Use Balanced when you already know the shot. Use Quality when you have a short idea and want the prompt-expansion step to develop it before generation. Quality is not a higher-resolution render or an automatic “make this better” switch. It spends more time producing a richer rewrite.
That distinction matters because prompt expansion happens before H3 Max generates the video. If the brief contains two camera moves that fight each other, a longer rewrite still begins with a contradiction. Fix the direction first; then choose how much interpretation you want.
Our H3 Max generator currently offers Balanced and Quality. They use the same credits for the same duration and resolution. The trade-off is control and preparation time, not a different credit rate.
What prompt expansion changes
The current fal H3 Max Turbo schema describes prompt expansion as a rewrite step. Balanced usually returns that rewrite in about one second. Quality can spend up to roughly 30 seconds preparing a richer prompt before video generation begins. Those are provider descriptions of the rewrite stage, not promises about total wait time or the final picture.
The richer rewrite may spell out details that a short seed left open. That can help a one-line idea become directable, but fal does not guarantee which details it will add. The more exact your original brief is, the less useful that extra interpretation may be.
| Your prompt before generation | Better first choice | Why |
|---|---|---|
| One sentence with a subject and mood | Quality | The idea still needs a shot, movement, and sound direction. |
| A structured brief with one action and one camera move | Balanced | Most important decisions are already present. |
| Exact dialogue, product geometry, or a carefully composed first frame | Balanced | Additional interpretation can compete with details you need to preserve. |
| A result ignored part of a contradictory prompt | Neither solves it | Remove the conflict before paying for another take. |
| You are still deciding framing or motion | Balanced | Faster prompt preparation makes early iteration easier. |
There is also a Disabled mode in fal's public API. It sends the original prompt without the expansion step. The standard AI Live Generator interface only exposes Balanced and Quality, so this guide focuses on those two choices.
Balanced is for a brief that already has a spine
Consider this five-second product shot:
A small amber perfume bottle stands on pale limestone. One narrow band of warm light travels slowly from left to right across the glass, revealing the curved shoulder and a single condensation trail. Locked medium close-up, bottle centered, no camera movement and no extra objects entering the frame. Audio: quiet room tone and one soft glass chime as the light reaches the label; no music.The prompt already answers the useful questions: what is present, what changes, how the camera behaves, what must stay clear, and what should be heard. Balanced is the sensible first choice because the shot does not need another layer of scene design.
Before generating, decide what would make the result usable. For this brief, the bottle should keep its shape, the camera should stay fixed, and the light should cross the label once. Those checks are more valuable than asking whether the clip feels “more cinematic.”
If the camera drifts, do not move straight to Quality. First check whether another phrase suggests a reveal, orbit, or push-in. The camera movement guide shows how to diagnose viewpoint changes without rewriting the whole scene.
Quality is for an idea that still needs direction
Now start with a deliberately smaller seed:
A glassblower creates a tiny orange bird at night. Intimate and magical, with workshop sounds.This has a subject, an action, and a mood, but it leaves the viewpoint, timing, light, and sound sequence open. Quality can be useful when you are comfortable letting the rewrite develop those choices.
That freedom is the point—and the risk. If the bird must have a particular silhouette, the furnace must remain in a fixed screen position, or a spoken line must stay exact, put those requirements in the original prompt. Quality cannot know which unstated detail is commercially important.
You can also write the direction yourself and return to Balanced:
Inside a dark glassblowing workshop, a craftsperson turns a glowing orange bead at the end of a blowpipe and pinches it into a tiny bird. Begin on the hands and molten glass, then make one slow pull back to reveal the furnace behind them. The bird keeps a simple rounded body and two clear wings. Warm furnace light, deep black shadows, one continuous shot. Audio: steady furnace roar, metal tools touching once, and the maker's quiet breath; no dialogue or music.Quality is not automatically the better choice for this longer version. The prompt now has its own camera and sound plan. Balanced avoids waiting for a richer rewrite when the brief already supplies those decisions.
Compare the modes without changing the experiment
If a shot is important enough to compare, keep the test understandable:
- Save the exact prompt outside the generator.
- Keep duration and resolution fixed. For text-to-video, keep the aspect ratio fixed; for image-to-video, reuse the exact same source image or images.
- Generate the first take with Balanced.
- If the idea still feels under-specified, generate a second take with Quality using the same prompt and controls.
- Compare the requirement you named before starting: camera path, identity, product shape, spoken wording, or sound cue.
This is not a laboratory A/B test. The interface selects a random seed, so the comparison does not isolate prompt expansion. Holding the visible settings steady simply stops you from confusing an expansion change with a new duration, resolution, aspect ratio, or source image.
The fal API may return an expanded_prompt in its response, although the field can also be null when expansion leaves the prompt unchanged or occurs inside MiniMax's hosted service. Our standard generator does not display expanded_prompt, so it cannot be inspected or edited here. Judge the choice by the returned clip and your original acceptance criteria, not by assuming the hidden rewrite was longer.
Prompt expansion cannot repair the wrong input
For image-to-video, the uploaded frame supplies the composition, identity, clothing, product details, and aspect ratio that the model starts from. Use the prompt for movement, camera behavior, timing, and sound. If the source image has a distorted hand or an unwanted logo, Quality does not turn it into a clean source asset. Prepare the image first.
The same boundary applies to exact captions and approved brand copy. Generate clean visual space, then add final text in an editor. The on-screen text guide explains why editable typography is safer than spending more generations on exact lettering.
Prompt expansion also does not create a separate negative-prompt channel. Our interface has one prompt field. If you need a constraint, write the positive shot first and keep only the exclusion that changes the result. The negative prompt guide has a practical example.
The cost stays the same; the decision does not
On the current site tariff, a five-second standard H3 Max take costs 15 credits at 480P or 25 credits at 768P. Balanced and Quality cost the same number of credits. Two five-second 480P comparisons cost 30 credits; two at 768P cost 50.
That does not mean every idea deserves two takes. If the brief is already specific, start with Balanced and spend the second generation only when the first result reveals a concrete question. If the idea is still a sentence fragment, either choose Quality or spend a minute writing the missing camera, action, and sound decisions yourself.
The useful rule is simple: the more authorship already present in the prompt, the less expansion it needs. Use the H3 Max prompt library to build the brief, then open the generator with one result you know how to judge.
The next shot is yours
Take the idea into the studio.
Start with a prompt or an image. Choose the duration and resolution, then check the credit cost before you generate.
Create an H3 Max video