Skip to main content
NSFW LoRA guide

Best MiniMax H3 NSFW LoRAs — the packs that work and how to run them

MiniMax H3 is the fastest-growing base for NSFW LoRAs: 75 of the 300 most-downloaded LoRAs on Civitai over the past month support it. The files are small, the leading author mirrors them on Hugging Face, and we ran them on all seven H3 LoRA endpoints.

Get an API keyOpen MiniMax H3 Singularity LoRA
From
$0.05/ output second
Image to video · 480p
LoRAs compared
8
Tested on SpicyAPI
3

What we measured on MiniMax H3

  • 7 of 7 endpoints, 0 refusals

    HMNSFW AIO v2.5 at 1.0 on text-, image- and reference-to-video, both Singularity endpoints and both H3 image endpoints: every run reached L4. 26 Sep 2026.

  • 0.6 keeps your first frame

    At 1.0, with a prompt that set a different scene, image-to-video replaced the first frame. In the 0.6 run, with a prompt describing what happens next, the opening frame matched the reference, as in the no-LoRA baseline. 26 Sep 2026.

  • Style LoRAs: strong from text, faint from photos

    The VHS and 1939-cartoon LoRAs took over text-to-video and image edits completely, but barely showed when image-to-video started from a photo, even at 2.5. 27 Sep 2026.

Tests ran on each family's main route with the exact request shape the API sends — SpicyAPI passes LoRA links through unchanged. Outputs stay in our test archive and are not published.

The most-used MiniMax H3 NSFW LoRAs

Public download counts from Oct 10, 2026, the strength each author recommends, what we used, and whether the file runs here.

Tested
we ran this exact file
Sign-in file
loaded without a token in our tests
Should run
public link, right base, under 1 GB

The fastest-growing base: 75 of the 300 most-downloaded LoRAs on Civitai over the past month list MiniMax H3. Files are small (60–300 MB for most), and the leading NSFW author mirrors his packs on Hugging Face so they download without an account.

Best NSFW LoRAs for MiniMax H3
LoRATypeDownloadsStrengthFileLicenseOn SpicyAPI
HMNSFW AIO v2.5HearmemanAINSFW all-in-one66.3K1.9K likes0.5–0.9 author1.0 · 0.6 our test82 MBAny serviceTestedSep 26, 2026
Mystic (MMH3)alcaitiffNSFW all-in-one99.2K3K likes0.2–1 author148 MBAny serviceShould run
MiniMax H3 NSFW UnlockedThe_Midnight_LabNSFW all-in-one32.7K1.1K likes0.3–1 author284 MBCivitai onlyShould run
2D Anime Style NSFWtazmannner379Style17K811 likes—569 MBAny serviceShould run
H3 Motion BoosterFourBunnyBody motion44.5K1.7K likes0.6–0.8 author148 MBAny serviceSign-in file
Better Human MotionAdaptiveVisionBody motion14K619 likes0.4–0.8 author296 MBAny serviceSign-in file
vh5tape (worn VHS)KennethFalStyle5.5K59 likes0.7–1.0 author1.0 · 0.8 our test125 MBNot statedTestedSep 26, 2026
Studio 1939 Old Animationlovis93Style6.3K79 likes1.0 our test63 MBNot statedTestedSep 26, 2026

Counts come from the public Civitai and Hugging Face APIs on Oct 10, 2026; a Civitai count can combine several versions of one LoRA. LoRAs whose names spell out specific acts or body parts — most of Wan 2.2's top 50 — are left off by name, but they load the same way. Repositories built around real people, or anyone who looks underage, are excluded entirely.

Load an H3 LoRA

All three H3 LoRA families take one loras slot with up to three files. Some authors publish separate files for image-to-video (FL2VA) and reference-to-video (REF2VA) — use the one that matches your endpoint. The H3 image family renders single frames from the same weights, so video LoRAs work there too.

Hugging Face, Civitai and private link formats
Example request
POST /api/v1/jobs/createTask
Authorization: Bearer $SPICY_API_KEY
Content-Type: application/json

{
  "model": "minimax/h3-singularity-lora/image-to-video",
  "input": {
    "prompt": "<trigger word>, she turns from the window and walks toward the camera, warm tungsten light",
    "image_url": "https://example.com/first-frame.jpg",
    "loras": [
      {
        "path": "https://civitai.com/api/download/models/<versionId>",
        "scale": 0.8
      },
      {
        "path": "https://huggingface.co/<owner>/<repo>/resolve/main/<file>.safetensors",
        "scale": 0.6
      }
    ]
  }
}

Tips for MiniMax H3 NSFW LoRAs

  • Stay between 0.5 and 0.9

    HMNSFW's author moved from “0.5 or below” on V2 to 0.5–0.9 on V2.5; Mystic's author gives 0.2–1. On image-to-video, start at 0.7.

  • Stack small, not strong

    H3's leading NSFW author splits packs into adapters designed to stack, but users of the popular realism LoRA report deformations when it's paired with others. Keep the combined strength modest.

  • Use text-to-video when style is the point

    From a photo first frame, style LoRAs barely show on H3. Start from text, or restyle a still with the H3 image edit endpoint and animate that.

MiniMax H3 LoRA models on SpicyAPI

Guides and tests for these models

MiniMax H3 NSFW LoRA questions

NSFW LoRA guides for other models

All LoRA models, the full test log and how-to