Skip to content

Presets and LoRAs#

A preset is a recipe: a base model plus the encoders, VAE, adapters and LoRAs it needs, with tuned defaults. Image Studio applies a model's default settings when you pick it; Video Studio picks everything through presets. This page covers both, plus LoRAs.

Model defaults in Image Studio#

Every main model carries default settings in the model manager: steps, CFG or guidance, scheduler, size, VAE. Selecting the model on the Generate tab applies them; edit them in the model's panel on the Models tab so your preferred values load every time. Catalogue models ship with sensible defaults; imported models start empty and use the family defaults.

Video model packs#

The video Models page (Video Studio, left rail) lists the Vireo presets: for each one, the slots it needs (transformer, text encoder, VAE, LoRAs, upscalers) with a per-slot file report: found, missing, or changing. A missing slot shows the expected locations, in root search order, and a copy button for the path; catalogued components have an install button. A slot may list alternatives: the Ref2VA preset accepts its int8 pack, the merged bf16 hybrid or the pruned hybrid, whichever is on the machine, and the row names the one that matched. The starter models dialog installs the whole LTX set and the MiniMax H3 reference-to-video pack (two official sources are downloaded and merged on your machine). Either H3 bundle also installs what every H3 preset shares when it is missing: the Qwen3-VL text encoder (downloaded as 14 shards and converted to a 34 GB int8 pack on your machine, one shard at a time), its tokenizer files and the video and audio VAEs. The same pack is also offered from Comfy-Org's pruned sources: about 60 % of the download and a 40 GB hybrid instead of 62 GB. Its output is a close variant of the full bundle's hybrid, not a pixel match: the adaLN timestep network is folded into a table, the two official pruned files carry different tables, and the hybrid's donor blocks keep the Ref2VA timestep response (same scene, framing and motion in the side-by-side of 2026-10-17). Install one of the two Ref2VA bundles, not both; character sets and int8 packs build from whichever hybrid is installed. Copy a file into one of the listed locations and refresh the report from disk. Catalog downloads resume where they stopped after a network interruption or a cancel, and every file is checked against the publisher's size and checksum before it is used.

Custom presets#

Change a slot and save the result under a new name to derive a custom preset. The base preset is untouched; the custom one appears wherever presets are chosen and sends its slot assignments to the runtime with each generation. Preview before saving: a slot that shows as missing is rejected before the job enters the queue, and so is a file that is found but of a format the slot does not load — the preview reads each MiniMax H3 file's header and names what it is (a pruned bf16 single file, a quanto pack, the Comfy-Org int8 ConvRot encoder, an nvfp4 or fp8 variant) and, when the slot cannot use it, what it reads instead. Deleting a custom preset never touches model files.

Custom presets are how you swap a checkpoint variant (FP8 for bf16), try a different upscaler, or bake a LoRA into a preset for a project. Character sets are a special case built for you; see Character sets.

Type a slot path relative to a models root, folder included (for example minimax_custom/my_h3.safetensors for a file kept in that folder under the root). A bare file name is looked up only in the preset's own family folder, and the preview lists every location it tried, so a "not found" slot tells you where the file is expected. MiniMax H3 accepts the Comfy-Org single-file bf16 checkpoints in both layouts: the full files and the smaller pruned ones (*_pruned_*, 40 GB instead of 66 GB), including community merges published in the pruned layout. A character set built from a pruned file stays pruned and streams about 40 % less per step.

The MiniMax H3 text encoder and video VAE slots take the Comfy-Org *_int8_convrot files as well (qwen3vl_32b_minimax_h3_int8_convrot and minimax_h3_video_vae_int8_convrot), so a ComfyUI recipe can be rebuilt here component for component. They run through the same int8 kernel ComfyUI uses (an NVIDIA card of the RTX 20 series or newer); the nvfp4 and fp8 variants of those files are not accepted, and neither is the Comfy-Org int8 ConvRot transformer file — the transformer slot takes the bf16 single files and the packs built from them. The text encoder then runs with int8 activations like ComfyUI (Maystro's own quanto file keeps them bf16), so a shot's details move at the same seed. Without the overrides the stock files do the equivalent work: the video VAE decoder is quantised to the same int8 form when it loads.

LoRAs#

A LoRA is a small add-on that steers a base model towards a style, a subject or a person. It only loads with the base model it was trained for.

  • Images: add LoRAs in the Generate panel and set a weight each. Style LoRAs sit around 0.6 to 0.8; identity LoRAs often want less. A LoRA panel shows its trigger phrases, which go into the prompt.
  • Video: shoot and scene LoRA lists take video LoRAs from the video LoRA library; character LoRAs from Training Studio land there automatically and default to strength 0.5, which is proven for MiniMax H3. Strength 1.0 breaks motion and lighting.
  • Edit tab: the Identity Edit LoRA for Krea 2 is required and chained automatically; see Edit.

Install image LoRAs from the Models area (Image tab) like any model; trigger phrases can be recorded in the model panel so the prompt box suggests them.