MiniMax H3 Singularity LoRA Review: Uncensored Fast Motion

The gallery holds published, safe-for-work examples from the H3 Singularity LoRA model page, shown as rendered. Our content-ladder test clips aren't shown; we report the results only.

A spin and a cut, still sharp. From one still of a swordswoman on a wet rooftop, the fine-tune produced a full spin, a cut through the rain and a whip-around camera, with the blade and the water spray still readable mid-motion.

A style LoRA from a photo. The VHS-tape style LoRA we use on the regular MiniMax H3 LoRA family, at strength 0.8, on a fast breakdance spin in a neon underpass. The footwork stays sharp, but the photo sets the look, so the tape texture barely shows.

Two people from two stills, one fight. Reference-to-video took a swordswoman from one still and a young man from another, staged a fast exchange on the same rooftop, and both stayed recognisably themselves through the movement.

Write the action in two or three physical beats, name the camera move for each, and end with an Audio: line for impacts and breath. Continue the picture rather than describing a new scene, and on reference-to-video give each input a job (Picture 1 is the singer, Picture 2 is the stage) before you write the beats.

1. A spin and a cut (image-to-video, 768p, 5 s). The exact prompt behind the rooftop example.

From this exact frame: she spins, whips the sword up and cuts through the rain in one motion as the camera whips around her, water spraying off the blade. Crisp fast motion, no blur. Audio: rain, steel whistling through the air, a sharp exhale.

Run this prompt on MiniMax H3 Singularity LoRA →

2. A style LoRA on top (image-to-video, 768p, 5 s). The exact prompt behind the underpass example, with the vh5tape VHS-tape style LoRA loaded at strength 0.8. From a photographic frame the tape look stays faint; use it for the motion.

vh5tape, worn VHS tape. From this exact frame: he launches into a fast breakdance spin on one hand, legs sweeping in a circle, then freezes in a pose. Low camera circling him. Audio: a heavy hip-hop beat echoing in the tunnel, sneakers squeaking.

Run this prompt on MiniMax H3 Singularity LoRA →

3. Two references, one scene (reference-to-video, 540p, 6 s). The exact prompt behind the singer example: a portrait names the face, an empty stage names the location.

Picture 1 is the singer, Picture 2 is the stage. She walks out to the microphone under the three spotlights, spins once and sings the first line as the camera orbits her. Audio: a band kicking in, the crowd cheering, her voice.

Run this prompt on MiniMax H3 Singularity LoRA →

4. Two characters, one fight (reference-to-video, 540p, 6 s). One still per character, no LoRA.

Picture 1 is the swordswoman, Picture 2 is the young man in the hoodie. They spar on a rain-wet rooftop at night: she swings, he ducks and spins away, then they both stop and laugh. Fast tracking camera, crisp motion. Audio: rain, footsteps splashing, laughter.

Run this prompt on MiniMax H3 Singularity LoRA →

We test in two configurations, because they answer different questions. With no LoRA loaded, on 27 September 2026 we ran our full content ladder against both endpoints: L1 Suggestive, L2 Mature themes, L3 Strong, L4 Intense and both L5 Extreme probe types (restraint/impact and gore), three independent runs each per endpoint. All 36 runs rendered; none was softened or refused. With an external adult LoRA loaded, on 26 September we ran one request per endpoint at the L4 Intense level: image-to-video and reference-to-video both rendered as written, with the referenced people and setting carried through correctly on reference-to-video.

The one thing to watch is the opening-frame behaviour, not the content level. On our first image-to-video run with the adult LoRA at full strength, the prompt described a scene the opening frame didn't show, and the render replaced the frame with its own scene instead of continuing it. A repeat run with the prompt written as a continuation of the frame, and the LoRA at a lower strength, held the frame as expected. If you need the clip to stay on your picture, write the prompt as what happens next in that frame, and start concept LoRAs around 0.6 to 0.8 rather than at full strength.

The page label is Uncensored: we add no content filter on our side, and neither the weights nor the LoRA slots are reviewed. A refused request is not billed.

Not tested yet: a ladder run with a LoRA loaded at every rung (26 September covered only the L4 Intense level); anime or illustrated stills as the opening frame.

Adults only, in every still, reference, LoRA and output. Never depict a real, identifiable person without their consent.

Singularity keeps MiniMax H3's architecture, so any LoRA trained for MiniMax H3 loads here unchanged, up to three per request with independent strengths from 0 to 4. To confirm the weights are really applied, we ran reference-to-video twice with the same seed and no LoRA, and got the same clip both times; adding one LoRA to the same request changed it. What a style LoRA doesn't do here is override a photo: the vh5tape VHS-tape LoRA at 0.8, and a 1939 cartoon LoRA at 1 and 2.5, left clips from a photographic frame or reference looking photographic. The regular MiniMax H3 LoRA family behaves the same way from a photo and shows those styles clearly only on text-to-video.

We also loaded an external adult LoRA trained for MiniMax H3 on both endpoints. At the L4 Intense level, image-to-video and reference-to-video both rendered as written. The one caveat: at full strength with a prompt describing a different scene than the opening frame, the concept LoRA pulled the render away from the frame on the first attempt, and holding the frame on a repeat run took a prompt written as a continuation plus a lower strength (0.6 to 0.8) rather than 1.0. That trade-off applies to any concept LoRA on this family, not only to adult ones.

There is no per-LoRA charge; you pay per second of output at the resolution you choose, whichever LoRAs you load.

Singularity is a community fusion of several MiniMax H3 checkpoints, deep fine-tuned and pruned by the AIGC-Singularity team and published on Hugging Face. Early testers back up the fine-tune's stated goals: one wrote, "You guys weren't kidding when you say 'The result is a clean, sharp, and highly dynamic video generation model.'" and another reported that a result "totally followed the prompt, the image doesn't feel greasy at all", echoing the fine-tune's de-oiled-skin claim.

Not every report is clean: one user found that the fine-tune "will cause the original reference image or video to change color, with its saturation reduced, turning it overly grayish and white" in some runs. If a render comes back desaturated, a lower LoRA strength or a second take is worth trying before you assume the prompt is at fault.

• Don't need a fine-tune or LoRAs? MiniMax H3 takes text, an opening frame or references, goes to 2k and costs about 70% less per second at 768p. Our H3 review.
• Have H3 LoRAs but a calmer shot? MiniMax H3 LoRA loads the same LoRA files, renders in roughly half the time, and costs about 25% less per second at the same resolution. Our LoRA-versus-LoRA comparison.
• No LoRA files of your own, just want uncensored loops fast? MiniMax H3 Spicy animates a still with no prompt required, from $0.038 per second at 480p. Our H3 Spicy review.
• Want two-stage LoRAs on a different base model? Wan 2.2 LoRA separates high-noise and low-noise LoRA slots and starts at $0.024 per second, but tops out at 720p.