Will AI Change Cinema Forever? · VFX vs Seedance 2.5
ErikDoesVFX films a dragon sequence in Iceland and produces two versions: one with 3D assets, animation, Houdini simulations, rendering and compositing; another with Seedance 2.5 on Higgsfield. The first better preserves vision and continuity. The second works quickly and surprises when it can reinvent everything. This guide breaks down the comparison, shows the hidden costs on both sides and proposes a hybrid workflow to decide shot by shot.
Video credit: ErikDoesVFX.Source video: Is AI About to Change Movies Forever?.
Quick Summary
- Credit: the experiment, filmed footage, VFX version and embedded video belong to ErikDoesVFX. The video credits Rowan for dragon animation, Luke for drone shots, Eye Candy for the Ymirak asset and the Higgsfield team for the AI version.
- Disclosure: the video is sponsored by Higgsfield. Erik claims to have negotiated editorial freedom, without scripts or brand talking points, and presents flaws in the AI result.
- The challenge uses the same concept, a man jumps on a dragon, crosses a waterfall and flies through a valley, but does not use the same method or the same degree of control on both sides.
- The traditional version starts with filming in Iceland, a 3D Ymirak model, specialized animation, procedural waterfall, collision and velocity field interaction, lighting, rendering and compositing.
- To reverse the drone's movement without making the water rise, Erik stabilizes the shot, isolates the waterfall, inverts only the necessary part and applies the camera movement again.
- First video-to-video attempt with Seedance 2.5 fails. The team switches to the first frame, generates motion with image-to-video, repeats prompts until they get useful parts, and stitches together shots in post-production.
- The first version of AI loses identity, scale, and anatomy of the cross-plane dragon. The set does not form as cohesive a sequence as traditional VFX, although some movements and visual replacements are impressive.
- A second version, created from scratch with prompts, is much better in acting, sound, and pacing. This reveals the central point: AI performs more when it can propose a solution than when it has to execute a vision that is already blocked in detail.
- The test does not demonstrate that the AI produces final VFX in seconds. Demonstrates a cycle with trials, references, mode switching, output selection and stitching. Generation time is not the same as production time.
- The practical conclusion is hybrid: film what needs identity and performance, preserve 3D and simulation where physics and continuity matter, and use AI for predictions, variations, substitutions, extensions and shots where interpretation is acceptable.
--what to try first
- Credit ErikDoesVFX for the expertise and Higgsfield for producing the Seedance 2.5 version.
- Register the business relationship. A sponsored test can be useful when the methodology and flaws remain visible.
- Compare the same editorial objective, not just the time of a render with the time of a generation.
- Define before the test what it means to win: fidelity to vision, realism, continuity, speed, cost, or creative surprise.
- Use real location when scale, light, and environmental interaction are essential.
- A production-ready 3D asset transfers control of anatomy, rig, materials and continuity to the team.
- Give complex animation to an expert when the creature's movement is the center of the shot.
- For stylized water, a procedural approach may be faster than a full physical simulation.
- Build interaction with colliders, speed fields, spray and mist, not just with a creature placed on the plate.
- Do light matching and contact interaction before trying to hide problems with color grading.
- When video-to-video destroys the shot, test image-to-video from an approved frame.
- The first frame resolves appearance and composition, but does not guarantee trajectory, identity or scale over time.
- Expect multiple generations and editorial selection. The right attempt is also part of the cost.
- Separate difficult shots into beats and stitch only on motivated cuts or moments of occlusion.
- Lock the scale of the dragon in relation to the sprite, waterfall, and frame width.
- Use your own reference for the creature and another for movement, without mixing the functions.
- AI can correct an inverted cascade by generating new movement, but this is a visual synthesis, not a physical recovery of the original plate.
- Consistency between shots remains more difficult than a single impressive moment.
- A free generation may seem better and, at the same time, move away from the achievement originally called for.
- Use AI as a contributor to variations, not as an automatic replacement for direction, animation or compositing.
- Store plates, masks, references, prompts, seeds or job IDs, versions and assembly decisions.
- Perform human review for identity, anatomy, physics, continuity, audio, and rights before publishing.
Credit, sponsorship and state of the evidence
The embedded video was published by ErikDoesVFX on August 7, 2026. VFX footage, experience, and master version is by Erik. The video credits Rowan for the complex dragon animation and Luke for the drone shots in Iceland.
The dragon used in the 3D version is Ymirak, created by Eye Candy. The official page describes a production asset with Maya and Blender versions, IK/FK rig, facial detail, muscle simulation options, UDIM textures and flight cycle cache.
The generative version was produced by Higgsfield with Seedance 2.5. The video is sponsored by the company. Erik says that he did not receive a script or talking points and that he formalized the freedom to criticize the result in the contract. This information is a statement from the creator, not an independent audit of the contract.
The analysis distinguishes between the creator's demonstration, official product claims and editorial recommendations. The search was updated on August 9, 2026.
The short answer: AI changes the process, but it still doesn't replace control
Yes, generative video already changes the way a small team can prototype creatures, environments, acting, sound and editing. In this test, Seedance 2.5 creates visually strong footage without building the dragon, waterfall, or 3D lighting.
No, the video does not prove that a final cinema shot went from hours to seconds. The Higgsfield team tests video-to-video, switches to image-to-video, extracts frames, repeats prompts, chooses results and puts together shots in post-production.
Traditional VFX wins when the question is which version best matches the filmed vision. The AI comes closer when given the freedom to write another achievement. The decisive difference is controllability, not just realism.
The challenge: the same idea, two production systems
The concept is ambitious: Erik runs to a cliff, jumps, lands on a dragon, crosses a waterfall and flies through a valley. The public is asked to decide between VFX and AI.
The two versions share an intent, but not an equivalent pipeline. The traditional side uses plates, tracking, rotoscoping, 3D assets, rigs, animation, simulation, lighting, rendering and compositing. The AI side switches between video-to-video, image-to-video, references, repeated generation and stitching.
A fair comparison needs to measure what each system received, how many attempts were made, how many people worked, how much human and computational time was spent and how close the result was to the brief.
Five criteria for evaluating VFX against AI
Realism measures whether the shot looks photographed. Continuity measures whether actor, creature, scale, lighting and setting remain coherent between frames and shots. Fidelity measures whether execution matches the vision and blocking orders.
Speed should include preparation, retries, waiting, culling, and post-production, not just the time shown by a spawn bar. Cost must include people, travel, assets, licenses, hardware, credits, failed versions and finishing.
The fifth criterion is creative value. A different version might be great, but it shouldn't receive the same loyalty score just because it surprises.
1. Filming a base that already looks like cinema
Erik travels to Iceland to find a waterfall that you can walk behind. The location provides scale, atmosphere, water, rock, diffuse light, and a valley for flight plans.
Real footage solves at once dozens of problems that would have to be generated or simulated: stone texture, spray, atmospheric depth, vegetation movement, reflections and drone parallax.
The decision also limits further work. The dragon, the digital double, the additional water and the color need to comply with the photograph that already exists.
2. Ymirak: anatomy and detail as infrastructure
The traditional version doesn't start sculpting a dragon from scratch. Uses Ymirak, a production-ready asset from Eye Candy, provided to production for the competition.
The asset page lists complete rig for Maya or Blender, IK/FK, facial controls, physical shader, versions with 24 texture UDIMs, 8K displacement and, in the Pro package, anatomy, flight-cycle and advanced ragdoll.
This detail matters because a good asset is not just about appearance. It is a persistent structure that preserves proportions, materials and articulations in each shot, even when the camera and animation change.
3. Specialized animation to sell weight and intent
Erik hands over the animation to Rowan, a collaborator on previous projects. The work includes the complex flight, the approach to the waterfall, the transport of the actor and a finale in which the creature passes close to the camera.
The advantage of the rig is the possibility of reviewing trajectories, timing, banking, wingbeats and distance without changing the dragon's identity. A correction is directed to a parameter or curve, not a complete resample.
This is the difference between explicit control and probabilistic control. Both can produce a good pose, but only one allows you to request exactly five frames earlier without risking another face, another anatomy or another scenario.
4. A procedural waterfall instead of full physical water
For additional interaction, Erik creates a sphere that emits particles, applies gravity, wind turbulence, and random resistance. It then uses local density to make clusters fall faster than single particles.
It is a result-oriented approach. It does not try to reproduce every property of water. Try to produce the silhouette, speed, spray and breakup necessary for the shot.
The Houdini documentation confirms that particles can receive forces such as gravity, wind and drag, and that velocity fields can drive other simulations. The exact video technique is an artist's construct, not a universal preset.
5. Fix the drone without making the water rise
The aerial shot has the opposite camera movement to that desired. Reversing the video resolves the drone's trajectory, but also makes the waterfall rise.
Erik's solution is to track and stabilize the plane, rotoscope the water, invert the necessary layer and then apply camera movement again. The plate maintains real detail and the intervention is localized.
This shot becomes a good test for AI because the objective is semantic, moving the camera forward while gravity remains correct, but traditional operations are spatial and separable.
6. Make the water react to the dragon
One creature overlapping the cascade is not enough. Erik uses the dragon as a collider, runs the simulation again and realizes that the effect is still not convincing.
The fix creates a smoke simulation from the dragon to achieve a fluid motion field. This field is linked to the cascade particles together with the collision, producing water that curves and surrounds the passing wings.
SideFX's Pyro documentation describes collision fields, velocity, density, and forces. The result of the video shows why interaction is more valuable than simple occlusion: the environment recognizes the creature's weight and displacement.
7. Light matching before rendering
Erik uses an image of a cloudy sky as a light source and creates a sphere with an opening to simulate the cave. The object does not appear in the render, but it blocks light and helps reproduce the real photograph.
The generated water is given a smoke shader to approximate the waterfall mist. The digital double uses the same lighting logic.
The lesson is simple: realism comes from relationships. The creature must share direction, smoothness, exposure, color, contact shadow, and atmosphere with the plate. More detail in the model does not compensate for incompatible light.
Why hours per frame don't mean hours of waiting
The video opens with the idea that Toy Story took around seven hours per frame and modern productions continue to spend many hours despite much faster computers. The general point is correct: ambition grows with the ability to calculate.
The official RenderMan evolution page explains that Toy Story's REYES system simulated effects that the hardware at the time could not calculate with ray tracing. Modern renderers use much richer global illumination and path tracing.
However, hours per frame usually represent normalized time on a machine or core, not necessarily the human wait on a render farm. Pixar itself explains, in the case of Coco, that 50 hours per frame would mean one hour with 50 colors. Compare core-hours, farm size, resolution, samples and complexity before drawing conclusions.
In Erik's project, the main render takes twelve hours according to the creator, including a new pass after increasing the spray. This is test data, not an industry average.
8. The shot is born in compositing
After rendering, Erik integrates dragon and water on the plate, removes its real presence, adds the digital double, mixes real and generated waterfall with masks and does color matching.
On the side, paint people, add a mist asset and improve integration. At the end, add camera shake when the dragon passes close to the lens.
The first green screen is not convincing because of the lighting and compositing. Erik decides to reshoot in a real cave. This choice is as important as any technique: when the practical element is accessible, capturing reality can be faster and better.
9. The first attempt at AI fails
The Higgsfield team starts with video-to-video, a prompt written with Claude and references from actor Adil and the dragon. The result clearly fails according to the video.
This exposes a common limit: video-to-video receives too many obligations at the same time. It has to preserve camera, timing, person and geography, but also introduce creature, contact, physics and a new relationship of scale.
When a generation fails, repeating the same order is not the only way. The team changes the representation of the problem.
10. First frame and image-to-video as fallback
The team extracts the first frame and uses image-to-video to reconstruct the movement from scratch. Achieves an acceptable jump, but the dragon's entry remains weak.
It then generates variations until it finds a useful version of the dragon and puts the parts together in post-production. The real workflow is no longer a prompt: it is frame selection, prompting, sampling, evaluation, cuts and assembly.
Image-to-video reduces the number of things the model must preserve. In exchange, it abandons the movement and performance of the original video. It is a good alternative when the composition matters more than the exact trajectory.
11. How AI solves the inverted waterfall
For the drone challenge, the team also uses the first frame and asks for the correct movement with image-to-video. The waterfall falls and the camera moves forward without the manual process of tracking and rotoscoping.
This is impressive as a result, but the operation is different. Traditional VFX corrects parts of the existing plate. The model generates a new plausible sequence from a frame.
If the obligation is to preserve pixel, environmental identity or documentary value, the synthesis may not be acceptable. If the goal is a plausible fantasy plan, it may be the most efficient option.
12. Video-to-video returns in flight shots
In the lateral and final shots, Higgsfield returns to video-to-video. The first outputs have problems, but the team continues until they choose versions that they consider usable.
In the cave plan, the model replaces Erik with Adil. The exchange demonstrates the power of transformation and, at the same time, breaks the identity of the original performer.
The process confirms that different shots require different modes. A single preset does not resolve creature input, flight, actuation, reverse motion, contact and identity with the same reliability.
The first result: strong moments, weak sequence
In the reaction, the AI version features a different actor, a dragon that drastically changes size, inconsistent proportions and a shot in which the creature looks more like a hawk.
Some details work. The fall has energy, the water reacts and the hair replacement is considered convincing. The problem is not the total absence of good frames, it is the lack of continuity between them.
The first comparison is won by traditional VFX because it maintains creature, scale, performance, light and intention throughout the entire sequence.
13. The surprise: a fully generated version
Higgsfield presents a third solution, created from scratch with prompts in Seedance 2.5, without trying to reproduce Erik's plates.
The initial acting, sound design, jump progression, and overall direction are better than in the first AI version. The result is more cohesive because image, movement, performance and audio are born within the same generative logic.
There are still inconsistencies and a recognizably synthetic look. However, the version is strong enough to change the tone of the discussion.
Why AI improves when given freedom
Reproducing a blocked view is a problem of constraints: person, clothing, camera, timing, waterfall, dragon, trajectory, scale, acting and editing must remain specific.
Generating from scratch is a problem of intention: courage before the jump, huge creature, passage through water, epic flight and cinematic sound. The model can choose the visual solution that best corresponds to these concepts.
Freedom increases perceived quality, but reduces direct authorship over every detail. The director chooses between good samples rather than building exactly the same solution.
Seven seconds is not the cost of production
The opening sentence contrasts hours of rendering with seconds of AI. It's a good provocation, but it measures different steps.
The 3D side pays upfront for assets, rigs, animation, simulation and lighting, then produces deterministic frames that can be reviewed. The generative side pays for reference preparation, prompts, queues, multiple attempts, selection, and less predictability of corrections.
Record human and computational time per shot, number of attempts, percentage of usable output and post-production minutes. Only in this way is it possible to compare total cost and not just interface latency.
What Higgsfield announces for Seedance 2.5
On the direct page verified on August 9, 2026, Higgsfield presents Seedance 2.5 as now live and offers a link to generate with the model.
The page advertises up to 30 seconds in a take, up to 50 multimodal references, output up to 4K, audio generated with the image and editing by region. These are official platform claims, not independent measurements of this article.
The page shared in the video redirects to the Seedance 2.5 landing with the campaign identification of ErikDoesVFX. Price, limits, queue, regional availability and actual options must be confirmed in the account before planning a production.
The hybrid workflow that this test suggests
Start with a shot list and classify each shot by identity, physical interaction, continuity, risk, and tolerance for interpretation. Do not choose the tool for the entire film before this screening.
Use real footage for performance and environments that are difficult to synthesize. Use 3D for hero creatures, recurring props, actions that need a precise trajectory and plans that will be revised. Use AI for predictions, concepts, extensions, substitutions, variations and isolated shots where surprise is welcome.
Combine the methods in compositing. A generation can provide background, mist or motion proposal; the 3D asset can preserve the creature; the plate can maintain face and body; the editing decides what the audience sees.
Choice of technique by type of shot
Close-up dialogue with an identifiable actor favors real plate, cleanliness and controlled composition. A recurring creature in multiple angles favors 3D asset. A brief, dreamlike or distant vision can favor generation.
Interaction with water, fire, destruction or tissue needs to assess contact. If the contact moves the story or will be seen up close, simulation or hybrid passes offer greater control. If it's secondary atmosphere, AI might be enough.
A transition plan can accept variation. A product shot, customer face or continuity of action normally cannot. Error tolerance is the most important variable in the decision.
Rights, identity and provenance
Use plates, faces, voices, 3D models, HDRIs, music, fonts and assets with permission. Ymirak's page features different licensing options and packages; Please confirm permitted use at the time of purchase.
Replacing one person with another is technically simple and editorially sensitive. Obtain consent from performers, inform clients about transformations and save the origin of references.
For professional work, maintain a manifest with file, owner, license, date, tool, template, prompt, job ID and edit performed. The final output must undergo human and legal review appropriate to the territory and use.
Verdict: The future isn't VFX or AI, it's direction over both
The traditional version meets Erik's challenge because it executes the original vision with consistency, interaction, and scale. The first version of AI brings together good moments, but does not keep them as a sequence.
The fully generated version proves something else: Seedance 2.5 can create a convincing little spectacle when the direction is expressed as intent and the model chooses the realization.
The change in cinema will not just be cheaper frames. It will be a new distribution of decisions. Good directors will know when to ask for precision, when to accept interpretation and how to combine actors, assets, simulation, generation and editing without losing authorship.
Prompt kit · Hybrid workflow for creatures and real footage
The transcript does not provide Higgsfield's full prompts. The templates below are original and transform the test lessons into a reusable system. Use only authorized footage, identities, voices and assets. Replace the fields in braces and adjust the syntax to the available template.
Prompt pack credit: ErikDoesVFX.
01 · Brief and definition of victory
Decide what the shot should preserve before choosing the tool.
Analyze the shot described below and create a comparable brief for traditional, video-to-video and image-to-video VFX.
SHOT
{DESCRIPTION, DURATION, CAMERA, ACTION AND CONTEXT}
LOCKED ELEMENTS
{IDENTITY, CLOTHING, LOCATION, PROP, CREATURE, TIMING, DIALOGUE, MOVEMENT AND FINAL FRAME}
TOLERANCE TO INTERPRETATION
{LOW, MEDIUM OR HIGH}, because {REASON}.
DEFINE
1. Five measurable success criteria.
2. Elements that may vary.
3. Faults that require regeneration or new rendering.
4. Best primary technique and an alternative.
5. Required inputs and licenses.
6. Comparison plan with human, computational time, attempts and post-production.02 · Capture plan for VFX and AI
Film a base that serves both workflows.
Creates a shooting plan for the {EFFECT} effect in {LOCATION}.
DELIVER
Main plate with full action.
Clean plate without actor.
Composition reference frame.
Passage with scale marker.
Close-up of actor with neutral expression and final expression.
360-degree lighting reference or exposure bracket.
Footage of texture, atmosphere and environment.
Clean ambient sound.
FOR EACH TAKE
Indicates duration, lens, height, camera movement, shutter, light direction, eyeline, starting and ending position of the actor, scale reference and margin for the creature.
Includes a safety checklist and a non-physical stunt alternative.03 · Reference manifesto
Prevent identity, movement and style from conflicting.
REFERENCE MAP
@source_video exclusively controls the camera, timing, blocking and actor movement.
@actor_ref exclusively controls {ACTOR}'s face, hair, body and clothing.
@creature_ref exclusively controls {CREATURE}'s anatomy, scale, color, skin, wings, and silhouette.
@location_ref exclusively controls geography, rock, vegetation, waterfall, light and atmosphere.
@motion_ref only controls flight mechanics and wing rhythm.
@audio_ref only controls environment, impact, water and authorized voice.
PRIORITY
Identity > anatomy > scale > timing > geography > light > style.
Do not transfer people, clothing, scenery, text or cuts between references. If there is a conflict, follow the priority and maintain one function per input.04 · Video-to-video with source lock
Preserve the plate and limit the transformation to the necessary effect.
VIDEO-TO-VIDEO ABOUT @source_video
SOURCE LOCK
Preserve 1:1 the face and body of {ACTOR}, clothing, performance, timing, camera, lens, framing, speed, cuts, scenery, waterfall, light and duration.
ONE-TIME CHANGE
Add {CREATURE OR EFFECT} based on @creature_ref. The creature starts at {POSITION}, takes {ACTION}, and ends at {POSITION}. Do not replace the actor or rebuild the environment.
SCALE
{CREATURE} measures {RELATION TO ACTOR, CASCADE OR FRAME}. The proportion of the head, torso, wings, limbs and tail remains identical in all frames.
INTEGRATION
Same cloudy light, contrast, temperature, atmospheric depth, motion blur, grain and plate focus. Correct contact shadow, spray and occlusion.
PROHIBITED
New actor, variable face, growing or shrinking creature, changeable anatomy, new camera, zoom, extra cuts, inverted waterfall, different clothing, text, music or watermark.05 Image-to-video fallback
Reconstruct motion when preserving the entire video is too difficult.
IMAGE-TO-VIDEO FROM @start_frame
FRAME LOCK
Preserve composition, actor, clothing, face, location, light, waterfall and scale of the first frame.
TIMELINE | {DURATION}
00:00-{T1}: {ACTOR} starts {ACTION}.
{T1}-{T2}: {CREATURE} enters from {DIRECTION}, maintaining the defined scale.
{T2}-{T3}: exact contact between {ACTOR} and {CREATURE}.
{T3}-FINAL: both follow {TRAJECTORY} and end in {FINAL FRAME}.
CAMERA
{SIMPLE AND CONTINUOUS MOVEMENT}. No cropping, unexpected zooming or axis shifting.
PHYSICS AND AUDIO
Realistic weight, inertia, wind, water and spray react to passing. SFX only: {LIST}. No dialogue unless explicitly stated.
Avoid teleportation, morphing, duplication, weightless contact, variable scale, rising waterfall and different character.06 · Scale lock of the creature
Define measurable relationships rather than just saying big.
CREATURE SCALE LOCK, TOP PRIORITY
@creature_ref is the only anatomical reference.
The head measures {X}% of the length of the torso.
The wingspan measures {N} times the actor's height.
When placed, the eye line becomes the actor's {RELATION}.
In the open plane, the creature takes up {PERCENTAGE}% of the frame width.
In the close-up, the snout measures approximately {RATIS} the width of the actor's shoulders.
Maintain exactly the same head, neck, torso, wings, paws and tail relationship in all frames and shots. A single solid animal, with constant mass.
Prohibited: giant head, thin body, disappearing limbs, variable wings, different species, bird, falcon, colossal creature, miniature, growth, shrinking, ballooning or morphing.07 · Cascade and gravity
Lock water direction and flight reaction.
WATERFALL PHYSICS LOCK
The waterfall always flows vertically downwards under gravity, with continuous speed. Never rises, freezes or reverses.
The camera moves {DIRECTION AND SPEED} independently of the movement of the water.
When {CREATURE} passes through the water in {TIMECODE}, the wings shift spray to {DIRECTIONS}, open a short cavity in the water veil, and create turbulent mist in the wake. The water closes after the passage.
Maintain volume, origin, width and falling point of the waterfall. The spray inherits the light, motion blur and depth of the plate.
No solid wave, gel, fabric, water passing through the creature, weightless splashes or new waterfall.08 · Identity and performance
Protect the performer in close-up and during transformations.
IDENTITY LOCK FOR {ACTOR}
@source_video controls performance and timing.
@actor_ref controls identity.
Preserve {ACTOR}'s face, age, hair, skin, eyes, proportions, body, clothing and expression in all frames.
PERFORMANCE
{TIMECODE}: {GESTURE, LOOK, BREATHING AND EMOTION}.
Maintain microexpressions and eye direction from the source video. Mouth closed unless specified speech.
CONTACT
Hands and feet maintain correct anatomy and contact with {SURFACE OR CREATURE}. Don't hide your face to resolve your identity.
Do not substitute someone else, change hairstyle, rejuvenate, alter body, create facial mask, duplicate performer or add dialogue.09 · Pass generation plan
Stop ordering everything in one sample.
Splits the {DESCRIPTION} shot into generative passes with safe cutoff points.
PASS TO
Objective: performance and approach.
Input: {INPUTS}.
Mandatory final frame: {FRAME}.
PASS B
Objective: entry and contact of the creature.
Input: last approved frame from A + @creature_ref.
Mandatory final frame: {FRAME}.
PASS C
Objective: flight and departure.
Input: last approved frame from B + @motion_ref.
Mandatory final frame: {FRAME}.
For each pass, define duration, locks, a main action, acceptance criteria and maximum attempts. Places cuts in occlusion, whip, spray, passage next to the camera or motivated change of plane. Don't hide a continuity gap with random cuts.10 · Hybrid compositing brief
Combine plate, 3D, simulation and AI by function.
Creates a compositing plan for {SHOT}.
REAL PLATE
Preserves actor, chamber, rock and main waterfall.
3D
Uses {ASSET} for anatomy, animation, shadow, reflections and position passes, normal, depth and motion vector.
SIMULATION
Uses {HOUDINI/PYRO/PARTICLES} for spray, mist, collision and air displacement.
AI
Use only for {EXTENSION, CLEANUP, BACKGROUND, TEXTURE, PREVIS OR VARIATION}. Does not change identity or locked geometry.
COMPOSITE
Order of layers, masks, holdouts, deep or Z, light wrap, contact shadow, atmospheric perspective, grain, motion blur, color match and camera shake.
DELIVERY
Pass list, responsible person, resolution, color space, frame range, naming and QC checklist.11 · VFX vs AI Scorecard
Evaluate without favoring the most spectacular moment.
Compares VERSION_VFX and VERSION_AI using the same brief. SCORE 0-10 1. Fidelity to the vision. 2. Actor's identity. 3. Anatomy and scale of the creature. 4. Physics and contact. 5. Light and integration. 6. Continuity between shots. 7. Acting and emotion. 8. Chamber and rhythm. 9. Audio. 10. Global realism. REGISTER Human time, computational time, number of attempts, direct cost, number of artists, post-production minutes and percentage of usable material. For each score, include timecode and observable evidence. If the version changes the realization, it separates aesthetic quality from fidelity to the brief. It ends with a winner by criteria, not just an outright winner.
12 · Final audit of continuity and rights
Transform the chosen cut into a list of corrections and proofs.
Audit the final cut shot by shot. FORMAT TIMECODE | CATEGORY | EXPECTED | OBSERVED | SEVERITY | CORRECTION | RESPONSIBLE CATEGORIES Face, body, clothing, creature, anatomy, scale, eyeline, camera, geography, waterfall, gravity, contact, spray, shadow, light, grain, flicker, audio, cut and text. RIGHTS For each plate, face, voice, 3D asset, HDRI, music, SFX, model and output, confirm owner, license, consent, supplier, date and restriction of use. CLASSIFICATION Accept, correct in composite, edit audio, use region edit, regenerate pass, revive 3D, reshoot or delete. It ends with an export checklist, credits, sponsorship disclosure, provenance and human approval.
Video Timestamps
FAQ
Who created Is AI About to Change Movies Forever?
ErikDoesVFX. The embedded video, experience, and traditional version are credited to the creator.
Is the video sponsored?
Yes. Higgsfield sponsors the video and produces the AI version. Erik claims to have maintained editorial control and shows flaws in the result.
Which AI model was used?
According to the description and the video, the entire generative side uses Seedance 2.5 inside Higgsfield.
Which 3D dragon was used?
Ymirak from Eye Candy. It is a production-ready asset available in Maya and Blender versions with different packages.
Who animated the dragon?
The video credits Rowan, Erik's collaborator on previous projects, for the complex flight and interaction animation.
Where was the sequence filmed?
In Iceland, in a waterfall where it was possible to walk behind it, and in a valley filmed by drone. The final close-up was reshot in a real cave.
Why create a waterfall if a real one already existed?
To generate interaction when the dragon crosses the water. The simulated version can collide, curve and produce spray around the creature.
Was a full physical water simulation used?
No. Erik builds an approximation with particles, gravity, wind, resistance, local density and a near mist shader.
How was the drone plane inverted without inverting the water?
Erik stabilizes the plane, isolates the waterfall with rotoscoping, inverts the necessary layer and reapplies the camera movement.
How does the dragon affect the waterfall?
The dragon works as a collider. A motion field derived from a smoke simulation is attached to the particles to create displacement and spray.
How long did the traditional render take?
Erik reports about twelve hours for his render, with an additional pass after increasing water interaction.
Toy Story took seven hours per frame?
It is the number used in the video hook. The safest data is that Toy Story used RenderMan REYES and that times per frame vary by shot, hardware and form of normalization.
Do hours per frame mean the team waits that long?
Not always. Render farms divide frames and samples across many machines. Pixar explains that per-frame metrics can represent equivalent time on a single machine or core.
Did the first video-to-video attempt work?
No. According to the video, it clearly failed. The team switched to image-to-video from the first frame.
Why did image-to-video work better?
Because I needed to preserve only the initial composition, not the entire movement of the plate. The counterpart was to generate a new performance.
Has AI solved the inverted waterfall?
Yes, visually. It generated new movement from the first frame with water falling and camera moving forward. It did not correct the original pixels like the traditional technique.
How many attempts did it take?
The video does not present a total. Shows failed attempts, new generations, mode change and result selection before stitching.
Which version won?
In the first comparison, the traditional version wins due to consistency, quality and fidelity to the vision. The video's audience is invited to vote on their own.
What failed in the first version of AI?
Identity of the actor, scale and anatomy of the dragon, continuity between shots and the final flyby close to the camera.
What did AI do better?
A few movements, the water reaction and a hair replacement. In the version generated from scratch, acting, sound, pacing and general direction are greatly improved.
Why is the version created from scratch better?
Because Seedance can choose image, performance, camera and audio as a coherent system, without exactly reproducing the existing plates and blocking.
Seedance 2.5 generates 4K video?
The Higgsfield page advertises output up to 4K. Confirm the option, cost and actual availability in your account and inspect the original file.
How many references does Seedance 2.5 support?
Higgsfield announces up to 50 multimodal references per generation.
Seedance 2.5 generates audio?
The official Higgsfield page states that audio and image can be generated on the same pass. The video highlights the sound design of the fully generated version.
Did the AI create the result in seven seconds?
Not as a complete final production. The experiment includes prompting, references, multiple trials, mode switching, selection, and post-production editing.
What is the best workflow for a small team?
A hybrid workflow: footage for performance, 3D for recurring creatures, simulation for contact and AI for predictions, variations, extensions and interpretation-tolerant shots.
Can I use the Ymirak model commercially?
The page presents packages and license options. Confirm the terms of the chosen product at the time of purchase and keep proof.
Can I replace an actor with another AI person?
Only with consent and appropriate rights. Document the authorization, purpose and origin of references.
--sources and credits
- ErikDoesVFX · Original video on YouTube
- ErikDoesVFX · Official Channel
- Higgsfield · Seedance 2.5 project associated with video
- Higgsfield · Seedance 2.5
- Higgsfield AI · YouTube Channel
- Eye Candy · Ymirak, Dragon 3D Model
- VFX Store · Assets mentioned in the video
- SideFX · Pyro effects, collisions and velocity fields
- SideFX · Forces in Houdini
- Pixar RenderMan · REYES evolution for path tracing
- Pixar RenderMan · How to interpret render times per frame
- McKinsey · The impact of AI on the future of film and television
- ErikDoesVFX · Instagram
- ErikDoesVFX · TikTok