Most people have watched a generative video go sideways. You nail a great opening frame, a product hero shot or a portrait, feed it into a standard image-to-video model, and then you watch it wander. The camera drifts somewhere you never asked for, the background starts to melt, a character simply vanishes halfway through. Nobody planned any of that.
The two-keyframe approach in MiniMax H3 Max, served on fal, is the fix for exactly this. Instead of pinning only the first frame and hoping the model invents a sensible path, you lock the opening frame and the closing frame. The endpoint becomes an interpolation engine: it renders the motion, the camera move, and a matching audio track that carry Frame A cleanly into Frame B. You get 768p or 480p output at 24 frames per second, with synchronized stereo sound baked in.
One honest caveat before the how-to. Locking both ends guarantees the boundaries. It does not guarantee a clean, physically believable bridge if your two stills imply geometry that could not exist, or lighting that contradicts itself, or mismatched aspect ratios. Feed the model two incompatible worlds and it will cheat with warping, a blurry cross-dissolve, or a smeared morph. Everything below is about avoiding that trap.
Decide What the Transition Is For
Before uploading anything, be clear about the cinematic job the bridge needs to do. Interpolation is happiest connecting two states of one continuous event, not stitching together two unrelated scenes.
A few patterns that reliably work:
A product mechanism reveal, such as a sealed cosmetic jar on a pedestal ending with the lid hovering above it. This shows a physical feature without a hard cut.
A controlled camera move, starting on an extreme close-up of a face and pulling back to a medium shot that reveals the wardrobe. It reads as a smooth dolly out.
A lighting or time progression, a skyline under midday sun resolving into the same skyline at blue hour. The environment transforms inside a single shot.
A character action landing, an athlete caught mid-leap ending in a balanced crouch. The trajectory gets bridged deliberately rather than guessed.
Build Two Frames That Can Actually Coexist
Whether the bridge holds together is decided almost entirely by how compatible your two inputs are. Treat the stills as two photographs from the same shoot, not two random images you happen to like.
Keep the aspect ratio and resolution identical on both. If one is 1344 by 768 and the other is something else, you will see canvas warping or letterbox bars pop in. Keep the subject consistent: same facial geometry, same hair, same clothing. Change the coat between frames and the model dissolves the fabric mid-motion just to reconcile them. Keep the lighting and grade matched, same color temperature, same shadow direction, or you get exposure flicker and accidental cross-fades. And keep the physical distance plausible, ask yourself whether a real camera or a real actor could travel from A to B in five to ten seconds. A forty-five degree head turn or a slow orbit is fine. Jumping from a sunny street to a submarine interior in five seconds just forces the engine to hallucinate geometry that does not exist.
The practical version of all this: generate or capture both keyframes from the same source, the same seed, the same checkpoint, the same photo session. Consistency at the input is what buys you coherence in the middle.
Prompt the Motion, Not the Subject
Here is the mental shift that trips people up. With two keyframes, the images already say what the subject looks like. Your prompt should stop describing static details and start directing only the change between the frames: the action, the camera vector, and the sound.
Set the resolution to the native 768p tier, which balances speed against structural hold. Keep duration at five seconds and only stretch to eight or ten for genuinely slow moves, because shorter clips drift less. Leave prompt expansion on its balanced default so the motion description gets tidied without a long rewrite delay. And describe one clear sound, a mechanical click or an environmental tone, so the model lays synchronized foley and room tone in the same pass.
Then write a single verb and a single camera direction. For a product shot, something like: the camera orbits thirty degrees right while the lid unscrews and lifts away, with a soft metallic click and a faint studio hum. For a character shot: the subject turns smoothly to face camera and exhales, a light breeze moving the hair over quiet street ambience.
What you must not do is stack competing beats into a five-second window. Walk forward, grab a coffee, drink it, turn, sit down. The model will either sprint through the frames or abandon the end frame entirely trying to fit it all. One action, one move, one sound.
Inspect the Bridge Before You Trust It
Generation itself is quick, roughly three to eight seconds on fal's inference stack. The real work is the review pass, and it is worth running deliberately rather than glancing at the result.
Scrub through at quarter speed and check four things. First, the arrival: does the final second settle cleanly into Frame B? If it misses, your prompt is probably too busy or your aspect ratios do not match. Second, intermediate drift: does the face, the product label, or a hand warp around the middle of the clip? If so, the gap between your two stills is too wide and needs shrinking. Third, fake dissolves: has the model quietly cheated the whole transition with a blurry cross-fade? Counter it by telling the prompt you want continuous physical camera motion, no dissolves, no cuts. Fourth, audio sync: does the sound land with the on-screen action? Use the generated audio as a guide, but plan to master the final foley in post.
Watch specifically for tail-snapping, a favorite failure of diffusion interpolation. The clip looks fine for three seconds, then the last ten frames contort violently to reach Frame B. If the motion snaps at the end, throw the take away and reduce the physical distance between your stills.
Prep the Approved Shot for Your Edit
Once a bridge passes inspection, three small steps make it behave inside a non-linear editor. Trim the first two frames and the last three to clear minor boundary artifacts and get clean cuts. Conform the clip to a constant 24 or 30 frames per second so it matches your sequence settings instead of confusing them. Keep the native room tone on the first audio track but layer real commercial sound effects on a second track to emphasize the physical action.
Know the Limits
Two-keyframe control is powerful but bounded, and expecting more than it gives is how shoots get frustrated. You direct the endpoints, not the journey: H3 Max honors the composition of Frame A and Frame B, but the path between them is the model's call, and you cannot keyframe individual joints mid-motion. Output tops out at 480p or 768p, so a 4K commercial deliverable still needs an upscaling pass afterward. Fine text and small labels can soften during fast rotations before snapping back into focus at the end frame, so keep critical typography out of heavy camera moves.
Questions That Come Up
Does a bad-looking result get refunded?
No. On pay-as-you-go platforms like fal, billing tracks the raw GPU seconds consumed. A generation that completes without a system error gets billed whether or not it matches what you wanted creatively.
Can I swap just the end frame on a rerun?
Yes. In both the playground and the API you can keep the original start image and change only the end-image field, testing several destination angles against one opening still.
Will it take a transparent PNG?
It accepts the file but flattens the alpha against a solid background on the way in. For predictable color, composite your subject onto a clean flat backdrop before uploading.
Do uploaded frames stay in my history?
Images staged through temporary URLs follow the platform's CDN retention rules. You can clear session history in the sandbox, but API request payloads are kept under the standard dashboard data-retention policy.
Will the API tell me why a keyframe failed?
Only for basic input problems. An unreachable URL, an unsupported format, or a corrupt file header returns a normal 4xx error with a message. It will not diagnose creative failures like an extreme perspective jump or a lighting clash, because those are not errors, just bad input design.
The Takeaway
Locking both ends of a take removes the guesswork that makes text-only generation feel like a slot machine. H3 Max on fal gives you a real bridge between two frames, with sound, in seconds. Get the two keyframes compatible, keep the motion prompt to one action and one camera move, and inspect the result honestly before it reaches your timeline. Do that and the shots stop drifting and start landing exactly where you placed them.
