MiniMax-H3 VFX Edit LoRA

Experimental

This is a research LoRA, not a finished tool. Everything below reflects observed behavior: what it does well, where it works, and where it tends to break.

What It Is

This is an edit LoRA for MiniMax-H3 that takes your clip as a frame-aligned track rather than as a reference.

You provide:

• a source video;

• one sentence describing the change.

The goal is to return the same take with only that change applied.

Quick Start

Trigger Word

Every prompt must begin with:

vfx_edit:

Without it, you get the base model's behavior rather than the edit LoRA.

LoRA strength: start at 1.0.

Source Video

Put the source video into Add Guide, not into a reference slot.

The Add Guide node keeps your footage aligned frame by frame. A reference slot places the clip on its own timeline instead, which loses the synchronization that makes this model useful.

Important Settings

• Width / height: Match the clip's original aspect ratio.
  A vertical clip rendered on a horizontal canvas will be cropped and zoomed.

• Length: Follow the node's supported frame steps.
  124 frames ≈ 5 seconds at 24 fps.

• Resolution: Use the highest resolution your GPU allows.
  Detail lost at generation time cannot be recovered afterward.

• Reference image: Use an image at least as large as the intended output.
  References are shrunk to fit, never enlarged. Small references may effectively be ignored.

• VAE on Add Guide: Must be connected.
  Without it, the guide cannot be encoded.

Minimal Prompt

vfx_edit: Replace the living room background with a desolate, post-apocalyptic urban scene with burning debris and a dark, moonlit sky, while preserving the subject's pose, motion, framing and unrelated background elements.

How This Differs From Plain H3

H3 already accepts video. You can place a clip into its reference channels, write a prompt, and get something v2v-like out.

That workflow works, but it behaves differently.

In Plain H3, a Reference Is a Suggestion

A reference clip exists on its own timeline, separate from the generated timeline.

The model decides how much of it to use.

Sometimes it follows the clip closely. Other times it takes the general idea and effectively performs the scene again.

The harder your prompt pushes the model, the more likely it is to drift away from the original footage.

The first things to drift are often the details that nobody explicitly asks the prompt to preserve:

• Timing — a gesture happens earlier or later.

• Framing — a locked-off shot develops camera drift, or handheld motion gets smoothed.

• Sync — mouth movement stops matching the audio.

• Continuity — faces or clothing shift slightly throughout the clip.

What This LoRA Changes

This LoRA trains a different path.

The clip is pinned to the same grid as the generated video, so each frame of the source corresponds directly to the output frame being produced.

The source stops being merely a suggestion and becomes the structure the model generates on top of.

When It Works, This Usually Helps Preserve

• Timing and synchronization.
  The mouth remains in the same place at the same moment, so dialogue and music tend to stay aligned.

• Camera movement.
  The original movement—including handheld shake—is usually preserved rather than reinterpreted.

• Frame alignment.
  The generated result lands on the same frames as the source, making it easier to cut against untouched material or mask back over the original without resynchronizing.

• Pose.
  "Keep the pose" refers to the pose you actually filmed, rather than a similar reconstructed pose.

None of this is absolute.

Strong edits can still pull the result away from the source. Thin or fast-moving details can still break. Some clips simply do not cooperate.

The claim here is a change in probability, not a guarantee.

The model also inherits the limitations of your footage:

• it will not fix a bad take;

• it will not invent a camera move that was never filmed.

Shorter, Simpler Prompts

One practical advantage of this workflow is that prompts can be much shorter.

Prompting the base model for this kind of edit often requires long, structured descriptions covering:

• who is in the shot;

• what must remain unchanged;

• what happens moment by moment.

Those prompts may run for several hundred words because the text must describe everything the model otherwise does not know.

With this LoRA, the video already provides most of that information.

The prompt only needs to say:

1. what should change;

2. what must remain unchanged.

For example:

vfx_edit: Replace the living room background with a desolate, post-apocalyptic urban scene with burning debris and a dark, moonlit sky, while preserving the subject's pose, motion, framing and unrelated background elements.

That is the entire instruction.

Training prompts range from approximately 16 to 68 words, with a median of about 38 words.

Longer prompts do not necessarily improve the result. They can make it worse because they begin re-describing the source footage—and every re-description creates another opportunity to contradict what is actually on screen.

The Cost: Long Videos Are Expensive

Keeping the original footage in the generation loop increases compute requirements.

The model is effectively handling two videos at once:

1. the video it is generating;

2. the source video you supplied.

As clip duration increases, processing cost rises faster than you might expect.

Doubling the duration can require more than twice the processing time, and memory requirements rise similarly.

In practice, VRAM is often the first limitation.

A clip that runs comfortably at four seconds may simply fail at eight seconds.

There is also no inexpensive preview for a long take: shortening the input necessarily shortens the edit.

Recommended Workflow

Feed the model shots, not sequences.

For longer material:

1. cut the source at its existing edit points;

2. process each shot separately;

3. reassemble the edited shots afterward.

This approach is:

• faster;

• easier on VRAM;

• generally more reliable;

• often higher quality.

What It Tends to Do Well

Location Changes

Changing the location is one of the most reliable use cases.

Examples:

• same person, different room;

• same subject, different street;

• arena;

• alien planet;

• entirely new environment.

Relighting and Color

Examples:

• new key light;

• different mood;

• flat interior → cinematic lighting.

Time, Weather, and Atmosphere

Examples:

• day → night;

• clear → storm;

• smoke;

• dust;

• embers;

• flares.

Object Replacement

The model can replace an object that the subject:

• holds;

• sits on;

• stands beside.

Subject-Attached Effects

It can add effects attached to:

• the subject;

• a body part;

• a specific region of the frame.

Copying a Look From One Still

If you manually edit a single frame from the clip, the model can often propagate that visual treatment across the shot.

When this works, it can provide extremely precise control.

Instead of describing a look with many adjectives, you art-direct a single image using the tool of your choice and give the model that image as a visual target.

This is also one of the modes whose performance varies the most from clip to clip.

What It Does Not Do Well

Face Swapping

This is not a face-swap tool.

If you provide a photo of a specific person:

• clothing and body characteristics may transfer well;

• the resulting face may look similar;

• the face will generally not be the exact same person.

Measured using face-recognition similarity, results reach roughly one-third of the similarity score observed between two photos of the same person.

Do not use this workflow when exact likeness matters.

Inventing Camera Moves

The model does not create new camera movement.

Camera movement comes from the source footage.

Long Clips

This model is not designed for long sequences.

Use short takes and cut longer material first.

Text and Logos

The model does not reliably preserve:

• printed text;

• engravings;

• logos;

• tiny distant faces.

Expect approximations.

Compositing

This is not a compositor.

It does not output:

• layers;

• alpha channels;

• roto masks.

It returns finished pixels.

How to Write the Prompt

Every prompt begins with:

vfx_edit:

After that, write one sentence with two parts.

1. Say What to Change

Start directly with the action.

Approximately 80% of the training prompts begin with:

• Replace

The remainder commonly begin with:

• Relight

• Add

• Transform

• Transition

• Enhance

• Adjust

Include everything associated with the intended change in the same sentence.

For example, if the change includes:

• debris;

• smoke;

• reflections;

mention all of those as part of the edit.

2. Say What to Keep

This is easy to omit, but it strongly affects whether the output remains usable.

Common preservation terms in the training prompts include:

• lighting;

• pose;

• background;

• motion;

• framing.

A typical closing clause looks like this:

..., while preserving the subject's pose, motion, framing, lighting, and unrelated background elements.

This clause tells the model where the edit stops.

Without it, unrelated elements may drift.

If something is particularly important, name it explicitly.

Examples:

• hands;

• the logo on a shirt;

• an orange strap;

• a reflection in a window.

Explicitly naming something makes it more likely to survive the edit.

Avoid Contradictions

Never ask the model to preserve something the edit itself must change.

For example, if replacing the background:

Bad

Replace the background ... while preserving the background.

Better

Replace the background ... while preserving all unrelated background elements.

A self-contradicting prompt often produces a weak or partially applied edit.

Writing It in Your Own Language

You do not necessarily need to write the final prompt yourself.

An auto-prompter can use:

• several frames from the clip;

• your request written naturally in any language.

A vision model can then convert that input into a concise English prompt in the format expected by the LoRA.

A good auto-prompter can also:

• resolve who "her" or "the guy on the left" refers to;

• add preservation instructions you forgot;

• phrase reference-image instructions correctly;

• keep the prompt within the length range the model handles well.

For many users, this can make the difference between the model feeling reliable and feeling random.

Using a Reference Image

You can provide one reference image.

It can serve one of two distinct purposes.

Mode 1 — An Object or Subject to Insert

The reference may show:

• an object;

• a garment;

• a head;

• another visual element you want inserted.

In this case:

1. name the referenced element in the sentence;

2. point to the image;

3. briefly describe only what is visibly present.

Do not invent adjectives that are not supported by the image. Extra descriptive language can pull the result away from the actual reference.

Example

vfx_edit: Replace the original chair with <Subject 1>, the red futuristic chair shown in <Picture 1>, while preserving the subject's motion and the room.

Mode 2 — A Finished-Look Reference

The reference can instead be an already-edited frame from the same clip.

In this mode, the rule is different:

Do not describe the image.

Simply tell the model to match it.

Example

vfx_edit: Replace the office background with an abstract glass and prism environment, add vibrant light flares, transforming the source video to match <Picture 1>, a guide frame of this clip with the edit already applied.

Describing a finished-look reference in detail is one of the most common ways to degrade the result.

The model may begin following your written description instead of the image itself.

Two Common Reference-Image Failures

1. Mentioning an Image That Was Never Provided

If no reference image exists, describe the intended result directly in words.

For example:

with the head of a different woman

Do not refer to <Picture 1> if there is no actual image.

Otherwise, the model is waiting for information it never receives.

2. Using an Undersized Reference

Reference images are shrunk to fit but never enlarged.

A thumbnail or tiny face crop therefore reaches the model with very little useful information.

It may appear as though the model ignored the reference entirely.

Use a reference that is at least as large as the intended generation resolution.

Starting From an Edited First Frame

Another option is to manually edit frame one of the clip and supply it as the starting frame while still providing the original video normally.

When compatible with the shot, this offers extremely direct control.

Instead of describing the desired look, you show the model exactly what the first frame should look like.

When It Does Not Work Well

The Change Is Not Visible in Frame One

Examples:

• a person enters later;

• a window only appears after the camera moves;

• a light switches on later.

There is nothing in the first frame for the model to anchor the change to.

Frame One Is Unrepresentative

Examples:

• extreme shoulder close-up;

• motion blur;

• whip pan;

• black frame.

In these cases, select a representative frame from the middle of the shot and use it as a reference image instead.

The Edited Frame Conflicts With the Video

Do not:

• move the subject;

• change the crop;

• change the pose.

The first frame and video would then describe the same moment differently.

Possible results include:

• ghosting;

• a jump on frame two;

• the edit fading out.

The Clip Is Too Long

The influence of a single frame decreases over time.

The end of a long shot may drift away from the look established at the beginning.

Rule of Thumb

Use an edited first frame when the change:

• is global;

• already exists at the beginning of the shot.

Examples:

• new location;

• new color grade;

• different time of day.

Use a reference image when:

• the important change appears later;

• the selected still does not perfectly align with frame one.

Getting More Original Detail Through

In order of importance:

1. Explicitly say what must remain unchanged.
   Name specific details that matter.

2. Match the output shape to the source footage.
   Rendering a vertical clip on a horizontal canvas causes cropping and zooming. This is easy to mistake for model failure.

3. Generate at the highest resolution your GPU allows.
   Missing detail cannot be recreated afterward.

4. Use an edited still when the exact look matters.
   One strong visual reference is usually more effective than several paragraphs of adjectives.

5. Keep each edit focused on one idea.
   Two unrelated changes in one prompt tend to interfere with each other. Two separate passes are usually more reliable.

6. Cut long material into individual shots.

Hardware

Testing for this release was performed on an RTX 5090.

That should be treated as the practical baseline for a comfortable experience at useful resolutions and clip lengths.

Smaller GPUs can still run the workflow, but the primary adjustment available is lowering resolution.

That usually does not remove the requested edit itself.

Instead, it reduces fidelity.

Typical degradation includes:

• increased flickering between frames;

• softer edges;

• mushier small details;

• visible artifacts around fast motion;

• problems with thin structures such as:

  • hair;

  • fingers;

  • text;

  • railings.

If you must generate at low resolution:

• keep clips short;

• plan a separate upscale pass afterward.

Known Failure Modes

Specific Faces

A particular person's face may look similar, but not identical.

Very Small Subjects

If a person occupies only a few percent of the frame, detailed edits are unlikely to survive.

Crop in first.

Self-Contradicting Prompts

Trying to replace a region while simultaneously asking the model to preserve that same region usually results in a weak, incomplete edit.

Wrong Output Shape

An aspect-ratio mismatch can create an unexplained-looking zoom or crop.

Undersized Reference Images

Very small references may appear to be ignored.

Text and Logos

Text and logos tend to be rewritten into approximations rather than preserved exactly.

Under the Hood, Briefly

For readers interested in the mechanism:

The source clip is encoded and packed as an aligned guide.

Each source latent is assigned the same position identity in space and time as the corresponding output latent.

In other words, the model is not viewing your clip as a separate reference running on its own clock.

It is effectively generating on top of the source clip.

That shared geometry is what helps preserve:

• timing;

• framing;

• motion.

It also explains why processing cost increases with clip duration:

both the source video and generated video are present in the sequence at the same time.