00Start
Start here: working in Lemonpeel
This handbook starts with one-object swaps and builds to full creature VFX and professional relighting. Read this chapter once; every later chapter assumes it.
Every method in this handbook runs on a Lemonpeel canvas. A canvas is a board of connected cards, called nodes. You drop your footage and images on as asset cards, add an image or video node, pick a model, paste a prompt, connect the inputs and run it. Each chapter lists which nodes to build and which model to pick.
The canvas routine for every method
- Open your project and canvas. Keep one canvas per shot so references never get mixed up.
- Upload your raw clip and any reference images. They land on the canvas as asset cards.
- Add the node the chapter asks for (image or video), then open its model picker and choose the model named in the chapter. Leave it on Auto only when the chapter says so.
- Connect each asset card into the node. For video edit models, connect the video first, then the reference images in the order the prompt describes them.
- Open the node settings and set aspect ratio and duration to match your source clip before you run anything.
- Paste the prompt, check the credit estimate, and run. Keep every take on the canvas so you can compare them side by side.
Three rules that apply to every prompt
- One prompt, 2,500 characters maximum. Lemonpeel has no separate negative-prompt field. Everything you want to avoid goes at the end of the main prompt as plain instructions, like "no morphing, no flicker, no text". Every prompt in this handbook already follows this and shows its character count.
- Say what to keep before what to change. Video edit models change too much when you don't lock things down. Open with "Keep his face, movement and camera exactly as filmed", then describe the change.
- Use frame directions only. Write "frame left" or "toward the camera", never "his left hand side". When an actor faces the camera, his left is the viewer's right, and the model gets confused.
Model map
The original class notes used OpenArt. Here is where each tool lives in Lemonpeel.
| Job | Originally used | Use in Lemonpeel |
|---|---|---|
| Edit a still frame, build a keyframe or remove an object | Nano Banana Pro, Adobe Firefly | Nano Banana Pro Edit (takes up to 14 reference images) |
| Change something inside a video and keep the performance | Kling 1 Video-to-Video, Seedance 2.0 V2V | Kling o3 Video Edit (up to 4 reference images). Kling o1 Video Edit is the cheaper fallback. |
| Make a still image copy the motion of a video | Kling 3.0 on OpenArt, DreamActor M2.0, Motion Sync | Kling 3 Motion Control. Wan Animate Move is the alternative. |
| Swap a character in existing footage | Kling O1 Replace Character | No one-click tool. Use Kling o3 Video Edit with a character sheet split into face, front and back images (Chapter 6). |
| Build a new shot from references and a motion clip | Vidu Q2 Elements to Video | Kling o3 Video Ref to Video. Seedance 2.5 Ingredients takes up to 30 reference images. |
| Final relight and realism polish | None | Luma Ray 3.2 Video Edit |
| Google's video edit model | Gemini Omni Flash | Gemini Omni 1.1 Edit (text only, no reference images) |
01Beginner
Object replacement
Swap one object in your hand for another, with realistic light from the new object.
The simplest AI VFX move: film yourself holding one object, then turn it into something else in the video. Here a pen becomes a lit candle, and the candle's glow lights the face and wall.
Canvas setup
- Film a short clip (5 to 10 seconds) holding the object steady in good light.
- Take a screenshot of one clear frame and drop it on the canvas.
- Add an image node with Nano Banana Pro Edit, connect the frame, and run the image prompt below. This makes the look reference.
- Add a video node with Kling o3 Video Edit. Connect the clip first, then the candle frame you just made. Run the video prompt.
Step 1: reference image
Edit this frame only where the pen is. Replace the black pen in his hand with a slim white wax candle with a lit flame, the same length and angle as the pen, held in the same grip with the same fingers. Keep his face, expression, hair, beard, shirt, pose, the room, the camera angle and the framing exactly as they are. The flame is small and steady, with a bright yellow core and a soft orange edge. Add the warm glow it would really cast: warm light on his fingers and the candle, a soft warm lift under his chin and on his cheeks, and a faint warm pool on the wall behind him. Darken the rest of the room slightly so the flame reads as the main light. Photoreal, same lens, grain, sharpness and colour as the original photo. No other changes, no extra objects, no text.
Step 2: apply it to the video
Video edit of the reference clip. Replace the pen in the man's hand with the lit candle from the reference image, in every frame. The candle follows his hand exactly: same grip, same position, same angle and every movement, frame for frame. Keep his face, expression, clothes, the room and the camera completely unchanged. The flame flickers gently and leans slightly against the direction he moves the candle, with faint heat shimmer above it. The candle's warm glow lights his fingers, the candle, his face and the wall behind him, and the light on him moves and breathes with the flame. The rest of the room sits slightly darker so the candle is the main light source. Photoreal, same grain, sharpness and colour as the original footage. Strictly no other changes to the video: no extra objects, no change to his face, no flicker in the room, no text.
02Beginner
Put yourself in a new world
Rebuild the world and costume around you in one frame, then drive it with your real motion.
Take one frame of yourself, rebuild the world and costume around you as a still, then make that still move exactly like your original footage. The example turns a stairwell shadow-boxing clip into a Muay Thai championship ring.
Canvas setup
- Screenshot a frame from your clip where your pose is clear and your face is sharp.
- Image node with Nano Banana Pro Edit: connect the frame and run the image prompt. Generate three or four and keep the best face match.
- Video node with Kling 3 Motion Control: connect the chosen keyframe as the image and your original clip as the motion video. Run the video prompt.
Step 1: keyframe
Keep this man's face exactly as in the photo: same identity, features, beard, hairline and expression, with no beautification or smoothing. Keep his exact body pose, fighting guard, gloves and the camera angle. Change everything else. Place him standing in the centre of an international professional Muay Thai championship ring: white canvas floor, coloured ropes and corner pads, a lighting rig overhead with bright spotlights, sponsor boards around the ring apron. Behind the ropes, a packed crowd in tiered seats, softly out of focus. Change his outfit to a professional Muay Thai fighter: shirtless, traditional satin Muay Thai shorts with embroidered waistband, ankle supports, hand wraps visible at the wrists, keeping his gloves. Give him a lean, conditioned build that suits his frame. Adjust the framing only as much as needed to show him full body on the canvas, with correct physics: feet planted, weight on the canvas, soft contact shadows under his feet. Relight him for the arena: hard white top light from the rig, cool blue edge light from the side, warm natural skin tones, light sweat sheen. Photoreal, ultra-detailed, cinematic sports-drama look with deep contrast and cool steel-blue shadows. No text on screen other than the sponsor boards.
Step 2: animate it with your motion
Animate the fighter from the reference image with the exact motion of the reference video: every guard shift, bounce, step, punch and head movement, matching the timing frame for frame. Keep his face, body, outfit, gloves and the arena exactly as in the image. Keep the camera framing from the image, with a subtle handheld feel. The crowd in the background moves naturally, standing and cheering. The rig spotlights shimmer and sweep slowly. His shorts and hand wraps react to his movement, and sweat catches the light as he moves. Photoreal, consistent identity in every frame, feet grounded on the canvas with contact shadows. No face morphing, no extra limbs, no warped gloves, no text.
03Intermediate
Scene, wardrobe and prop swap
Replace the location, outfit and prop in one pass while the performance stays yours.
Change three things at once while keeping the face and camera: the location, the outfit and the prop. The example turns a home staircase with a toy rifle into a special forces soldier in a bombed building.
Canvas setup
- Screenshot a clear frame. Find or generate a clean image of the exact rifle you want.
- Image node with Nano Banana Pro Edit: connect the frame first, then the rifle image. Run the image prompt.
- Video node with Kling o3 Video Edit: connect the clip first, then the keyframe. Use video prompt A.
- For a period piece with separate uniform, weapon and location references, connect the clip, then the rifle, uniform and environment images in that order, and use video prompt B.
Step 1: keyframe
Use Image 1 (the frame from my footage) and Image 2 (the rifle). Keep the man's face exactly as in Image 1: same identity and features, no beautification. Keep the camera angle, lens, framing and his body pose exactly as in Image 1. Change only three things: 1. Location: a narrow concrete stairwell inside a partially destroyed building in a war zone. Cracked walls, exposed rebar, rubble and dust on the steps, daylight coming through a broken opening in the wall. 2. Outfit: modern special forces multicam combat uniform, plate carrier with pouches, tactical helmet with headset and comms mic, tactical gloves. 3. Weapon: replace the gun he is holding with the exact rifle from Image 2, at correct real-world scale, in the same grip and aim. Lighting: hard directional daylight through the broken opening, dust floating in the light, cool shadows and warm highlights, light bouncing off the concrete onto his gear. Photoreal, ultra-detailed, gritty modern war-film look, slightly desaturated teal-and-sand grade.
Step 2A: video from one keyframe
Video edit of the reference clip. Use the reference image as the look for the whole shot. Keep the man's face, his performance, every movement, the camera motion and the timing exactly as in the video, frame for frame. Transform his clothing into the uniform, plate carrier, helmet and comms from the reference image. Replace the gun in his hands with the rifle from the reference image, at correct scale, following his grip and aim exactly in every frame. Replace the background with the destroyed concrete stairwell from the reference image, matching the same camera angle and perspective, with parallax that follows the original camera motion. Dust drifts through the light and small debris sits on the steps. Relight him with hard directional daylight from the broken opening, soft bounce from the concrete, and dust catching the light around him. Gritty, slightly desaturated teal-and-sand grade, photoreal. Keep his face, uniform and rifle identical in every frame. No morphing, no extra people, no added cuts, no text.
Step 2B: video from separate prop, uniform and location references
Video edit of the reference clip, using three reference images: Image 1 is the rifle, Image 2 is the uniform, Image 3 is the environment. Keep the man's face, performance, every movement, the camera motion and the timing exactly as in the video, frame for frame. Transform him into a WWII infantry soldier wearing the worn uniform from Image 2: mud-stained fabric, field belt and pouches, the same textures in every frame. Replace his weapon with the rifle from Image 1: wooden stock, correct real-world scale, following his grip and aim exactly. Replace the environment with a muddy WWII battlefield based on Image 3: shell-damaged buildings, scattered debris, wet ground with puddles, light smoke drifting across the frame. The opening frame must match the tone and layout of Image 3, and the background must move with the original camera motion and parallax. Lighting: overcast daylight, soft shadows, light diffused through the smoke, a little mud splash and grime on his boots and hands. Grounded WWII realism, slightly desaturated grade, photoreal. Keep his identity, uniform and rifle stable in every frame. No added cuts, no extra characters, no reflections, no text.
Connect the images in the same order the prompt names them: rifle, uniform, environment.
04Intermediate
Full environment change
Replace every surrounding surface, relight to match, and keep the performance untouched.
Keep the person and the performance, and replace everything around them. The example upgrades an everyday car interior to an ultra-luxury sports car, with the street still moving outside.
Canvas setup
- Find a clean reference photo of the target interior, shot from roughly your camera angle.
- Video node with Kling o3 Video Edit: connect the clip first, then the interior image. Run the prompt.
Video edit of the reference clip. Keep the man exactly as filmed: his face, expression, clothes, every movement of his head, hands and steering, and the camera position, frame for frame. Replace the car interior with the ultra-luxury sports car interior from the reference image: its steering wheel under his hands, the dashboard and screens, seats, stitching, trim and panoramic glass roof, at the correct scale and perspective for this camera angle. His hands grip the new steering wheel exactly where they gripped the old one and follow every turn. Keep the street moving past the windows with natural motion and parallax that matches the original drive. Relight him to match the new cabin: soft daylight from the panoramic roof, faint glow from the dashboard screens, and the colour of the leather and trim bouncing softly onto his clothes and skin. Photoreal, same grain and exposure as the original. No changes to his face, no extra passengers, no text.
05Intermediate
Motion transfer
Make any character image copy your exact performance, with a living background.
Drive a finished character image with the motion from any clip, including a phone video of you in your room. This is Chapter 2 pushed further: a POV shot where the fighter calls the camera into the fight.
Canvas setup
- Use the character keyframe from Chapter 2, or make a new one with Nano Banana Pro Edit.
- Record the motion you want, facing the camera, in one continuous take.
- Video node with Kling 3 Motion Control: image in, motion video in. The prompt is optional but it adds crowd, lights and camera feel.
POV shot at an international Muay Thai championship. The fighter from the reference image faces the camera, calling his opponent into the fight, following the exact movement and timing of the reference video. The camera has a slight handheld feel, as if seen through the opponent's eyes. Behind the ropes the crowd is on its feet cheering, arms raised. The rig spotlights sweep slowly across the ring and flare softly into the lens. Sweat glistens on his skin, his shorts move with his steps, and his feet stay planted on the canvas with contact shadows. Keep his face, body and outfit identical to the reference image in every frame. Photoreal sports-broadcast realism. No face morphing, no extra limbs, no text.
06Advanced
Character replacement
Replace the person in a shot with your own character, frame-accurate and lit like the original.
Swap the person in existing footage for your own character while keeping every movement, the lighting and the background. There is no one-click "replace character" tool in Lemonpeel, and that is fine: Kling o3 Video Edit does it well when you feed it a character sheet split into separate images.
Build the character sheet
A good sheet has a front view, a back view and a clean face close-up, all on a plain grey background. Then split it into three separate images. Kling matches identity far more reliably from three clean images than from one busy sheet.
Create a character sheet of this exact person on a flat mid-grey studio background, in three panels side by side: 1. Full body, front view, standing straight, arms relaxed. 2. Full body, back view, same pose. 3. Head-and-shoulders close-up of the face, looking straight into the camera. Same face, hairstyle, outfit, skin tone and body proportions in all three panels. Soft, even studio light with no strong shadows, neutral expression, sharp focus. Photoreal and highly detailed, with no text, labels or watermark.
Connect one or two clear photos of your character. Then crop the result into three images: face, front, back.
Canvas setup
- Cut the source footage into shots of 10 seconds or less. Long shots drift in identity.
- Video node with Kling o3 Video Edit: connect the shot first, then face, front and back references, in that order.
- Run two or three takes of every shot and keep the one where the face holds through head turns.
Shot 1: wide running shot
Video edit of the reference clip. Replace the running man with the exact character from the reference images. Change nothing else. Identity: the reference character's face must replace the man's face completely and stay identical in every frame. Young Indian man in his early twenties, lean face, warm light-brown skin, thick messy black curly hair falling over his forehead, thick dark eyebrows, dark brown eyes, straight nose, light stubble and thin moustache, slim build. Match the reference face exactly, not a lookalike, with no trace of the original man's face, hairstyle or features. Wardrobe from the reference: olive green button-up cotton shirt with sleeves rolled to the forearm, charcoal grey slim trousers, clean white sneakers, thin silver bracelet on his wrist. The shirt and curls react to the running motion and the sea wind, fabric fluttering naturally. Performance: keep the original man's running motion exactly, every stride, arm swing, body lean and head movement, frame for frame. Keep his exact position, size and path in the frame, running toward the camera across the black pebble beach. Same camera, framing and timing. Keep the environment completely unchanged: the tall black sea stack, the flock of white birds and their exact flight, the grey overcast sky, the sea and the black pebble shore. Lighting match: soft, flat, cool overcast daylight from the grey sky, gentle top light, low contrast, cool blue-grey tones on his skin and clothes, soft shadows, matching the scene's grade. His feet touch the pebbles with natural contact shadows and pebbles shifting under each stride. Birds passing in front of him still pass in front of him. Photoreal, seamless, same film grain, focus and motion blur as the original footage. Keep face, hands and body shape stable and consistent in every frame: no face morphing, no identity drift, no flicker, no extra limbs, no warped fingers, no cartoon or CGI look, no text.
Shot 2: close two-shot
Video edit of the reference clip. Replace both people with new characters while keeping every movement, pose, touch and camera move exactly the same. The man becomes the exact character from the reference images. Young Indian man in his early twenties, lean face, warm light-brown skin, thick messy black curly hair, thick dark eyebrows, dark brown eyes, straight nose, light stubble and thin moustache. Match the reference face exactly in every frame, including in profile and when partly covered by her hands, with no trace of the original man's face or hairstyle. Wardrobe: olive green button-up cotton shirt with the top button open and sleeves rolled, thin silver bracelet. The woman becomes a new, original young Indian woman who does not resemble the original actress in any way: a different face, soft oval features, long wavy dark brown hair, small gold earrings, a sleeveless gold sequinned top and gold rings, keeping the same styling mood. Her face must stay consistent in every frame. Performance: keep both performances exactly as filmed: the forehead kiss, her head tilting back, her hands resting on his chest and then holding his face, his hand on her neck, eyes closing, the lean toward each other, and the hair moving in the wind, frame for frame. Keep their exact positions, sizes and framing, with the same camera drift and timing. Keep the environment unchanged: the soft-focus green mountain slopes and the pale blue sky. Lighting match: warm low golden-hour sunlight from behind them, a glowing golden rim on both characters' hair, cheeks and shoulders, with soft warm fill on their faces, matching the original exposure and colour grade. The warm light and shadows of her hands fall correctly on his face and curls, and the collar of his shirt catches the rim light. Photoreal, seamless, same shallow depth of field, lens softness and grain as the original footage. Hands and fingers stay natural with correct finger count, and faces stay stable through every frame and every head turn: no face morphing, no identity drift between frames, no blending of the old and new faces, no flicker, no cartoon or CGI look, no text.
This prompt replaces both people. Never keep a real actor’s face in a romantic or intimate scene with someone they did not film with. If you have a sheet for your female lead, attach it as a fourth image and swap her description for "the exact woman from the fourth reference image".
07Advanced
Recreate a trailer shot
Rebuild a famous shot with your own character, matching its action and camera beat for beat.
Take a shot from any trailer and rebuild it with your character in the same world, with the same action and camera move. You make a clean background plate, get an exact description of the shot, then generate a new video that follows the original's motion.
Canvas setup
- Trim the trailer shot to 5 seconds or less and screenshot one frame.
- Image node with Nano Banana Pro Edit: connect the frame and run the removal prompt to get a clean plate. Then run the quality prompt on the clean plate.
- Upload the trailer clip to Gemini (or any assistant that can watch video) with the description prompt below. It writes the shot prompt for you.
- Video node with Kling o3 Video Ref to Video: connect the trailer clip as the reference video, then the clean plate and the character sheet. Paste the generated prompt, or adapt the sample.
- If Kling ignores the plate, try Seedance 2.5 Ingredients with the clean plate, the original frame and the sheet as its reference images.
Remove the character
Remove the person completely from this frame and rebuild what was behind them: the ground, fires, smoke, debris and building lights continue naturally through the space where they stood, with matching perspective and lighting. Keep everything else exactly the same: framing, composition, light, colours, smoke density and grain. No person, no silhouette, no shadow of a person, no smudge or blur where they were.
Sharpen the clean plate
Recreate this exact image at the highest quality: same composition, same objects, same lights, same colours and the same camera angle. Sharpen the detail, clean up compression artefacts and noise, and make textures crisp and natural. Do not add, remove or move anything.
Get an exact shot description
I am recreating the attached video shot with an AI video model. I will give the model three references: the original video clip as the motion reference, a clean background plate with the character removed, and a character sheet of my new character. Watch the whole clip carefully and write one generation prompt, under 2,400 characters, in plain English, with these sections in order: Setting, Character, Action with timestamps, Camera, Lighting, Look. Rules: - Describe exactly what happens, second by second, including where the character enters, lands, turns and what the hands do. - Describe the camera move precisely (static, push-in, pan, handheld), and its speed. - Use only frame directions (frame left, frame right, toward camera), never the character's own left or right. - Refer to the references as "the reference video", "the background image" and "the character sheet". - Describe the character only by what is in my character sheet, not by the original actor. - End with one line of quality guardrails: consistent face, no morphing, no extra people, no text.
The instruction to stay under 2,400 characters leaves you room to add a line in Lemonpeel.
Sample shot prompt
Recreate the reference video as a new shot. Follow its motion, timing, framing and camera move exactly, frame for frame. Use the background image as the environment and the character sheet as the only character. Setting: a dramatic night scene filled with thick, drifting smoke and dust. An old gothic stone church stands in the background, outlined with draped strings of glowing yellow fairy lights. The dusty ground is scattered with small pools of burning fire. Warm, flickering orange and yellow firelight against a dark night sky. Character: the Indian man from the character sheet. His face, clear aviator glasses, moustache and goatee stay identical in every frame. He wears the detailed tactical full-body red armour suit from the sheet, with raised muscle panels, glowing red-orange light strips and a metallic gold "E" emblem on the chest. Action with timing: 0.0-0.8s: he drops fast from above the top of the frame and lands in the centre in a deep bent-knee crouch, kicking up a burst of dust on impact. 0.8-2.0s: he rises from the crouch, squares his shoulders to the camera and settles into a wide-legged power stance, arms slightly flared. 2.0-4.0s: he pulls a red cloth mask up over his mouth and nose with both hands and locks his eyes on the lens in an intense stare. Camera: a smooth, continuous dolly-in for the whole shot, ending on his face and upper body. Lighting: firelight from below and behind, warm rim on the armour, glints on the gold emblem and glasses, embers and smoke passing in front of him. Look: photoreal action-film realism, anamorphic feel, fine grain, warm orange grade with deep shadows. Consistent face and suit in every frame, no morphing, no extra people, no text.
08Advanced
Green screen creature fight: the two-pass Kling method
A step-by-step pass structure for adding a creature that fights your actor, tested and debugged.
This is the full workflow from our Titan test: an actor on green screen, turned into a warrior fighting a colossal creature on Saturn's moon. The big lesson from testing is that one prompt cannot do everything. When a single prompt asks for a new world, relighting, a creature and a fight, the model drops half of it. So each pass does one job, and each pass feeds its output video into the next.
Before you prompt: time the plate
Scrub through the plate and write down when each action happens. Our plate: guard until 1.4s, step across until 1.8s, shield overhead from 1.8 to 2.4s, a spin until 2.8s, first lunge at 2.8s, then repeated thrusts to the end. Creature prompts tied to these timestamps sync far better than ones that say "when he raises his shield".
Pass 1: world and relight, no creature
Video edit of the reference clip. Replace only the background and relight the performer. Do not add any creature or other character. Keep the performer exactly as filmed: same face, red hair, full red beard, brown leather long coat, olive trousers, black boots, sword and round wooden shield. Keep his exact position in the frame, his path across the frame, his size and every movement, frame for frame. Same static wide camera and framing. Replace the green screen and all tracking markers with the surface of Titan, Saturn's moon, built in deep layers. Foreground: out-of-focus dark ice boulders in the bottom corners of the frame. Ground: wet black ice pebbles, frost crust and reflective puddles that mirror the orange sky and reflect his legs. Midground: a black glassy methane lake with slow ripples across the back of the frame, jagged ice spires and eroded ice cliffs. Background: tall mesas and a distant ice volcano releasing vapour, fading into thick layered orange-amber haze. Saturn only as a faint pale ringed shape behind the haze. Light drizzle and low mist drifting across the ground and swirling around his boots as he moves. Relight the performer completely so he belongs in this world. Remove the flat studio lighting and every trace of green spill on his outline, hair, beard and coat. Soft dim orange-amber skylight from above, darker exposure than the original, low contrast, no hard sunlight. Warm orange bounce from the ground on the underside of his coat and shield. Wet sheen and small droplets on his leather coat, hair, beard and shield. Dark contact shadow under his boots that moves with his feet, ice grit and mist kicked up by every step. Photoreal feature-film look, large-format cinema camera, anamorphic lens, shallow depth of field, fine film grain, cool orange-teal grade, deep blacks. Seamless integration, no cutout edges, no flicker, no cartoon look, no text.
Pass 1B: fix the lighting and grounding
Our first Pass 1 gave good depth but left the actor lit from the front, like the studio, against a backlit sky. It also looked like the Arctic instead of Titan. Pass 1B fixes only those things and keeps everything else.
Video edit of the reference clip. Keep the composition, the environment layout, the camera, and the performer's position, size, face, costume, sword, shield and every movement exactly as they are. Do not add any creature or character. Only fix the lighting, grounding and colour. Relight the performer to match the backlit scene. The glowing light source in the hazy sky is behind him, so his front, face, coat and trousers fall into soft shadow, darker than now, with low contrast. Add a warm orange rim light tracing his outline, hair, beard, shoulders and the edge of his shield. Soft orange-amber ambient fill from the hazy sky, with no flat frontal studio light. Add a faint wet sheen on his leather coat and shield, catching the rim light. Ground him in the scene. Add a soft dark contact shadow under his boots that moves with every step, plus small bursts of ice grit and low mist kicked up by his feet as he moves. Add a faint blurred reflection of him in the wet ground and puddles, and thin low mist drifting across the shoreline around his legs, with light drizzle falling. Shift the whole world toward Titan: reduce the Earth-like blue. Tint the ice cliffs, lake and rocks with muted amber, brown and grey-teal tones under the orange haze, with the water dark and glassy, reflecting orange. Increase the haze density with distance so the far mountains are faint silhouettes. Photoreal feature-film look, anamorphic lens, shallow depth of field, fine film grain, cool orange-teal grade, deep blacks. No flicker, no cutout edges, no changes to his motion, no text.
Pass 2: add the creature, locked to timestamps
Video edit of the reference clip. Keep the environment, camera and the man exactly as they are: face, costume, sword, shield, position and every movement, frame for frame. Add one colossal monster that fights him, perfectly timed to his moves below. "Frame left" means the left side of the screen as the viewer sees it. The monster: alien creature twice the man's height, head near the top of the frame, dark frost-crusted armour plates, faint cyan veins, long segmented arms with scythe claws, heavy grounded feet. It always faces the man and never faces the camera. It never overlaps him and he never walks through it. EXACT TIMING: 0.0-1.4s: the man stands in guard at frame left. The monster charges out of the haze from the far background toward him, growing larger as it approaches, and stays behind and beyond him in depth, never beside him. 1.4-1.8s: the man steps toward the centre of the frame. The monster arrives at frame left, directly behind where he started, towering and raising one claw high. 1.8-2.4s: the man lifts his shield above his head. At that exact moment the monster's claw comes down from above and smashes onto the shield. Ice bursts from the impact and his body dips under the weight. 2.4-2.8s: the man spins around. The monster pulls its claw back and rears up, roaring. 2.8-3.4s: the man lunges and thrusts his sword toward frame left. The sword tip strikes the monster's reaching forearm with sparks and ice spray, and the monster jerks back. 3.4-7.0s: the man holds his lunging stance and keeps thrusting toward frame left. The monster stays at frame left, facing him and trading blows the whole time. Each time his sword pushes forward, the monster swipes at him with a claw, he parries, and it staggers back one step, then lunges again. It never freezes and never stands with its arms hanging. Lighting: backlit orange haze with a warm rim on the monster's armour. Its cyan glow falls on the man's side facing it, on his shield and on the wet ground. Both cast shadows on the ice. Photoreal feature-film VFX, fine grain, orange-teal grade. Keep the man's hands, sword and shield stable. No morphing, no flicker, no cartoon look, no text.
Replace the timestamps with the ones you measured from your own plate. If sync is still off after two takes, split the plate at 3.4s and run each half with only its own timing lines.
Optional: Gemini Omni version
Omni blocked our first attempts. The likely triggers were combat words ("monster", "slashes", "blade connects") and real film titles. This version frames the scene as a staged fantasy duel. Note that Omni Edit takes no reference images and tends to re-stage the actor, so use it as a backup rather than the main path.
Cinematic fantasy video edit of the reference clip. Keep the performer exactly as filmed: same face, red hair, full red beard, brown leather long coat, olive trousers, black boots, prop sword and round wooden shield. Keep his position in the frame, his path and every movement unchanged, frame for frame, with the same static wide camera. Do not move, resize or re-stage him. He is in a staged fantasy duel with a giant creature on the frozen surface of Titan, Saturn's moon. The creature is three times his height and always faces him, staying close enough for its long arms to reach him. It never looks at the camera and never stands still: it breathes, shifts its weight, follows him with its head and keeps reaching for him. When he raises his shield overhead, its arm comes down from above and lands on the shield, releasing a spray of sparkling ice crystals. When he sweeps his sword toward it, the sword meets its arm with a flash of light and ice sparkles, and it sways back before reaching for him again. Its heavy steps press into the ice and stir the mist. Creature design: smooth dark grey armour-like shell dusted with frost, faint glowing cyan lines, long segmented arms with curved tips, and a wet, glistening surface. World: deep and layered. Out-of-focus dark ice boulders in the foreground, wet black ice pebbles with reflective puddles, a still black lake, ice cliffs and spires, and distant mesas fading into thick orange-amber haze. A faint ringed Saturn is barely visible through the haze. Gentle drizzle and low mist move with the action. Lighting: relight the performer to match this world. Dim, soft orange-amber skylight, darker exposure, no studio light and no green tint on his outline, hair or beard. A gentle cyan glow from the creature falls on the side of him that faces it and on the wet ground. Damp sheen on his coat and shield, and natural shadows under both figures. Style: photorealistic, large-format cinema camera look, anamorphic lens character, fine film grain, cool orange-teal colour grade, rich shadows, seamless integration. Keep hands, sword and shield steady, no flicker, no large mist bursts covering the performer, no cartoon look, no text or watermark.
What went wrong in testing, and the fix
| What we saw | Why | Fix |
|---|---|---|
| The creature stood behind the actor or on the wrong side | The prompt mixed the viewer's left with the actor's left, and put the creature where the actor starts | Frame directions only, defined once, and the creature placed where the actor is not |
| The actor walked through the creature | The creature was placed in the actor's path | Map the actor's path first, then write "never overlaps him" |
| The creature froze in the second half | It had no actions of its own for that stretch | Give every second of the clip a creature action |
| Hits didn't land | The actions were tied to moves the model couldn't detect | Tie the actions to measured timestamps |
| The actor looked pasted on | The relight was asked for in the same pass as everything else | A separate relight pass (1B) before the creature |
09Pro
Beeble: SwitchX and Beeble Studio
Keep your actor pixel-perfect and change the world and light around them.
Generative video edit models redraw your actor on every frame, which is where identity drift and flicker come from. Beeble works the other way. It keeps your original pixels and changes the world and light around them. That makes it the pro choice for relighting and background swaps, with Lemonpeel still handling anything that has to be invented, like a creature.
SwitchX: video-to-video with the subject locked
- Upload the clip to the Beeble cloud app and open SwitchX.
- Pick a mask mode. Auto finds the subject and tracks it; use it for most background swaps. Select lets you click the parts to keep, for example the face and hands while the costume changes. Fill selects the whole frame, for relighting or restyling a scene without moving anything. Upload alpha takes a matte you made in Nuke or After Effects, for pixel-exact edges.
- Add a reference image of the finished look. The easiest way is one frame of your plate edited in Lemonpeel with the keyframe prompt below. SwitchX takes the environment, light and grade from it.
- Paste the shot prompt. Keep camera tracking on (follow source) if your camera moves.
- Render, then bring the result back to Lemonpeel if you still need to add a creature or effects.
Reference keyframe for SwitchX
Turn this frame into the reference image for SwitchX. Keep the man exactly as he is: face, hair, beard, costume, sword, shield, pose, size and position in frame. Replace the green screen with the frozen surface of Titan, Saturn's moon, in deep layers: dark ice boulders in the foreground corners, wet black ice pebbles with reflective puddles, a black glassy methane lake, ice cliffs and spires, distant mesas fading into thick orange-amber haze, and a faint ringed Saturn high in the haze. Relight him to match: dim soft orange-amber skylight, darker exposure, a warm rim on his outline and shield, no studio light, no green spill, a wet sheen on his coat, and a dark contact shadow under his boots. Photoreal, large-format cinema look, fine grain, cool orange-teal grade, deep blacks.
SwitchX shot prompt
Keep it a plain description of the finished shot, with no left or right. SwitchX keeps the actor's performance from the source, so the prompt only has to describe the world, the fight and the light.
Hyper-realistic cinematic shot of a red-haired, bearded warrior in a brown leather long coat, with a sword and a round wooden shield, fighting a colossal monster on the frozen surface of Titan, Saturn's moon. The monster is three times his height, with dark obsidian armour plates crusted with frost, faint glowing cyan veins, long segmented arms with curved claws, and a wet, dripping body. It is constantly attacking him, slamming its claws down onto his shield and recoiling when his sword strikes it, while ice shards and mist burst from every impact and footstep. The environment is deep and layered: dark ice boulders in the foreground, wet black ice pebbles with reflective puddles, a black glassy methane lake, jagged ice cliffs and spires, and distant mesas fading into thick orange-amber haze. A faint ringed Saturn is barely visible through the haze, and light drizzle and low mist drift across the ground. Lighting: dim, soft orange-amber skylight from the hazy sky with low contrast and no hard sunlight, a cool cyan glow from the monster spilling onto the warrior, his shield and the wet ground, a wet sheen on his coat, hair and shield, and dark contact shadows under both figures. Photoreal feature-film VFX, large-format cinema camera, anamorphic lens, shallow depth of field, fine film grain, cold orange-teal grade, deep blacks.
Beeble Studio: relighting with real light controls
Beeble Studio is the desktop app. It runs on your own GPU, so files stay on your machine and no cloud credits are used. It uses the SwitchLight model to turn footage into PBR passes: surface normals, albedo, depth and alpha. With those passes you light the actor like a 3D object, using an HDRI of your background, area lights, a sun, or a video playing as a light source. Then you composite the result over the plate.
- Import the plate and generate the PBR passes and alpha.
- Load an HDRI or a still of your AI background as the environment light.
- Add a key light that matches the direction of light in your background, and a rim light for separation.
- Animate lights if the scene demands it, for example a flickering cyan light for the creature glow.
- Export the relit actor with alpha for compositing, or render over the background inside Studio.
10Pro
The pro pipeline: shoot, generate, composite
How professionals shoot, generate and finish AI VFX so it holds up on a big screen.
AI only fixes what it is given. The biggest quality jumps come from how you shoot and how you finish, not from the prompt.
Shoot for AI
- Light the actor like the final scene. If the shot will be backlit by a sunset, backlight the actor on set. Relighting a flat-lit actor is the hardest thing to fake.
- Lock the camera or add tracking markers. A static camera is the easiest to work with. If you move the camera, keep tracking markers on the green screen.
- Give the actor an eyeline. Put a tennis ball on a pole at the height of the creature's head, and plan the choreography so strikes go toward where the creature will be.
- Choreograph in beats. Count the moves ("block on 2, spin on 3, lunge on 4"). Clean, readable beats become clean timestamps in your prompt.
- Shoot short takes. Keep takes to 10 seconds or less, at one frame rate (24 or 25fps) for the whole project.
- Record a clean plate of the empty set, plus a few photos of the lighting setup for reference.
Generate
- Conform every clip to one frame rate and trim to the action.
- Build or relight the world first (Chapter 8 Pass 1, or Beeble SwitchX).
- Fix the lighting and grounding in a separate pass.
- Add characters and creatures last, with timestamped actions.
- Run two or three takes per pass and keep the best before moving on.
Composite and finish
Bring the AI result into After Effects, Nuke or DaVinci Resolve Fusion for the last ten percent that makes it look filmed:
- Light wrap and edge blend. Let the background light bleed a few pixels onto the actor's edges.
- Grain match. Add the same grain over the whole frame so AI areas and real areas feel like one piece of film.
- Contact shadows. Paint soft shadows under feet and at points of contact if the model missed them.
- Atmosphere layers. Add haze, smoke or dust between foreground, actor and background for depth.
- Black level match. The darkest parts of the actor and of the background should match.
- Camera shake on impacts. A small shake when the creature strikes the shield sells the weight.
- Sound. Impacts, roars and footsteps do half of the work of VFX. Never review a VFX shot on mute.
QC checklist before you deliver
- The face matches the reference in the first, middle and last frame
- Hands have the right number of fingers and props don't change shape
- No flicker in the background when you watch at full screen
- Light direction on the actor matches the background
- Every foot touches the ground with a shadow
- No leftover green on hair or edges
- Nothing passes through anything else
11Reference
Troubleshooting and prompt rules
Quick fixes for everything that went wrong while building this handbook.
The anatomy of a strong video edit prompt
- Keep: what must not change (face, performance, camera, timing).
- Change: exactly what to replace, numbered if there are several things.
- Action: for anything new that moves, a timestamped list of what it does each second.
- Light: where the light comes from, its colour, and how it falls on the actor.
- Look: camera, lens, grain and grade described in words, not film titles.
- Guardrails: one closing line of "no ..." instructions, since there is no negative field.
| Problem | Fix |
|---|---|
| Face changes between frames | Split the character sheet into separate face, front and back images, cut shots to 10s or less, and add "no identity drift, no blending of old and new faces". |
| New element on the wrong side | Use frame directions only and define them once. Place new things where the actor is not standing. |
| Creature or character is static | Give it an action for every stretch of time, and write "never freezes". |
| Action doesn't sync to the actor | Measure timestamps from the plate and write the timeline in seconds. |
| Actor looks pasted in | Run a separate relight pass that describes the light direction, a rim light, contact shadows and green-spill removal. |
| Flat background with no depth | Describe four layers: foreground, ground, midground and background, plus haze that gets thicker with distance. |
| Output is 16:9 when the source is wider | Set the aspect ratio in node settings before running. |
| Gemini Omni blocks the prompt | Remove combat words and film titles, and describe it as a staged fantasy scene. If it still blocks, test without the new element to see whether the real person in the footage is the trigger. |
| A pass changed something it shouldn't | Start the prompt with "Keep ... exactly as filmed" and add "Change nothing else". |
| Prompt too long for Lemonpeel | Cut adjectives before instructions. Keep, change, action and light always stay; style words go first. |
Handbook complete
You have the full AI VFX ladder: object swaps, new worlds, character replacement, trailer recreation, creature VFX and pro relighting. Pick one shot from your own footage and take it through Chapter 8 this week.