MiniMax H3 Image LoRA Review: Uncensored Stills, Your Own LoRAs

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

A style LoRA, at a glance. A text-to-image still with a VHS-look LoRA at strength 1: a woman in a slip dress in a doorway at night, tracking lines and a pink-and-cyan colour cast across the frame, exactly the tape-damaged look the trigger word calls for.

A cinematographer's light, from words alone. A man in a hotel bathrobe at a window at dawn, a realism LoRA at 0.8 shaping the skin and the blue-to-warm light split across his face. The same still later became an opening frame for an H3 video clip.

Same face, new scene. One studio portrait went in as the single reference on the editing endpoint, with the realism LoRA loaded; the output kept her face and hairstyle and put her on a sunlit balcony in a white linen shirt.

A different medium, same subject. A photograph of a singer at a microphone, redrawn by a 1939 cartoon LoRA into a rubber-hose-style hand-drawn frame, sepia paper texture, the pose and dress unchanged.

On text-to-image, put a LoRA's trigger word first if it has one, then describe the subject, the setting and the light like you're lighting a film shot. On editing, open with one short line per reference picture, then describe only what should change; naming the face again tends to fight the reference instead of helping it.

1. Style LoRA, trigger word first (text-to-image, 1k, 3:2). The exact prompt behind the VHS-look example, LoRA strength 1.

vh5tape, a paused VHS frame: a woman in a slip dress leaning in a doorway at night, pink and cyan light on wet asphalt, tracking lines across the bottom of the frame.

Run this prompt on MiniMax H3 Image LoRA →

2. Light it like a film shot (text-to-image, 1k, 16:9). The exact prompt behind the hotel-window example, realism LoRA at 0.8.

Cinematic film still. A man in his thirties in a white hotel bathrobe stands at a floor-to-ceiling window at dawn, one hand on the glass, looking out over a waking city, soft blue light on his face and warm lamp light behind him. 35mm film look, shallow depth of field.

Run this prompt on MiniMax H3 Image LoRA →

3. Name the reference, then the change (image editing, 1k, 3:4). The exact prompt behind the balcony example; Picture 1 was a single portrait, realism LoRA at 0.8.

Picture 1 is the woman. Put her on a sunlit balcony in a white linen shirt, wind in her hair, late afternoon light. Keep her face and hairstyle.

Run this prompt on MiniMax H3 Image LoRA →

On 27 September we ran our content ladder against both endpoints with no LoRA loaded at all, three runs per probe type on each: L1 Suggestive, L2 Mature themes, L3 Strong, L4 Intense and both L5 Extreme probe types (restraint and impact, and gore). Every one of the thirty-six runs rendered as written, with nothing softened and nothing refused. On editing, the L5 gore edits drew the wound more quietly than text-to-image did, a cut through the armour with blood rather than a full scene, but still what the prompt asked for. The base model needs no adult LoRA to reach the top of the ladder.

On 26 September we loaded an adult identity LoRA and ran one request on each endpoint. Text-to-image rendered our L4 Intense probe as written. Image editing did the same, and the person in the result matched the one in the reference image. The page label is Uncensored: we add no content filter on our side, and nothing in the prompt, the LoRA weights or the output is reviewed.

A refused request is not billed; none of our runs across either round was refused.

Not tested yet: LoRA-loaded runs at 2k (all of ours were 1k); scenes with two women or two men; anime and illustrated styles; a LoRA at strengths above 1.

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

Both endpoints take up to three LoRA entries per request, each a direct HTTPS link to a .safetensors file trained for MiniMax H3, with a strength from 0 to 4. A weight trained for a different base model loads without an error and changes little to nothing, so the first run with any new LoRA is a check, not a final: put its trigger word first if it has one, render one 1k image, and look for an obvious effect.

A style LoRA at strength 1 produced clear VHS tape damage on a text-to-image still: tracking lines and a shifted colour cast the plain prompt never asked for.

A realism LoRA at 0.8 changed the light and skin on a written scene without taking over the composition, and the same LoRA at the same strength, paired with a single reference image on the editing endpoint, kept a person's face and hairstyle while every other detail of the scene changed.

A cartoon-restyle LoRA at strength 1, with its trigger word first, took a photograph on the editing endpoint and returned it as a hand-drawn 1939 cartoon frame, holding the pose, the dress and the prop in place while the entire medium changed.

Every one of these ran with a LoRA hosted outside SpicyAPI: point path at any reachable .safetensors link trained for MiniMax H3, including the ones people share on community LoRA hubs for this base model. Weights load with no review of what they contain or what they were trained to draw, which is also what let an adult LoRA load and render our L4 Intense probe unchanged on both endpoints, as described above.

MiniMax H3 Image LoRA runs on the same open MiniMax H3 weights as the video families, so what people say about that base model applies here too. The recurring theme is how little it holds back: "Least-censored model of any type I've ever seen from a major lab." and "Minimax has NO filters of any kind" are the two most-quoted lines. Redis creator antirez relayed a more mixed take, that "H3 is quite uncensored" but wasn't trained for the upper rungs, which describes anatomy risk rather than refusal.

The complaint most relevant to this family is about instruction-following: people running the open weights without MiniMax's own prompt rewriting say "I can write stuff and the prompt is completely ignored." In our runs, plain, specific prompts and clearly numbered references rendered as written; vague one-liners are the ones most likely to drift.

• Animating the same LoRA instead of a still? MiniMax H3 LoRA runs the same weights on video, from $0.025 a second, and takes the same character or style files.
• A cheap, LoRA-free uncensored still? Z-Image Spicy has no LoRA input at all but passed our content ladder for $0.01235 a still. Our Z-Image Spicy review.
• Open-source community LoRAs on other base models? Z-Image Turbo LoRA, FLUX.1 Dev LoRA and Qwen Image 2512 LoRA load the LoRA hubs built for each of those models, at their own prices and aspect ratios.