@sogni-ai/sogni-creative-agent-skill
Version:
Sogni Creative Agent Skill: agent skill and CLI for Sogni AI image, 3D, speech, video, and music generation.
119 lines (94 loc) • 6.6 kB
Markdown
# MiniMax H3 video LoRA catalog
The adapters Sogni publishes for MiniMax H3 — the only video family that loads
LoRAs. Everything else (LTX, WAN, Seedance, HappyHorse) ignores `--lora`, and
the CLI refuses the combination rather than rendering as if you had not asked.
Video LoRAs are a different shape from the Krea 2 image sliders:
- **Positive-only, except Action Speed.** For the fine-tunes a negative strength
is not an inverse effect, and `0` is off. `h3-action-speed-slider` is the one
bipolar H3 adapter: negative strengths slow the action, positive strengths
speed it up.
- **One file per family**, not one per effect.
- **Availability differs by H3 mode.** Read the live catalog for the exact mode
you are rendering; the mode's canonical model id is the filter.
## Selecting a compatible model
For MiniMax reference-to-video, prefer Balanced R2V (minimax-h3-r2v-balanced), including LoRA requests; it holds Mystic X v5 at full strength. Use minimax-h3-r2v-balanced-2stage for two-stage, 1080p or 2K output. Keep other R2V tiers only when explicitly requested.
LoRAs attach to the resolved H3 model, so name the mode in the same command:
| Mode | Selector | Canonical model id |
|------|----------|--------------------|
| Text-to-video | `-m minimax-h3-t2v` (`-turbo`) | `minimax-h3-fl2va-fp8_t2v` |
| Image-to-video | `-m minimax-h3-i2v` (`-turbo`) | `minimax-h3-fl2va-fp8_i2v` |
| First+last frame | `-m minimax-h3-flf2v` (`-turbo`) | `minimax-h3-fl2va-fp8_flf2v` |
| Balanced reference-to-video (default) | `-m minimax-h3-r2v-balanced` | `minimax-h3-ref2va-fp8_r2v_balanced` |
| Standard reference-to-video (explicit choice) | `-m minimax-h3-r2v` (`-turbo`) | `minimax-h3-ref2va-fp8_r2v` |
| FastH3 text-to-video | `-m minimax-h3-fasth3-t2v-turbo` | `minimax-h3-fastvideo-int8_t2v_turbo` |
| FastH3 image-to-video | `-m minimax-h3-fasth3-i2v-turbo` | `minimax-h3-fastvideo-int8_i2v_turbo` |
| FastH3 first+last frame | `-m minimax-h3-fasth3-flf2v-turbo` | `minimax-h3-fastvideo-int8_flf2v_turbo` |
| FastH3 Reference | `-m minimax-h3-fasth3-r2v-turbo` | `minimax-h3-fastvideo-omni-int8_r2v_turbo` |
Bare `-m minimax-h3` is not enough to validate a LoRA request: it resolves by
frame arguments, and availability differs per mode. Name the explicit mode.
The Balanced text, image and first+last frame selectors take the same adapters
as their Standard modes (as of 2026-10); their built-in LightX2V FL2VA or Larry
v4 Ref2VA acceleration dependency is not a custom LoRA slot. The live catalog
remains authoritative if that changes.
FastH3 keeps the FastVideo engine when an H3 LoRA is attached. Base jobs
require 23 GB, while jobs with an H3 LoRA require 32 GB. Always check the exact
FastVideo model id's live catalog before relying on a particular adapter.
FastH3 Reference (`-m minimax-h3-fasth3-r2v-turbo` and its `-2stage` form)
takes the same catalog adapters and Personal LoRAs as Reference-to-video.
The FastH3 audio-to-video modes (`minimax-h3-fasth3-ia2v-turbo`,
`-flfa2v-turbo`, `-a2v-turbo` and their `-2stage` forms) take the same catalog
and Personal H3 LoRAs as the frame modes (as of 2026-09-23): pass `--lora` with
`--ref-audio`, and the CLI resolves them against the selected audio mode's live
catalog exactly as it does for every other H3 selector.
## Published adapters
`h3-realism-people` (fal) — a realism pass trained on live-action footage of
people. Restores skin texture and pores, stray hairs, fabric weave, and a fine
sensor grain the base model smooths away; holds up in close-up. Range `0-2`,
catalog default `0.8`, usable band `0.6-1`. **It needs its trigger word**: put
`r34l1sm` near the FRONT of the prompt, or the render comes back as ordinary H3
with no error. It also pulls the camera in as it climbs — at `1.5` and above the
shot reliably recomposes and the grade darkens, which on an image-conditioned
mode can crop the subject out of the frame you supplied. Prefer the default when
you supplied a first or last frame.
`h3-vbvr-video-reasoning` — a prompt-adherence pass that holds the model to what
was asked instead of improvising. Range `0-1`, default `1`, usable band
`0.7-1`. No trigger word. Opt-in mature-capable; requires `--no-filter`.
`h3-mystic-xxx-v5` — an uncensored all-round adult adapter that keeps motion,
temporal consistency, and fine detail; stronger than v4 at the same strength.
Range `0-1`, catalog default `0.5`, recommended band `0.2-0.6`. No trigger
word. Requires `--no-filter`. `h3-mystic-xxx-v4` remains available with the
same range and default.
`h3-action-speed-slider` — changes how fast limb actions such as punches,
kicks and sword swings play. Range `-8` to `8`, catalog default `4`,
recommended band `-4` to `4`; negative strengths give slow motion and `0` is
off. No trigger word.
On Reference-to-video only Mystic X v5 is offered, on the author's own Ref2VA file (as of 2026-10); v4 and Action Speed are not.
Only `h3-realism-people` is published on every mode. Ref2VA carries it alone.
## Usage
```bash
# Single adapter at its catalog default strength
sogni-agent --video --ref shot.png -m minimax-h3-i2v \
--lora h3-realism-people "r34l1sm, a slow push-in as she looks up"
# Explicit strength, and a stack (order matters — adapters do not commute)
sogni-agent --video -m minimax-h3-t2v --no-filter \
--loras h3-realism-people,h3-vbvr-video-reasoning \
--lora-strengths 0.8,0.9 "r34l1sm, a welder lowers her mask"
# Slow motion with the bipolar Action Speed Slider
sogni-agent --video --ref shot.png -m minimax-h3-i2v \
--lora h3-action-speed-slider --lora-strength -4 "a sword swing through falling rain"
```
Omitting `--lora-strengths` applies each adapter's **catalog default**. Sending
a LoRA to the worker with no strength at all falls back to `1.0`, which for
`h3-realism-people` is already the top of its band — so the CLI fills the
catalog default in for you rather than letting that happen. Stack up to 8.
## Live catalog is the source of truth
This page is a snapshot. Sogni publishes and retires adapters without a skill
release, so read the catalog before relying on an exact range:
```bash
sogni-agent --list-loras --lora-catalog-model minimax-h3-fl2va-fp8_i2v
sogni-agent --json --list-loras --lora-catalog-model minimax-h3-ref2va-fp8_r2v
```
The CLI validates `--lora` against exactly this catalog for the model you
selected, so an id it accepts is an id the render will actually load, and a
wrong id fails before it costs a render. A LoRA the model cannot load is dropped
silently server-side — that check is what stops it reaching that point.