MiniMax H3 Transformation LoRA: Morph Between Two Images

Tr@nsf0rmation_style (Hugging Face) is a community LoRA for MiniMax H3 that morphs one picture into another. Given a start frame and a target frame it deconstructs the first scene, element by element, and reassembles the pieces into the second, so subjects dissolve into one another, materials change as they travel, and outlines redraw themselves along the way. Ashmotv published the weights on September 25 with two checkpoints, ten sample clips and the prompt format the adapter expects.

A frame from a multi-element scene transformation generated with the MiniMax H3 transformation LoRA

A frame from one of the showcase clips, where each subject in the first picture is assigned its own counterpart in the second.

What it does

MiniMax H3's first-and-last-frame path normally reads two images as the ends of one continuous shot: the camera or the subject moves, and the scene stays recognisably the same. This adapter makes the two frames a beginning and an end state instead. Everything between them is the transition, and the model is trained to carry individual visual elements across it rather than crossfade the whole image.

Because the correspondence is spelled out in the prompt, one move can carry several transformations at once. In the sample below the person in the grey t-shirt becomes the girl in the second picture, the person in the yellow shirt becomes the yellow grass, and the person in the green shirt becomes empty background, all inside a single clip.

The prompt format

The adapter uses a trigger word plus an action connector, and works best when the prompt names how each subject, material or element turns into its counterpart:

Trigger word: Tr@nsf0rmation_style
Action connector: Tr@nsf0rmation into

The recommended template is:

integrated_multimodal_description: [Shot 1] Tr@nsf0rmation_style, Picture 1 Tr@nsf0rmation into Picture 2 with individual elements and subjects Tr@nsf0rmation into one another.

A more specific example, from the model card:

integrated_multimodal_description: [Shot 1] Tr@nsf0rmation_style, Picture 1 Tr@nsf0rmation into Picture 2 with individual elements and subjects Tr@nsf0rmation into one another. The bald person in grey T-shirt and blue jeans from <Picture 1> Tr@nsf0rmation into the Girl in <Picture 2>. The person in the Yellow shirt from <Picture 1> Tr@nsf0rmation into the yellow grass from <Picture 2>. The person in the green shirt from <Picture 1> Tr@nsf0rmation into the blank background in <Picture 2>.

Samples

Multi-element scene reassembly: each person in the first picture is morphed into a different element of the second.

A straight visual transformation between two scenes.

Pencil stroke evolution, using the more specific prompt variant where individual pencil strokes transform into one another.

Seven more clips, covering fluid transitions, semantic morphing, complex subject swaps and high-fidelity flow, are on the model card.

Recommended settings

setting	value
base weights	MiniMax H3 FL2VA or Ref2V, such as `minimax_h3_fl2va_pruned_int8_convrot.safetensors` from Comfy-Org/MiniMax-H3
checkpoints	`Transformation_style_c1-st2500` is the fully converged one, `-st1600` is softer and lighter
LoRA strength (st2500)	1.0 to 1.2
LoRA strength (st1600)	1.0 to 1.4
stacked setup	st2500 at 1.2 plus st1600 at 1.4, used for the showcase clips
sampler / scheduler	`res_multistep` or `euler` with `simple`; ResMultistep gives the most stable transitions
steps	12 with a DMD turbo adapter, or 25 to 30 standard
sigma shift	`shift_video` 8.0 and `shift_audio` 3.0
resolution	852 × 480 or 1024 × 576, upscalable to 1080p

Running it in ComfyUI

The files are distributed in ComfyUI naming (Transformation_style_*_comfyui.safetensors), so no custom nodes are required:

Download the checkpoint you want into ComfyUI/models/loras/.
Add a LoraLoaderModelOnly node after your MiniMax H3 UNet loader and set strength_model between 1.0 and 1.2.
Feed the start picture and the target picture into the first-and-last-frame conditioning nodes.
Write the prompt with the Tr@nsf0rmation_style trigger and the Tr@nsf0rmation into connector, as shown above.

Availability

LoRA weights: Ashmotv/transformation_lora
Base model repack: Comfy-Org/MiniMax-H3

Download Transformation_style_c1-st2500_comfyui.safetensors (36 MB)