I Mixed AI With Real Footage… And It's Actually Scary
AI Samson demonstrates a hybrid AI VFX workflow that keeps real performance and camera movement as the foundation, then uses Gemini Omni Flash, reference images, GPT Image 2, and structured prompts to transform the shot.
Video credit: AI Samson. Source video: I Mixed AI With Real Footage… And it's Actually Scary.
Quick Summary
- AI Samson's workflow starts with a real video plate, then uses Gemini Omni Flash inside Higgsfield to relight the footage, replace environments, change clothing or objects, add generated characters, create VFX, and synchronize text with a recorded gesture.
- The strongest results come from preserving what already works. Keep the performance, timing, camera path, framing, and source identity explicit, then describe one controlled transformation and support it with a reference image when needed.
- This remains a generative process, not conventional pixel-locked compositing. Faces, hands, text, logos, motion, and continuity can drift, so useful production work still needs versioning, human review, disclosure, and a conventional finishing pass.
--what to try first
- Film a clean plate with the exact performance and camera movement you want before asking AI to redesign the shot.
- Begin with one edit, such as relighting, a background replacement, an outfit change, or one VFX object.
- Write a source-lock section that lists everything the model must preserve, including identity, timing, framing, camera movement, and untouched scene elements.
- Use a first-frame concept image when the new environment, character, clothing, or object is difficult to describe with text alone.
- Describe the relationship between each input. Tell the model what comes from the video, what comes from the image, and what the prompt should change.
- Generate several versions and compare continuity frame by frame. A persuasive thumbnail is not enough if the effect morphs during motion.
- Finish the accepted shot in an editor for pacing, sound, titles, color consistency, disclosure, and final quality control.
What changes when AI starts from real footage?
The important idea in AI Samson's tutorial is hybrid production. The model is not inventing the entire shot from an empty prompt. A real recording already provides the actor, performance, gesture timing, lens perspective, composition, and camera movement. AI becomes a transformation layer built on top of that foundation.
Google describes Gemini Omni Flash as a multimodal model that can combine text, images, and video for generation and conversational editing. That makes a source video more than visual inspiration. It can act as the motion and timing reference while a still image defines a new environment, outfit, object, character, or visual style.
This approach is useful because directing a real performance is often easier than generating one from scratch. A creator can mime surfing, swing a bottle like a sword, point to an empty area for a title, or act beside an imaginary character. The model then has clear physical beats to reinterpret.
The tool stack used in AI Samson's demonstration
The central model is Google Gemini Omni Flash, accessed through the Video section in Higgsfield. The creator uploads a base video, references it inside the prompt, and can add one or more images to guide the new look. Google also provides Omni Flash through its own Gemini and Flow products, plus a public-preview developer model.
For difficult visual changes, AI Samson extracts a first frame and edits it as a still before returning to video. He demonstrates GPT Image 2 inside Higgsfield for this concept-frame stage. OpenAI documents GPT Image 2 as an image generation and editing model that accepts text and image inputs, not as the video renderer in this workflow.
Claude is used as a prompt-development assistant. AI Samson built a reusable custom skill that turns a simple creative request into a more precise video-to-video prompt. Anthropic describes skills as reusable packages of instructions, workflows, context, and supporting resources that give Claude domain-specific expertise.
Higgsfield sponsored the source video. The examples should therefore be understood as AI Samson's creator demonstrations inside a sponsored tutorial, not as an independent benchmark against every available video model.
AI relighting and cinematic color treatment
The simplest experiment changes the look of existing footage without intentionally replacing the action. AI Samson applies golden-hour lighting, an ethereal treatment, and a neon atmosphere to real clips. He also uses a screenshot from another studio as a visual reference for the desired lighting and color direction.
This is more than a traditional color grade. A conventional grade remaps recorded color and exposure, while a generative relighting pass may redraw shadows, highlights, materials, skin, or background detail. That can create a stronger transformation, but it can also introduce changes that were never requested.
A safe test prompt should preserve the subject's identity, expression, movement, camera, composition, and environment, then specify only the lighting direction, color temperature, contrast, atmosphere, and reference image. Compare hands, face shape, clothing edges, and background geometry before approving the result.
Background, outfit, and object replacement
For a studio transformation, AI Samson takes a still from the source clip, uses an image model to design a stronger set, and selects the most useful concept frame. He then uploads both the original footage and the redesigned still to the video model, asking it to keep the performance from the video while applying the environment from the image.
The same input relationship can change clothing or a moving prop. In the tutorial, a shirt becomes a different outfit, a motorbike becomes a hover bike or alien creature, and a simple room becomes a branded or cinematic environment. The real motion remains the driving reference while the requested appearance changes.
Start with one replacement at a time. Clothing folds, object contact, occlusion, and fast hand movement are difficult continuity tests. When several objects, a new environment, a new outfit, and a new lighting scheme all change in one generation, it becomes harder to identify why a shot failed.
Design the real performance for the AI effect
The lightsaber example works because the original action was filmed with the intended effect in mind. AI Samson swings a bottle through the frame, giving the model a physical object, hand grip, arc, speed, and timing to reinterpret as a glowing weapon.
The same principle supports environment and character transformations. Pretending to surf gives the model body movement for a wave scene. Football movement can be transported into a stadium. A riding pose can drive an imaginary creature. The source plate functions like motion direction rather than finished imagery.
Plan the effect before recording. Leave visual space for the generated element, use an object with a similar grip or silhouette when practical, avoid unnecessary motion blur, and capture a few variations of the action. Better source choreography gives the model clearer evidence.
How the AI text-animation method works
AI Samson records a talking-head performance with an explicit hand gesture, then asks the model to place bold text beside him at the peak of that movement. The prompt preserves the performance, framing, background, lighting, and grade while describing the words, typography, placement, entrance, and synchronization point.
Google currently lists text and action synchronization as a Gemini Omni Flash capability. That makes the method attractive for short social videos where one gesture can motivate a title, label, impact word, or explanatory graphic.
Legible text across moving frames remains a demanding quality test. Review spelling, letter shape, tracking, baseline stability, motion timing, and whether the graphic collides with the presenter. For exact brand typography or legal copy, a conventional motion-graphics overlay is still more predictable.
A prompt structure that gives video edits a better chance
Begin with the operation: edit the supplied source video rather than create a new unrelated clip. Then write a source-lock block that lists the identity, performance, timing, camera path, framing, edit points, background elements, and lighting that must remain unchanged.
Describe the requested transformation next. Name the new object, outfit, setting, character, effect, or text treatment. Map important changes to visible action beats, explain scale and physical contact, and define how the new element should inherit the plate's light, grain, perspective, shadows, and motion blur.
Finish with forbidden changes and a review target. State what must not be reframed, retimed, replaced, duplicated, or distorted. A custom Claude skill can make this structure reusable, but longer language is not automatically better. Every instruction should help the model understand either what to preserve, what to change, or how to integrate the change.
When the text prompt is not enough, add a reference image. Give each input a single responsibility, such as motion from the video, environment from the first image, clothing from the second image, and the exact transformation from the text.
Generated characters, scene expansion, and transitions
To add a person beside the presenter, AI Samson exports the first frame, asks an image model to redesign the studio and add a new character, then uses that still as the visual target for the video transformation. The original performance anchors the presenter while the reference frame supplies the added character and environment.
A similar workflow can expand a close shot into a wider view or transform one real clip into another. These are generative continuity tasks. The system must infer off-camera space, preserve perspective, and bridge two different moments, so they require more scrutiny than a contained relighting pass.
Inspect contact, gaze direction, scale, shadows, facial stability, and whether the added character reacts consistently throughout the shot. Never present a generated person as a real participant, and obtain permission before transforming another person's likeness or performance.
A practical AI VFX production playbook
First, define one effect and record a clean plate for it. Lock exposure and focus when possible, leave enough room for the transformation, and capture a few takes with readable motion. Second, export a representative frame and create a concept still only if the new visual world needs stronger guidance.
Third, upload the source footage and reference stills to Gemini Omni Flash in Higgsfield or another supported surface. Write a prompt with source lock, requested transformation, action timing, physical integration, and forbidden changes. Generate several versions rather than treating the first output as final.
Fourth, compare the full motion, not only the opening frame. Track identity, anatomy, edges, object contact, background continuity, text, and camera behavior. Branch from the strongest result or simplify the request if several elements drift.
Finally, bring the accepted shot into a conventional editor. Trim the generated clip, rebuild exact titles when necessary, match sound and color across shots, add disclosure, and preserve the original plate plus prompts and references as production records.
Current limits, disclosure, and the filmmaker's role
At the time of this research, Google describes Gemini Omni Flash as a public-preview model with 10-second generations. Google also documents limitations around scene extension, uploaded audio references in the API, video-reference processing, and character consistency during scene or camera changes. Higgsfield currently lists 720p output and 10-second clips for its implementation. These specifications can change.
Google says Omni-generated video includes SynthID watermarking, but creators should still disclose material AI transformations when viewers could mistake a generated person, location, product, event, or endorsement for reality. Keep consent, trademark rights, source-image licenses, and brand approvals in the production checklist.
AI Samson's conclusion is that filmmakers who combine existing craft with AI have the strongest opportunity. That is also the practical takeaway from his examples. Performance direction, shot design, visual taste, prompt structure, continuity review, sound, editing, and ethical judgment remain human production responsibilities.
Copyable AI VFX Prompt Pack
These prompts are adapted from the pack AI Samson supplied with the video. Replace the example-specific subject, setting, timing, aspect ratio, and sound instructions with details from your own footage. Use only people, brands, characters, and source material you have permission to transform.
Prompt pack credit: AI Samson.
Golden Hour Relighting
Preserve the performance and camera, then change only the direction, softness, color, atmosphere, and grade of the light.
A young man sits on the wooden deck of a tropical villa, legs splayed out barefoot, wearing an oversized muted purple tee and dark shorts, gently petting a tan and white dog that rests across his lap. Behind him: sliding glass doors, a wooden bench, dense ferns and palms, a rattan lounger, a folded parasol, and a pool at right. Preserve identity, face, expression, the petting motion, the dog's rest and every movement, wardrobe, deck geometry, framing, lens, and the static locked-off camera exactly. Change only the lighting and grade, relighting the whole frame from soft overcast daylight to warm golden-hour light. Photoreal. 16:9. 5 seconds, matching the source. Warm golden-hour relight, soft honeyed grade, gentle low contrast, warm highlight rolloff, and soft glow. SFX only. Continuous single shot, same wide eye-level framing and the same static locked-off tripod as the source. The man keeps petting the dog exactly as in the plate. The dog stays relaxed and half asleep on his lap. Every motion and micro-expression remains identical. Trade the flat overcast daylight for warm late-afternoon sun. A soft golden key enters low from screen left, raking gently across the deck planks and lighting one side of his face and the dog's fur in warm honey tones. Soft-edged shadows fall long and gentle. A warm bounce lifts from the sunlit wood back into the shadows so nothing becomes harsh. The ferns and palms turn lush and glowing, backlit with a soft rim of sun through the leaves. Skin reads warm, healthy, and even. The dog's coat glows amber. A faint warm haze and gentle light bloom drift through the frame, with soft dust motes catching the low sun. Apply a low-contrast, airy, soothing golden grade over everything, with highlights rolling off softly and no crushed blacks. Face and identity remain unchanged. Petting, dog, wardrobe, framing, lens, and camera stay identical to the source. Only light color, direction, softness, and grade change. SFX only: soft warm room tone, gentle birdsong, distant leaves rustling, the quiet contented shift of the dog settling, and a faint calm ambient pad low in the mix.
Cold Neon Horror Relighting
Use the same plate as the golden-hour test, but push it toward a high-contrast cyan and magenta horror treatment.
A young man sits on the wooden deck of a tropical villa, legs splayed out barefoot, wearing an oversized muted purple tee and dark shorts, gently petting a tan and white dog that rests across his lap. Behind him: sliding glass doors, a wooden bench, dense ferns and palms, a rattan lounger, a folded parasol, and a pool at right. Preserve identity, face, expression, the petting motion, the dog's rest and every movement, wardrobe, deck geometry, framing, lens, and the static locked-off camera exactly. Change only the lighting and grade, relighting the whole frame from soft overcast daylight to cold neon horror light. Photoreal. 16:9. 5 seconds, matching the source. Cold neon horror relight, teal and crushed-black grade, high contrast, saturated cyan and magenta sources, and near-black shadows. SFX only. Continuous single shot, same wide eye-level framing and the same static locked-off tripod as the source. The man keeps petting the dog exactly as in the plate. The dog stays half asleep on his lap. Every motion and micro-expression remains identical. Kill the soft daylight and relight the whole scene in cold neon. A hard cyan key rakes in from screen left through the glass doors, throwing long hard-edged shadows across the deck planks. A sickly magenta accent bleeds from off-screen right and catches the pool so it throws cold rippling caustics onto the boards. Fill drops away so shadows crush to near black beneath the man and dog and pool in the open doorway. Cold rim light traces the man's shoulders and hair and the dog's fur. Skin turns pale and desaturated under the blue. Green foliage darkens to a toxic near black. Faint cold haze drifts through the frame with thin volumetric shafts angling through the doors. Apply a high-contrast crushed-black horror grade over everything. Face and identity remain unchanged. Petting, dog, wardrobe, framing, lens, and camera stay identical to the source. Only light color, direction, hardness, contrast, and grade change. SFX only: a low sub-bass drone under the whole shot, a faint electrical neon buzz and flicker, one distant hollow thud near the midpoint, and the soft rustle of the dog settling.
Change the Outfit
A compact prompt for replacing clothing while protecting the performance, scene, and existing lighting.
Keep the video unchanged. Preserve the man's performance, camera framing, background, lighting, and color grade exactly as they are. Change only the man's outfit. Replace it with a flowy, colorful silk-like shirt in a bold festival print, such as paisley, tie-dye, or an abstract pattern, using warm tones such as orange, magenta, turquoise, and gold. Style it as a relaxed bohemian fit, paired with layered beaded or turquoise necklaces. Keep the fabric glossy and lightweight, with natural draping, folds, shadows, and movement that follow the original body motion.
Brand-Safe Merchandise
This version replaces the named fashion label from the supplied example with a logo you own or are authorized to use.
Keep the video unchanged. Preserve the person's identity, performance, camera framing, background, lighting, shadows, perspective, and color grade exactly as they are. Replace only the current outfit with a classic fitted short-sleeve polo shirt in navy, white, or black. Add my supplied approved brand logo as small embroidery on the left chest. Keep the shirt fit natural to the person's body and pose. Match the original fabric movement, lighting, shadows, camera perspective, and motion blur. Do not invent additional logos, text, labels, or brand marks.
Futuristic Hover Bike
Transform the moving vehicle while keeping the rider, action, timing, and original scene composition intact.
Keep the original video exactly the same. Preserve the person, identity, performance, movement, timing, camera angle, framing, facial expression, environment, and composition. Change only the bike and enhance the lighting. Replace the bike with a floating, menacing futuristic racing vehicle. It hovers just above the ground with an aggressive aerodynamic silhouette, exposed mechanical details, glowing energy elements, subtle heat distortion, and powerful propulsion effects. The design feels dangerous, cinematic, original, and highly realistic while matching the rider's original position and movement perfectly. Do not copy a protected franchise vehicle or branded design. Add subtle cinematic lighting with stronger directional highlights, deeper shadows, atmospheric contrast, soft volumetric light, and a faint glow from the vehicle reflecting naturally onto the rider and surrounding environment. The final result feels intense, futuristic, photorealistic, and seamlessly integrated while keeping the original action unchanged.
Football Stadium Transformation
Preserve every touch on the ball while replacing both the garden environment and the outfit.
A young man with short brown hair, barefoot, does keepy-uppies with a football on a raised wooden garden deck. He wears a blue open short-sleeve shirt over a white tee and loose light-grey linen trousers. Behind him is a tropical villa garden with a concrete wall, tiled roof, closed patio umbrella, wicker loungers, and dense green foliage. Soft flat overcast daylight, near-shadowless. Locked wide full-body shot. Preserve his identity, face, posture, and every foot and knee touch on the ball exactly, plus framing, lens, and camera. Change only two things: replace the whole garden with a packed football stadium, and swap his wardrobe for a generic light-blue and white striped football kit. Photoreal. 16:9. 4 seconds. Bright soft near-shadowless daylight grade matching the source's flat overcast light. Non-IP generic light-blue and white striped kit, with no team crest, sponsor, or brand marks. SFX only. Continuous locked wide shot, exact same framing, lens, and full-body composition as the source. His identity, face, and every touch on the ball remain identical to the source. Keep the same juggling rhythm, ball path, barefoot contact, and timing. Nothing about his movement or the ball changes. Swap the world around him. The wooden deck becomes a green stadium pitch, and the garden becomes a large open-air football stadium packed with fans in tiered stands, banners and flags, distant floodlight towers, and a roof line. The crowd shifts and moves with real depth and parallax. Keep the stadium under the same soft, flat, bright overcast daylight so the pitch and stands match the existing key light on the performer. Replace only his wardrobe: the blue shirt and white tee become a light-blue and white vertically striped short-sleeve football jersey, and the linen trousers become plain white football shorts. His feet stay bare exactly as in the source to protect the footwork and ball contact. Light and grade the new jersey, shorts, pitch, and crowd in the same soft shadowless daylight, with matched softness and micro-contrast. Add faint stadium haze over the far stands so nothing appears pasted in, plus believable contact where his bare feet meet the grass. Face and identity remain unchanged. Body, footwork, ball motion, timing, framing, and camera stay identical to the source. Only the environment and kit change. SFX only: full stadium crowd ambience, a steady low roar with scattered cheers and a distant drum chant, plus soft clean ball touches synchronized with each juggle.
Surfing Environment and Motion Transfer
Use the recorded balancing performance to drive a surfboard and ocean scene without changing the body motion.
A young man with short dark hair, clean-shaven, in a loose open blue linen short-sleeve shirt over a white tee and light-grey linen trousers, barefoot, stands in a wide athletic crouch on a wicker sun lounger in a lush tropical garden. His knees are bent, arms extended, and he sways and shifts weight to keep his balance. Static locked wide shot, full body in frame, soft flat overcast daylight, cool low contrast. Preserve his body motion exactly, frame for frame. Every weight shift, knee bend, torso lean, and arm movement remains identical to the source, with the same timing and no added or altered motion. Preserve his identity, face, hair, wardrobe, and the locked static camera framing exactly. Replace only the environment. Swap the garden for open ocean and transform the surface beneath his feet into a generic plain surfboard with no brand or logo. Photoreal. 16:9. 4 seconds. Bright but soft overcast sea daylight, with a cool low-contrast grade matched to the source. SFX only. Continuous single shot, exact same static locked wide framing and full-body composition as the source. His body movement is a perfect copy of the source: the same balancing crouch, arm sway, weight shifts, and timing. Nothing is added, smoothed, or re-posed. Under his bare feet, the wicker lounger becomes a plain matte surfboard, and the world becomes open ocean. The board and water follow his existing motion rather than driving it. A glassy green wave wall rolls and curls beside and behind him. Water rushes beneath the board in a fast planing wake. Thin sheets of spray peel off both rails, with foam and spindrift trailing past camera. A hazy flat sea horizon and soft overcast sky sit behind him. Keep the original soft flat key on the face and body untouched. Light and grade only the water and sky to sit under that same cool overcast, with matched softness and atmospheric haze. Add wet reflections and contact spray where his feet meet the deck so he is grounded on the board. No cut-out edges, halos, or mismatched rims. His body motion, face, identity, hair, wardrobe, and performance remain identical to the source in every frame. Only the world around him changes. SFX only: the rush and hiss of a board planing over water, spray sheeting off the rails, the low rolling roar of the wave behind, steady wind over open ocean, and occasional foam slap.
Generic Plasma Energy Blade
Add an original energy blade to a practical prop and make its light react naturally with the performer and environment.
A man with short brown hair wears a draped blue shawl over one shoulder and grey shorts, barefoot on the dark tiled lip of a pool, with a lush tropical garden behind him. He performs a two-handed sword form with a short wooden hilt: side guard, draw to a center chamber, diagonal thrust, high-guard chamber, and forward two-handed point. Preserve his identity, face, hair, blue shawl, shorts, bare feet, every pose, and the exact timing of the form, plus the wide locked framing, lens, and soft flat daylight on him. Change only this: a glowing energy blade ignites from the top of the wooden hilt. Photoreal. 16:9. 5 seconds. Soft overcast tropical daylight, natural grade. Only the blade and its glow add new light. Non-IP generic plasma energy sword, not based on any brand or character. SFX only. Continuous locked wide shot, same framing and lens as the source, static camera. He performs the identical two-handed form with the same body mechanics and timing. On the first frame, a straight blade of blue plasma snaps from the top of the hilt to about one meter. It has a blinding white-hot core wrapped in saturated electric-blue glow with a soft bloom halo. The edges shimmer with a faint heat wobble. The blade stays rigidly locked to the hilt and tracks his hands frame for frame through the whole form. It smears a blue light trail through each fast arc and settles to a steady glow in the guards. Its light is directional and reactive. Cool blue flicker rakes across his face and neck, catches the folds of the shawl, throws a hard moving highlight on his forearms and chest, glints off the wet dark pool tile at his feet, and sweeps a travelling blue wash over the green leaves behind him as the blade passes. Add soft contact bloom where the tip nears the foliage. Use real light falloff and colored shadows grounded in the scene, never a flat pasted overlay. Face, identity, hair, shawl, shorts, bare feet, pose, and timing remain unchanged. Everything else stays identical to the source. Only the blade and its glow are added. SFX only: a sharp electric snap and hiss as the blade ignites on the first frame, then a steady low plasma hum that rises in pitch on each fast swing, an airy whoosh on every arc, and faint continuous electrical crackle underneath.
Medieval Knight Character Replacement
Replace the performer with an original medieval character while preserving every action beat and the existing camera.
Keep everything in this video unchanged. Preserve the exact camera movement, framing, shot timing, background environment, lighting, color grade, and all motion and action beats of the original performance. Replace only the character with an original medieval knight: a broad-shouldered person in weathered chainmail beneath a white surcoat with a simple faded red geometric emblem, steel gauntlets, and a great helm tucked back or worn open-faced to keep the expression readable. The knight carries a massive two-handed greatsword with a long leather-wrapped grip and simple cruciform crossguard, held in both hands. The knight performs the exact same movements, gestures, and timing as the original character. Integrate the greatsword naturally into the motion, with its weight visible in the swings and stance. Chainmail shifts and surcoat fabric reacts to each movement. Match the original character's position in frame, scale, eyeline, contact points, and silhouette throughout. Maintain realistic material detail: dulled steel with scratches and wear, dirt on the surcoat hem, and matte metal that catches the scene's existing light sources. Do not copy a named historical figure, protected character, film costume, crest, or franchise design.
Gesture-Synchronized Text Animation
Add a concise kinetic title at the peak of a hand gesture while keeping the text stable and readable.
Keep the video unchanged. Preserve the man's performance, identity, camera framing, background, lighting, and color grade exactly as they are. As the man gestures, a modern motion graphic animates into frame beside him, synchronized to the peak of his hand movement. The words "LIKE THIS" appear in a bold, clean sans-serif typeface, all caps, crisp white, with a subtle soft shadow separating them from the background. Animation behavior: the two words snap in with a fast kinetic pop, one word after the other, with slight overshoot and settle, similar to contemporary YouTube kinetic typography. The text then holds steady in screen space for the remainder of the shot, locked in position and never tracking or drifting with the camera. Placement: upper third of the frame in the negative space beside the man's head. Never cover his face or hands. Make the title large enough to read instantly but secondary to the subject. Style: sleek, minimal, contemporary motion design. Sharp edges, perfectly legible letterforms, no distortion, no extra text, no misspellings, and no watermarks.
Video Timestamps
FAQ
What does mixing AI with real footage mean?
It means using a real recording as the performance, timing, composition, and camera foundation, then applying generative edits such as relighting, background replacement, outfit changes, new objects, characters, text, or VFX.
Which tools does AI Samson use in the tutorial?
He uses Google Gemini Omni Flash inside Higgsfield for video-to-video transformations, GPT Image 2 for concept-frame image edits, and a custom Claude skill to expand simple ideas into structured VFX prompts.
How do you preserve the original performance?
Use a source-lock section that explicitly preserves identity, expression, body movement, timing, framing, camera path, edit points, and untouched background elements. Make one controlled change and verify the entire generated clip for drift.
Can Gemini Omni Flash replace a video background?
Yes, background transformation is a documented and demonstrated use case. A reference image can guide the new location, but results are generative and may alter the subject, lighting, edges, or camera geometry, so review is still required.
Can AI generate synchronized text animations?
Gemini Omni Flash can connect text or graphics to actions in the source video. Film a clear gesture, describe the exact words, placement, style, entrance, and synchronization point, then check every frame for spelling and typographic stability.
What are the current Gemini Omni Flash limitations?
Google currently documents 10-second generations plus limitations around scene extension, uploaded audio references in the API, video-reference processing, and consistency during scene or camera changes. Higgsfield currently lists 720p output for its implementation.
Where can I get AI Samson's AI VFX prompts?
AI Samson links a free AI VFX Prompt Pack from the source video's description. The direct signup page is included in the sources and credits section below.
Who created and sponsored the source video?
The tutorial was created by AI Samson. Higgsfield sponsored the episode, and the article identifies the examples as creator demonstrations within that sponsored context.
--sources and credits
- AI Samson: I Mixed AI With Real Footage… And it's Actually Scary
- AI Samson on YouTube
- AI Samson: Free AI VFX Prompt Pack
- Google: Introducing Gemini Omni
- Google: Gemini Omni Flash developer release and limitations
- Higgsfield: Gemini Omni Flash
- Higgsfield: Gemini Omni Flash VFX and video-editing guide
- OpenAI: GPT Image 2 model
- Anthropic: Agent Skills