Skip to content

Presets#

A preset bundles a model pack with resolution, frame count, sampler settings and the LoRAs to fuse. Pick it per shoot. Two families ship with the starter set.

LTX (image-to-video, audio and video)#

Fast, starts from your reference image, generates picture and sound together.

  • Best for: establishing shots, motion from a still, short clips with ambient sound.
  • Clip length: up to about 10 seconds. Longer clips lose the audio envelope and the sound falls apart before the picture does.
  • The upsample variants render at a lower rate and stretch the frames: the result plays in slow motion. Use them only when that is the intent.
  • The reference image is the first frame, so the aspect-ratio rule in Reference images and voices is strict.

MiniMax H3 (reference-to-video, speech)#

Builds the shot from the prompt with a face reference and, optionally, a voice reference. Lip-sync and dialogue are its strength.

  • Best for: dialogue scenes, characters that must look the same across shots, any shot with spoken lines.
  • Frame counts follow a 17n + 5 rule (for example 124, 141, 311, 345, 362). The preset list only offers valid counts. Longer takes are supported up to 362 frames (15 s; the point ComfyUI's "15 seconds" rounds to); very long takes take tens of minutes on a 16 GB card.
  • Resolution: the 768-square presets are the sweet spot for identity and speed. Larger frames cost more than they give.
  • Character LoRAs are fused at strength 0.5 by default. Strength 1.0 destroys motion and lighting; do not raise it.
  • LoRAs added to a shoot are not the way to use a character on H3: a character set preset ignores them and the stock preset fuses them at load time on every job. See LoRAs on a shoot.
  • Checkpoints: the official full files and the Comfy-Org pruned layout (minimax_h3_*_pruned_bf16) both load. Pruned files are 40 GB instead of 66 GB and the character-set packs built from them are about 21 GB instead of 34 GB. Without the pruned download, Models → Video → Install → pruned base by local fold folds the installed full hybrid into the pruned layout on your machine (about 2 minutes, 40 GB). When a pruned hybrid is installed, character sets for reference shots are built on it by default. Community merges published in the pruned layout work the same way through a custom preset.
  • Turbo: the lightx2v Minimax-h3-Turbo LoRAs cut a shot to 4 or 8 model calls at the same per-step cost (a 124-frame reference shot: about 3.5 min instead of 5.5 on a 16 GB card). Fuse the LoRA into a pack at strength 1.0 (Models → Video → character packs, or the pack API), then derive a custom preset on that pack and fill its sampler defaults: steps 4 for the 4-step LoRA (8 for the 8-step one), and for the 768p LoRAs the schedule shifts video 6 / audio 3 (the 544p v0.1 LoRA keeps the model's 12 / 3). Shots on that preset use those values unless the shoot sets its own step count (or its scene a default count); the shoot's steps field names the count that will run and where it comes from, so a shoot value never overrides the preset's count silently. Merges with Turbo already baked in (the 8-step TURBO builds) keep the model's 12 / 3 too; a lower video shift on them washes out detail and prompt following. The shift fields show the model's own values as placeholders and warn when a value departs from them.
  • Steps count model calls, as ComfyUI and diffusers count them: 8 runs the model 8 times. (Runtimes before 0.1.29 counted the grid points, so 8 ran 7 calls and the Turbo guidance read 5 / 9; if you entered 5 or 9 following it, change them to 4 / 8 — a value of 8 or 10 was always meant as calls and now runs as such, one call longer than before.)
  • Rebuilding a ComfyUI recipe: the same checkpoint, text encoder and VAE files load through a custom preset (the Comfy-Org int8_convrot encoder and VAE included). ComfyUI's steps 8 is 8 here: its simple scheduler picks the same grid, and its res_multistep sampler is available as the custom preset's Sampler default (Euler is the model's own step). A seed maps to a different noise sample on each side by default, and that sample decides how a shot is staged (where it cuts, how the camera moves); set the custom preset's Noise source to ComfyUI and a seed lands on the same scene as in ComfyUI with the same files (frames still differ a little; the encoders are not bit-identical). Verified at a 384×384 test canvas; at 768×768 the same seed gives a prompt-faithful variant of the scene (small staging differences, as if the seed had changed). That gap is in the transformer's own arithmetic, not in anything you can set: the noise, the layout and the encoder outputs match ComfyUI, so a 768p recipe lands on the same staging only up to such a variant. Leave it on Maystro to keep seeds you have already used in Maystro.
  • Attention kernel: by default a MiniMax H3 shot runs its first two model calls on exact SDPA attention and the rest on SageAttention. The scene's composition is decided in those first calls; on a small test canvas (384×384) SageAttention there tipped the scene into a badly rendered composition in our tests, while after them it only produces a variant of the same scene. The default keeps most of the speed: a 15-second 768×768 take finishes in about 10.7 minutes on a 16 GB card, against 9 with SageAttention on every step and 16 with SDPA on every step. SageAttention on every step is the fastest (measured safe at 512² and above in our tests); SDPA on every step is the exact-attention reference; SDPA first, SageAttention after runs the first half of the steps on SDPA, the conservative middle (about 12 minutes). The choice lives in the custom preset's Attention default, and every shoot can override it in its own panel (Attention kernel (this shoot)) to compare kernels take by take without deriving presets.
  • Reference canvas: by default every reference image is scaled down (never up) to the shot's own size, match, the fastest: a 15-second 768p shot steps about 13 % faster than at the released model's canvas, with the same shots and identity in our tests. In a custom preset's sampler defaults, Reference canvas picks native (a 2048 px short edge with small photos upscaled, the way the released model does it; about 6,000 extra tokens per reference, so steps take about 1.5x as long at 480p and 1.2x at 768p) or max (up to 2048 px, never upscaled). Try native on a test shot when a character's likeness drifts with a small reference photo.

Choosing#

You want Use
A still to come alive, ambient sound LTX
Someone to say a line, lips in sync MiniMax H3
The same person across a whole scene MiniMax H3 with a character set
A longer take MiniMax H3 (up to 362 frames, 15.1 s)
A take longer than 362 frames MiniMax H3 chain clips; on a reference shoot the references condition clip 1 only

Custom presets (your own LoRA, a different resolution) are built in Models and presets.