Make one object in your video come to life with Higgsfield

Higgsfield.ai’s video editor is a video-to-video tool: you hand it a clip you already filmed, and it re-renders it following your instructions — keeping everything you tell it to keep, and changing only what you tell it to change. The example below takes an ordinary talking-to-camera clip and makes a small robot figurine on the desk briefly wake up, gesture, and deliver a line — while the person, the camera, and every other object stay exactly as shot. Point the same technique at any single object in your own footage.

The magic isn’t the model. It’s the prompt — the more precisely you lock down what must NOT change, the more believable the one thing that does.

Step by step

Open Higgsfield and start an Edit Video project

Go to higgsfield.ai and sign in. From the menu, choose Edit Video — that’s the mode that transforms an existing clip, rather than generating a new one from text or a still image.

Upload your video

Upload the clip you want to edit. Keep it short — the example is about 6.7 seconds. Short clips render faster, cost less, and give the model fewer chances to drift.

Film with the end in mind: a locked-off camera (no zoom or pan), a still background, and the object you want to animate clearly visible and well lit. The less that moves in your original footage, the cleaner the result.

Paste the prompt

Paste the full prompt below into the prompt box. It looks long, but every line is doing a job — and that length is exactly why it works. (The last step explains how to adapt it to your own clip.)

Notice the <<<video_1>>> tags: that’s the placeholder for the clip you just uploaded. Leave them exactly as written — Higgsfield maps them to your video for you.

Higgsfield video-to-video prompt

=== THE EXISTING ROBOT FIGURINE BRIEFLY COMES TO LIFE (SAME SINGLE OBJECT, NO DUPLICATE, NO ADDED PROPS) — V2V on <<<video_1>>> (~6.71s) === ⚠️ SOURCE LOCK: this is an ANIMATION PASS on ONE existing object in <<<video_1>>>, NOT a new generation and NOT an added object. KEEP 1:1 the woman's identity/face, her exact performance, entrance, movement, gestures and timing, and the exact CAMERA (framing, angle, lens, any move) and EDIT/cuts. Do NOT re-frame, re-time, re-angle or re-cut. Do NOT alter the whiteboard, the brown desk, the microphone, the keyboard, the mouse, the speaker, or any other object in frame. ⚠️ CRITICAL — SINGLE ANIMATED ELEMENT ONLY: there is exactly ONE small robot figurine on the desk, and it is holding one smaller robot. That one figurine is the ONLY thing in the entire frame that is allowed to move at any point other than the woman. Same object, same position on the desk, same size, same scale relative to the desk, same paint/material, same contact point with the desk surface. Do NOT add a second robot, a duplicate figurine, a swapped design, or any new prop anywhere in the frame. Do NOT let the figurine slide, drift, grow, shrink, or change pose between shots. There must be a total of ONE robot figurine (plus the one tiny robot it holds) visible in every frame, and it is the original one. LOCATION <<<video_1>>>: a plain white whiteboard background; a brown desk in the foreground; on the desk a microphone, a keyboard, a mouse, a speaker in the background, and the one small robot figurine holding a smaller robot. A woman jumps into frame holding a microphone and speaks into it. Everything else in the room stays completely still and unchanged for the full duration. FORMAT/CUTS: 16:9, 30fps, ~6.71s, one continuous shot, camera never cuts, camera never moves, matches <<<video_1>>> exactly. OPTICS: as in source plate. No change to framing, focal length, depth of field or focus pulls. ACTION TIMING (by seconds): 0–3.6s: the woman jumps into frame exactly as in the plate, holds the microphone and talks into it, performance and timing identical to the source. EVERYTHING ELSE IS FROZEN. The robot figurine is completely inert: a still object, zero motion, zero eye light, zero servo twitch, indistinguishable from a static prop. Keyboard, mouse, speaker, desk and whiteboard are perfectly static. No drift, no breathing motion, no micro-warping on any object. 3.6–3.9s (BOOT-UP BEAT): the robot figurine "wakes." Tiny stepped servo motion, like stop-motion: a single sharp head twitch, a faint flicker as its eyes light up, one shoulder settling as if power just reached it. Weight stays planted on the desk. Half a second at most. 3.9–6.4s (PERFORMANCE BEAT): the figurine tilts its head up toward the lens, then performs a short gesture sequence while delivering the line "Comment TUTORIAL to learn how to do it." Motion reads as small, mechanical, charming, slightly jerky, never fluid or human. Gesture menu (pick and sequence, all small-scale and mechanical): - looks down at the smaller robot it is holding, then snaps its head up to camera - lifts one free arm and points directly at the lens on the word "Comment" - both arms mime tapping/typing in the air, like typing a comment - a two-handed presentational sweep toward the desk, a tiny "ta-da" - a stiff double thumbs-up, held one beat too long - a small shrug with both shoulders, elbows locked - a short wave with the free hand on the last word - optional micro-beat: the smaller robot it is holding gives one tiny nod or single-hand wave, then freezes again - optional: the head does one quick left-right scan, as if checking nobody is watching, before it starts 6.4–6.71s (POWER-DOWN BEAT): arms drop back down, head settles, eye light fades out, and the figurine returns to its EXACT original pose from frame 1, then is completely inert again for the remaining frames. The final pose must match the 0–3.6s pose pixel-for-pixel in position, angle and silhouette. WOMAN'S PERFORMANCE: unchanged throughout. She does not look at the robot, does not react to it, does not change position or gesture, and her timing is identical to the plate. She simply continues or finishes her line as in the source. PHYSICS / MATERIAL REALISM: the figurine is a solid physical object made of plastic/painted metal. Rigid limbs, hinged joints, no squash-and-stretch, no rubbery bending, no morphing. Its base stays fixed to the desk surface with no sliding or floating. Contact shadow on the brown desk updates correctly with the limb motion. The smaller robot stays held in its grip the whole time and never detaches, falls, or changes design. LIGHT-MATCH / INTEGRATION: lit by the same practical room light as the rest of the plate, same color temperature, exposure, grade, grain and lens character. No added glow, no rim light, no spotlight on the robot, no vignette shift, no exposure change when it activates. The only permitted light emission is the faint internal eye glow during the active beat. ANTI-SLOP / FORBIDDEN (strict): exactly one robot figurine total, holding exactly one smaller robot, in the original desk position. A second figurine, a duplicate, a design swap, or any new prop appearing anywhere in the frame is a failed generation. No movement whatsoever from the keyboard, mouse, speaker, microphone, desk or whiteboard at any point. No text, logos, writing or drawings appearing on the whiteboard. No robot motion before 3.6s or after 6.71s. No sliding, floating, teleporting, scaling or drifting of the figurine. No melting, morphing or limb-count errors. No human-fluid animation. No change to the woman's performance, position, face, hands, microphone or wardrobe. No camera move, re-frame, zoom, cut, speed ramp or crossfade. No flickering between poses across frames. No warping of desk edges or whiteboard seams. ~6.71s, 16:9 (match <<<video_1>>>). Diegetic SFX only: her voice and continuous room tone 0–3.6s; from 3.6s a faint mechanical servo whir during the robot's motion and the robot's line "Comment TUTORIAL to learn how to do it" in a small synthetic voice. Room tone continuous across the whole clip. No music. No on-screen text.

Choose the Gemini Omni Flash model

In the model picker, select Gemini Omni Flash. It’s the model this prompt was written for and tuned against. Different models read the same prompt differently, so if you swap it out you may need to adjust the wording.

Then hit generate. A first pass usually takes a minute or two.

Review, then re-roll if it drifts

Watch the result at full size and check the things the prompt tried to protect: did anything other than your chosen object move? Did a duplicate appear? Did the camera stay locked? Did the object return to its exact starting pose at the end?

AI video is a slot machine — the same prompt gives a slightly different result every time. If a pass breaks one of the rules, generate again before you rewrite. Two or three rolls usually lands a clean one.

Adapt the prompt to your own footage

The prompt is really a template. Swap the specifics for your own scene and the structure keeps working. These are the parts worth understanding:

The source lock. The opening block tells the model this is an animation pass on one existing object — not a new generation — and lists every single thing that must stay identical: the person, the camera, the framing, the cuts, each object on the desk. Rewrite that list to match what’s in your shot.

The single animated element. The prompt names exactly one thing allowed to move, then repeats — bluntly — that there must be no duplicate, no second copy, no design swap. Naming one object and forbidding copies is the difference between a clean result and a mess.

Action timing by seconds. The clip is broken into beats with timestamps: frozen, boot-up, performance, power-down. Give your object a clear “before,” a moment it comes alive, something to do, and a return to stillness — and match the timestamps to your clip’s actual length.

Physics and material. Telling the model the object is solid plastic and metal with hinged joints — no squashing, morphing or sliding — is what stops the animation looking like melting rubber. Describe what your object is really made of.

Light-match. “Same room light, same grade, no added glow or spotlight” keeps the animated object sitting believably inside the original shot instead of looking pasted on top of it.

The forbidden list. That long “anti-slop” block at the end is the most important part. Spelling out the specific ways a generation can fail — duplicates, drifting, morphing, camera moves, stray text — tells the model exactly what “wrong” looks like. Keep this list long and specific.

Val's tip

The counterintuitive rule with AI video: you get more control by writing more restrictions, not fewer. Most of this prompt is a list of things that must NOT change — and that’s precisely why the one thing that does change looks real. When a result goes weird, don’t loosen the prompt; add the failure you just saw to the forbidden list and roll again.