MiniMax H3 Max Recast
MiniMax's video-to-video recast model — swap the people in an existing clip using reference photos while keeping the original motion, camera work, and cuts.
curl https://api.venice.ai/api/v1/video/queue \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3-max-recast",
"prompt": "Aerial drone shot over a misty mountain valley at golden hour"
}'
# Use the returned queue_id with https://api.venice.ai/api/v1/video/retrieve.
# Call /video/complete after downloading if needed.Overview
What is MiniMax H3 Max Recast
MiniMax H3 Max Recast is a video-to-video editing model from MiniMax that replaces the people in an existing clip with new identities drawn from reference photos — up to four people, one photo each — while preserving the original motion, camera work, and cuts. It outputs 768P or 1080P video.
Running it privately on Venice
On Venice, MiniMax H3 Max Recast runs under the private tier with zero retention — your source footage, reference photos, and prompts are never stored, profiled, or used for training. The model is also uncensored, so legitimate recast work that other platforms gate behind content filters runs without friction. You pay per use with credits instead of a subscription, and no generation history ties the work back to your identity.
Specifications
Datasheet
- Maker
- MiniMax
- Modality
- Video-to-video (identity recast from reference photos)
- License
- Proprietary (Recast variant); H3 base under MiniMax H3 Community License
- Clip lengths
- Auto
- Resolutions
- 768P, 1080P
- Mode
- video
- Audio
- No
- Prompt limit
- 2,000 chars
- Input images
- Up to 4 reference photos, one per person; accepted formats and size limits not published
- References
- Up to 4 reference images (one photo per person) plus the source video to be recast
- Reference video
- The source clip is the primary input; no published maximum duration
- Released
- October 2026 on Venice; the H3 base model was announced July 31, 2026
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Oct 2026
Assessment
Strengths and limitations
- Surgical identity replacement: recast up to four people in an existing clip from a single photo per person, applied left to right.
- Editing-grade control: the original motion, camera moves, and cuts are preserved rather than re-generated, so the edit feels like the same shoot.
- Up to 1080P output, with 768P available for faster, cheaper drafts.
- Uncensored on Venice: no content-filter gatekeeping on legitimate recast and localization work.
- Private by default: zero retention means your footage and reference photos are not stored or used for training.
- Requires a source video: this is not a text-to-video model, so it cannot invent a scene from a prompt alone.
- No audio track on Venice output; if you need the original sound carried through, plan to re-sync it in your editor.
- Clip length is chosen automatically rather than set by the user.
- The Recast variant is closed: only the H3 base model has published weights, and its community license excludes several territories including the US, EU, and UK.
Use cases
What it is good for
- 01Ad localization — swap on-screen talent so one shoot serves multiple markets without reshooting.
- 02Brand and influencer content — put a licensed character or spokesperson into existing footage.
- 03Film and commercial pre-visualization — recast scratch actors as final casting choices before committing to a shoot.
- 04E-commerce and product video — reuse a proven performance while changing who delivers it.
- 05Privacy-safe editing — replace identifiable bystanders in footage you have the rights to edit.
Prompting
Getting better results
Upload one clear, front-facing photo per person, ordered left to right to match how they appear in the video.
Never put two faces in a single reference image — the model maps one photo to one person.
Spend your 2,000-character prompt on the look change: wardrobe, hair, age, build, and lighting on the new identity.
State explicitly what must not change — camera moves, cuts, background, and framing — to protect the original edit.
Draft at 768P to iterate on the recast cheaply, then re-run the winning prompt at 1080P for delivery.
Choose reference photos shot in lighting similar to the source scene; mismatched lighting is the most common cause of a pasted-on look.
Alternatives
How it compares
| Model | Best for | Max duration | Max resolution | Native audio |
|---|---|---|---|---|
| MiniMax H3 Max Recast | Recasting people in existing footage | Auto | 1080P | No |
| MiniMax H3 Max | Text & image-to-video with sound | 15s | 1080P | Yes |
| Grok Imagine 1.5 | Fast generative clips with audio | 15s | 1080p | Yes |
| Kling O3 Pro | Cinematic generated scenes | 15s | — | Yes |
| Wan 2.7 Enhanced | Open-weights generation | 15s | 1080p | Yes |
MiniMax H3 Max Recast is the right pick when the footage already exists and only the people in it need to change — localization, casting swaps, and character work. For generating entirely new scenes from a prompt, use MiniMax H3 Max or Grok Imagine 1.5 instead.
API
Call it from your code
Venice exposes this model through the REST API. Queue a generation with the model id.
curl https://api.venice.ai/api/v1/video/queue \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3-max-recast",
"prompt": "Aerial drone shot over a misty mountain valley at golden hour"
}'
# Use the returned queue_id with https://api.venice.ai/api/v1/video/retrieve.
# Call /video/complete after downloading if needed.Pricing
What it costs on Venice
Pricing available on Venice.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
FAQ
Frequently asked questions
MiniMax H3 Max Recast is a video-to-video editing model from MiniMax, built on the H3 video model family. It replaces the people in an existing clip with new identities from reference photos — up to four people, one photo each — while keeping the original motion, camera work, and cuts, and outputs 768P or 1080P video.
On Venice, MiniMax H3 Max Recast is billed per use with credits — see the pricing page for the current per-second rate. Third-party hosts bill by output duration; for example, fal charges $0.30 per second at 768p and $0.45 per second at 1080p, which gives a rough sense of market pricing.
It is neither free-software nor open-weights: the Recast variant is closed and only available through hosted services like Venice. The underlying MiniMax H3 base model was released on Hugging Face under the MiniMax H3 Community License, but that license excludes territories including the US, EU, UK, and South Korea, and it does not cover the Recast variant itself.
They solve opposite problems. H3 Max Recast edits footage you already have, swapping the people while preserving motion, camera, and cuts. H3 Max generates new video from text or image prompts with native stereo audio. If you have a finished shot and need to change who is in it, use Recast; if you are creating a shot from scratch, use H3 Max.
Two output resolutions: 768P for faster, cheaper drafts and 1080P for final delivery. Clip length is chosen automatically by the model rather than set manually, and prompts can be up to 2,000 characters.
On Venice, output from MiniMax H3 Max Recast has no audio track. Some third-party listings describe the underlying recast technique as preserving the source clip's sound, so plan to re-sync your original audio or a replacement track in your editor after downloading the video.
Up to four reference photos, one per person, matched left to right against the people in the source video. Each photo should contain a single face; the model maps one image to one person. There is no published limit on the source video's duration.
Venice runs MiniMax H3 Max Recast under its private tier with zero retention: your source video, reference photos, and prompts are processed but never stored, profiled, or used for training. You can use it in the Venice app or via the Venice API without building a personal generation history.
Run MiniMax H3 Max Recast privately
No prompt logging. No data used for training.