Skip to content

I Challenged a Pro VFX Artist to Beat AI

2026-08-07 · case-study tutorial · Adil, ErikDoesVFX · recording, 25:38 · watch on YouTube

Topic: Post-production & finishing (also: Video generation & shot direction, Assets, references & consistency, Prompting & AI assistants, Cost, credits & plans) · mode: Tutorial · tags: guest or collaboration, fiction, prompts shared, one creator's experience, time-bound offer

In this case study, Adil challenges the VFX artist ErikDoesVFX by remaking Erik's Iceland dragon shots with AI only. He makes character sheets on a plain grey background, using Nano Banana 2 for faces and Seedream 5.0 for the dragon. He then rebuilds the shots with Seedance 2.5 video-to-video from prompts Claude writes, and fixes scale with a size-fit frame. Then he makes a dragon-flight short film from scratch. He picks Soul Cinema locations by their light and describes motion and emotion in physical detail.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — One shot took a VFX artist days, the other an hour with AI: Adil challenges Erik

  • ★★ Adil challenges the VFX artist Erik (ErikDoesVFX) to make the same shots two ways: Erik with a full VFX pipeline of 3D tracking, simulation and compositing, Adil with AI only. — 00:00:00

00:00:40 — Video call with ErikDoesVFX: same shot, same brief, two pipelines

00:01:14 — Seedance 2.5 goes live on Higgsfield, with a launch offer

00:01:34 — Erik's VFX shots from Iceland and the AI game plan

  • ★★ Erik filmed shots around a cave in Iceland, including a drone shot towards the waterfall; Adil's task is to put himself in the cave, jump off the cliff onto a dragon and fly through the waterfall, with AI in one day. — 00:01:54
  • ★★ Adil estimates that a video like his AI dragon short film would cost studios at least $50,000 to make. — 00:01:54

00:02:30 — Step one: character sheet of Adil in Nano Banana 2 and the dragon sheet in Seedream 5.0

  • ★★★ To make his character sheet, Adil uploads a photo of himself taken against a wall to Claude and asks for a prompt: him on a grey background from three angles, full body front and back, one close-up, in hiking gear with an attached logo. — 00:02:30 · Claude, character sheet
  • ★★★ Adil advises always generating character sheets on a plain grey background; he says he burned through a lot of credits learning this. — 00:02:53 · character sheet
  • ★★★ In a Cinema Studio project Adil picks Nano Banana 2 for his character sheet, calling it the best model at the moment for matching faces. — 00:02:53 · Cinema Studio, Nano Banana 2, character sheet
  • ★★ Adil's workflow for any project: Claude writes the prompt and Higgsfield makes the content, with no complex software or heavy pipeline. — 00:03:22 · Claude, Higgsfield AI
  • ★★ Adil says the face on his Nano Banana 2 character sheet is a 100% match and the lighting is balanced, as if shot in a real studio. — 00:03:51 · Nano Banana 2, character sheet
  • ★★ Adil saves his character sheet as an asset named "Adil hiking" so that Claude automatically includes it in the prompt of any scene he is in, and advises doing this for every final asset. — 00:03:51 · Claude, character sheet
  • ★★★ For the character sheet of a huge, realistic black dragon with close-ups of the mouth open and closed, Adil switches from Nano Banana to Seedream 5.0, which he calls the best model at the moment for fantasy creatures. — 00:04:19 · Seedream 5.0, Nano Banana, Claude, character sheet
  • ★★ For a creature with a lot of detail, Adil generates two close-ups on purpose (mouth open and mouth closed), which he says keeps the dragon's face from glitching between expressions. — 00:04:52 · character sheet
  • ★ Adil says only pros with years of experience can build a 3D dragon, which has to be sculpted, retopologised, UV-unwrapped and rigged before it can be animated. — 00:04:52

00:05:21 — Shot one: replacing the actor with video-to-video

  • ★★ Adil says his Seedream 5.0 dragon character sheet took one minute and cost about 40 cents. — 00:05:21 · Seedream 5.0, character sheet
  • ★★★ To recreate Erik's shots Adil uses the video-to-video method in Seedance 2.5: upload the original clip and make the edits with a prompt. — 00:05:21 · Seedance 2.5
  • ★★ Adil cuts the source footage into segments and freeze-frames the last frame of any clip shorter than 4 seconds before using it for video-to-video. — 00:05:21
  • ★★★ Seedance 2.5 needs at least 4 seconds of video input for video-to-video. — 00:05:21 · Seedance 2.5
  • ★★★ For shot one Adil gives Claude Erik's clip and his character sheet and asks for a prompt that replaces the person in the video with the person in the image, keeping the video one to one. — 00:05:49 · Claude, character sheet
  • ★★ Adil writes Seedance 2.5 prompts with a Claude prompt skill that breaks a shot down into camera angles, lighting and movement, and offers the skill free in the description. — 00:05:49 · Claude, Seedance 2.5
  • ★★★ For video-to-video Adil sets the Seedance 2.5 duration to 4 seconds to match the reference clip, saying that with anything longer the model starts guessing. — 00:05:49 · Seedance 2.5 · also in: AI vs VFX: Can Seedance 2.5 Beat VFX?
  • ★ Adil says that in traditional VFX, compositing the actor into the scene and matching light, shadows and film grain can easily take hours, even for a pro. — 00:06:32
  • ★★ Adil says his video-to-video shot one took one photo and a single prompt, with no green screen, live shoot or VFX software, and the model matched lighting and contrast to the cave by itself. — 00:07:14 · Seedance 2.5

00:07:47 — Shot two: the cliff jump and dragon catch, and the CG double in Erik's version

  • ★★ Adil's video-to-video prompt for the cliff jump with a dragon catch failed in all four batches: with no actual jump moment in the source footage, the shot turned into slop. — 00:07:58 · AI look, Seedance 2.5
  • ★★★ Adil says that when the same problem shows up in all four batches, the issue is the prompt or the source material, and batching further only burns credits. — 00:07:58 · rebatching
  • ★★ When video-to-video fails, Adil rewrites the prompt with Claude from a screenshot of the location, runs it as image-to-video and stitches the two shots together in the edit. — 00:08:28 · image-to-video, Claude
  • ★★ Adil says AI saves the most time on shots that need a CG double, which takes a long, complex process to build and animate; in Erik's shot the sneakers change between frames at the swap to the CG double, and the fall gives it away. — 00:09:19
  • ★★ Adil says a person goes off a cliff either feet first or by pushing off into a dive, while Erik's digital double tips over, locks into a stiff pose and slides down at one speed, like an asset dragged in software. — 00:09:53 · also in: AI vs VFX: Can Seedance 2.5 Beat VFX?
  • ★★ For shot three Adil attaches the real footage to Claude and asks it to analyse the video and write the prompt that places the dragon in it. — 00:09:53 · Claude

00:10:22 — Shot three: the dragon bursting through the waterfall

  • ★★ Adil's first Seedance 2.5 take of the dragon shot looked epic but changed the weather and the camera movement, so he rejected it as slop: the shot has to match the original plate. — 00:10:41 · AI look, Seedance 2.5
  • ★★★ To fix the dragon's movement Adil does not write "make it more natural", which he says does nothing; he has Claude rewrite the prompt with a physical picture: normal speed, shaking off the water, moving like a snake in midair. — 00:10:41 · Claude
  • ★★ Adil points out a dragon wing that looks low-poly, like origami, with flat surfaces, sharp corners and straight folds instead of smooth organic muscle curves, and says it has to be high-poly. — 00:10:41 · AI look
  • ★★ Adil says the rewritten prompt for the dragon shot made a day-and-night difference and the whole clip turned out realistic. — 00:11:31
  • ★★ In Adil's first take of the dragon flying at the camera, the proportions between the dragon and him were off compared to the previous shot, and the camera flew far too slowly. — 00:11:57 · Seedance 2.5

00:12:05 — Shot four: the size-fit frame technique

  • ★★★ Adil's fix for scale is a size-fit frame: he uploads a screenshot of the previous shot, his character sheet and the dragon's sheet to Claude, asks for a detailed prompt and generates a new image asset that fixes his scale on the dragon. — 00:12:36 · Claude, character sheet

00:13:20 — Side by side: AI versus traditional VFX

00:13:54 — The $50,000 AI short film is shown

  • ★★ Adil's dragon short film was generated entirely with AI with no real footage, with the same pipeline: images first, then video. — 00:14:03 · image-to-video

00:15:03 — Outfit assets with GPT Image 2 and Nano Banana, cropped close-up and merged size sheet

  • ★★★ Adil says GPT Image 2 handles clothing best, which makes it his model for changing a character's outfit, useful for anyone with their own clothing brand. — 00:15:03 · GPT Image 2
  • ★★ In Adil's GPT Image 2 outfit generations one warped the logo and another turned the clothes bluish; he says you have to learn to spot slop. — 00:15:33 · GPT Image 2, AI look
  • ★★★ For a small fix such as a warped logo, Adil says no big prompt is needed: he switches to Nano Banana, which he says holds input best, and writes "change the logo to the one in image 2". — 00:15:33 · Nano Banana
  • ★★★ To keep the face from turning to slop in video generation, Adil crops the head out of the full-body frame of the character sheet, so the model has one place to take the face from: the close-up. — 00:16:04 · character sheet · also in: AI vs VFX: Can Seedance 2.5 Beat VFX?
  • ★★ Adil merges a frame from the dragon's sheet with his new character sheet in Nano Banana 2 and saves it as its own character sheet, so Seedance 2.5 does not change the dragon's size or his. — 00:16:04 · Nano Banana 2, Seedance 2.5, character sheet · also in: AI vs VFX: Can Seedance 2.5 Beat VFX?
  • ★★★ For locations Adil always uses Soul Cinema, which he says gives the most cinematic frames. — 00:16:28 · Soul Cinema
  • ★★ Adil says that for locations GPT Image always skews yellow and Nano Banana makes them too clean and too symmetrical. — 00:16:28 · GPT Image 2, Nano Banana

00:16:40 — Cinematic locations in Soul Cinema, chosen by light

  • ★★ For the waterfall exterior Adil attaches a screenshot of the first frame of Erik's video in Soul Cinema and runs it in batches. — 00:16:53 · Soul Cinema, rebatching
  • ★★★ Adil says one GPT Image generation costs 7 credits, while in Soul Cinema the same 7 credits give 56 variations, so locations can be batched as much as wanted. — 00:16:53 · Soul Cinema, GPT Image 2, credit
  • ★★★ Adil tests candidate location stills in video and says to watch the light when making locations: bad light is why video generations come out as slop. — 00:17:38 · AI look · also in: AI vs VFX: Can Seedance 2.5 Beat VFX?

00:18:18 — Scene one: drone camera and the waterfall reveal from the clouds

  • ★★ Adil says the opening scene must grab attention from the very first second: when his drone shot came out too static, he started scene one inside the clouds, with the camera punching out of them before the waterfall reveal. — 00:18:44

00:19:31 — Scene two: the cave location from a prompt and an emotional performance

  • ★★ To build the cave location from a prompt rather than a photo, Adil asks Claude for a similar location based on the photo reference (a cave inside a waterfall, same angle, heavier water flow) and runs it in Soul Cinema as text-to-image with no images attached. — 00:19:31 · Claude, Soul Cinema
  • ★★★ To show a location such as a room or a street from the opposite angle, Adil describes the new perspective to Claude, attaches the locked location asset and runs the prompt in Soul Cinema. — 00:20:02 · Claude, Soul Cinema
  • ★★★ To get real emotion from a character in Seedance 2.5, Adil describes the face in detail: what the character is going through, what he says, how his voice sounds and what his hands are doing. — 00:20:39 · Seedance 2.5

00:21:48 — Scene three: one-take cliff jump to dragon flight, with a consistent voice

  • ★★★ To keep a character's voice consistent across generations, Adil uses the character sheet: he says Seedance 2.5 locks the voice along with the appearance, so the voice need not be described each time. — 00:21:48 · character sheet, Seedance 2.5

00:22:49 — Scene four: starting mid-action, water spray physics

  • ★★ For the shot coming out of the waterfall Adil starts mid-action, with no approach or empty intro: the dragon punches through the waterfall in the very first second. — 00:22:52
  • ★★★ For high-speed shots Adil suggests prompt words such as a 180-degree camera orbit, speed shake and lens droplets for extra realism. — 00:23:30
  • ★★ Adil advises always specifying physical movement in the prompt to get dynamic camera passes. — 00:23:30
  • ★★ On high-speed shots Adil lets the model improvise background detail, as long as it does not break the story. — 00:23:30

00:23:49 — Scene five: high-speed flight and camera orbits

  • ★ Adil says clouds are a huge pain in traditional VFX because they are made of millions of volumetric particles. — 00:24:22

00:24:35 — Scene six: volumetric clouds at sunrise, and wrap-up

In the description and comments

  • ★★ Higgsfield: Asked whether a reference video is needed for such shots, Higgsfield answers that anything can be generated from scratch without one. — creator comment

Threads in this video

Where the stories that moved here went next

Mentioned

Central: character sheet · Claude · ErikDoesVFX · Seedance 2.5 · Soul Cinema · video-to-video

Also: Concepts: AI look · credit · image-to-video · rebatching · size-fit frame

Products: Cinema Studio · Elements · GPT Image 2 · Nano Banana · Nano Banana 2 · Seedream 5.0

Also centred on ErikDoesVFX: AI vs VFX: Can Seedance 2.5 Beat VFX? · Can AI Beat VFX? (feat. ⁨⁨ErikDoesVFX)

Also centred on video-to-video: AI vs VFX: Can Seedance 2.5 Beat VFX?

Replied to by

← Previous: 33 Days of Unlimited Seedance 2.5 on Higgsfield (2) · Next → AI vs VFX: Can Seedance 2.5 Beat VFX?