.aivideo.

📈 1,631,845 views — 38× the account's median

Why it works:

• Put steam over her head. The cartoon tell reads before the face does, so the mood lands in the very first frame.
• Pair the fury with a grin. Two opposite faces in one frame is a joke that needs no dialogue to work at all.
• Repeat one line four times. The same words with a new face each time is a loop people watch to the end.

2. Seedance bath time for three baby dragons, plus an intruder

▶ Watch on Instagram · @mrbashr_ai

A red-haired woman in a green apron kneels at a wooden tub by a river, washing three baby dragons in suds while a soap bubble drifts between them. Behind her a full-grown dragon wades out of the water, wings raised, mid-stride, the whole scene lit by late sun through the pines. The caption calls it 'that annoying dragon' with an angry face. @mrbashr_ai made it with Seedance 2.5.

📈 109,588 views — 3.6× the account's median

Why it works:

• Light it with low sun. Warm backlight on wet scales and suds reads as filmed, which is what sells the fantasy.
• Keep the threat soft-focus. A blurred dragon behind the main action makes viewers wait for the moment she notices it.
• Make the fantasy a chore. Bathing a pet is a job everyone knows, so the impossible part lands without explaining anything.

Try this on your next idea: Put the angry one and the calm one in the same frame before you write a word of dialogue, and let the gap between their faces carry the first three seconds.

🤖 News & Updates

MiniMax-H3 gets a 4-step LoRA: 1344x768 video with stereo audio from a 50-step model

A set of LoRAs, small add-on weights, for MiniMax-H3, the 33B text-to-audio-video model, has landed. They turn the 50-step teacher into a 4-step generator that puts out 1344x768 video with native stereo audio, sampled at 124 frames at 24 fps with no classifier-free guidance. The inference code and a Diffusers pipeline example sit in a separate code repository. If you run open video models on your own hardware, this matters because four steps instead of fifty is the difference between waiting and iterating.

ElevenLabs hits a $22 billion valuation and ships Eleven v4 and v4 Turbo

The voice company behind Eleven v4 closed a $300 million employee tender offer that values it at $22 billion, double its Series D valuation from February. Its first round, in 2022, valued it at $9 million. Alongside the money it has released Eleven v4 and v4 Turbo, which it calls its fastest and most emotive text-to-speech models yet, with 3x credits on Creator+ until October 12. If you narrate ads with cloned voices, this matters because the vendor you depend on is well funded and still shipping models.

@Kling_ai explains how The Beat was made with Kling 4.0

Kling's own account has posted a breakdown of The Beat, a piece made with Kling 4.0. The post walks through the five things that carried it: shots that run 30 seconds, Omni Reference, multi-keyframe control, 10-bit HDR and stereo audio. It is a making-of from the model's own team, and the post is linked here. If you make serial content with a recurring cast, this matters because a 30-second shot with keyframe control is most of what a scene needs.

✨ Sponsored

🤫 The Daily Secret

Rooms that look lived in

Look at the dragon bath in today's clips and ask where the light is coming from. Low sun behind the pines. It hits the wet scales, the suds, the woman's hair and the big dragon's wing from the same side. That is most of the difference between a scene that looks shot and a scene that looks assembled. A viewer cannot name it, but they feel it in the first half second and decide whether to trust the next ten.

If you sell things, lived-in is not decoration. A product in a room with a history reads as something a person owns and uses. The same product in a clean grey void reads as a catalogue page, and people scroll past those. Give the room one light source with a reason to exist, a window or a lamp, and let everything else fall off from it. Then add the mess of use: a mug ring, a chair pushed back, a cable across the table. Those details say someone lives here, so the product belongs here.

See you inside.
— Lauren

Published by Internet Mastery