MiniMax H3 LoRA and H3 Max Recast solve different consistency problems. A LoRA keeps a trained style, character or motion across new videos. Recast keeps an existing video and replaces the people in it with the people in your reference photos. Both are available on Raywake as separate operations with their own inputs and prices.
This guide uses the request contracts available on Raywake on 23 October 2026. It describes inputs and limits, not output quality.
H3 with a LoRA: three entry points
The MiniMax H3 LoRA page offers three operations that share the same LoRA input.
| Operation | Starts from | Extra inputs |
|---|---|---|
| Text to video LoRA | A prompt | Aspect ratio 21:9 to 9:16, optional audio track |
| Image to video LoRA | An optional first and/or last frame | Optional audio track; the canvas follows the frame you supply |
| Reference to video LoRA | Reference files named in the prompt | Up to 9 images, 3 videos and 3 audio clips, 12 files in total |
Every operation requires a prompt and at least one LoRA. You can apply up to three LoRAs in one request. Each entry needs a path, either a direct file URL such as a .safetensors link or a Hugging Face repository ID in owner/repo form, and an optional strength from 0 to 4 (default 1). When a repository holds several weight files, name the one to use with weight_name.
Raywake does not offer H3 LoRA training. Bring weights you have trained elsewhere or are licensed to use, and keep the original training subject in mind: a character LoRA works best when the prompt describes that same character.
In reference mode, refer to files by their type and order: Image 1, Image 2, Video 1, Audio 1. Reference videos and audio clips must each be 2 to 15 seconds, and each list must total at most 15 seconds.
Sources: text to video LoRA, image to video LoRA, reference to video LoRA.
Resolution, duration and price
Choose a duration from 5 to 15 seconds. 480p and 768p are native generation modes. 2K and 4K upscale a 768p base result, so they add pixels to the same composition rather than generating more detail from scratch.
| H3 LoRA resolution | 5 seconds | 10 seconds |
|---|---|---|
| 480p | up to $0.55 | up to $1.09 |
| 768p | up to $0.66 | up to $1.31 |
| 2K | up to $1.42 | up to $2.84 |
| 4K | up to $1.75 | up to $3.49 |
For drafts, test the LoRA strength and prompt at 480p or 768p, then render the approved version at the size you need. Prompt expansion is on in balanced mode by default; set it to disabled when the LoRA trigger words must reach the model unchanged.
H3 Max Recast: same video, new people
H3 Max Recast takes a source video and one to four reference photos, one photo per new person. It keeps the source motion, camera, cuts and audio. By default the people in your photos replace the main people in the video from left to right in every shot where they appear. Use the optional prompt to say who becomes whom, or what else to keep.
The source video must be 5 to 30 seconds long, with no single shot longer than 15 seconds. Upload it rather than linking it, so its length can be measured and priced before you generate.
| Recast resolution | 5-second source | 10-second source |
|---|---|---|
| 768p | up to $2.62 | up to $5.23 |
| 1080p | up to $3.93 | up to $7.85 |
Recast is billed per second of the source video, and reference photos are included. Source: H3 Max Recast.
Only recast people with their permission and only with footage you have the right to edit. Deceptive depictions of real people are not allowed under the acceptable use policy.
Which one fits your brief?
Use a LoRA when you are producing new scenes and need the same character, product style or motion signature each time. Use Recast when the scene is already right, for example an approved performance, edit or camera move, and you need a different cast in it.
They can work in sequence: generate a scene with an H3 LoRA, approve its motion and timing, then recast the people for a regional version. For reference controls without a LoRA, see the H3 and H3 Max reference guide.
Open MiniMax H3 in the studio