Skip to content

How to Make Ultra Realistic AI Videos (28 Best Tips)

2026-05-16 · case-study tutorial · Adil, Joma Jurabaev, Higgsfield · recording, 28:37 · watch on YouTube

Topic: Productions (case studies) (also: Prompting & AI assistants, Assets, references & consistency, Pre-production & story, Studio work & business, Post-production & finishing, Cost, credits & plans, Video generation & shot direction) · mode: Tutorial · tags: one creator's experience, series, guest or collaboration, prompts shared

In this diary episode about making a feature for Cannes, Adil reviews the day-8 draft of Hell Grind, which had to be cut from 110 to 90 minutes, and shows three failed scenes fixed by day 11. Most of the video is the team's tips. They include a system prompt for prompt adherence, which he calls the most important, practical light only, an anchor in every location, and a slop director who checks every generation. It closes on the totals: 108,859 generations and about $400,000 in generation costs over 14 days, and half a million dollars for the whole film.

What is said

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

00:00:00 — Intro: a feature film for Cannes in less than a month

00:00:22 — Day 8: draft review and three failed scenes

  • ★★★ On day 8 of 14 the Hell Grind team had the full draft of the feature, as planned, but Adil judged it not good enough: 110 minutes of footage had to be cut to 90. — 00:00:22 · Hell Grind
  • ★★ Forty scenes of the Hell Grind draft still needed fixing; Adil follows three of them through to day 11. — 00:01:11 · Hell Grind
  • ★★ In the first failed Hell Grind scene, Jax and Rain look too plasticky with too much light on their faces, their reactions feel off and the portal looks cheap. — 00:01:38 · AI look, Hell Grind
  • ★★ In the third failed scene the skateboard is wrong: its first-episode draft was still on the canvas, and the person generating the scene grabbed that wrong asset. — 00:03:45 · Canvas

00:04:12 — Day 11: the three scenes fixed

  • ★★ During three days of fixing scenes, up to day 11, the Hell Grind team almost ran out of credits but did not. — 00:04:12 · Hell Grind
  • ★★ In the fixed first scene the green lines on the characters' backs are gone, the lighting makes sense and the characters look less plastic. — 00:05:02 · AI look

00:07:07 — Chapter 1, prompting: emotional scenes, prompting like a photographer, a system prompt for adherence

  • ★★★ For emotional scenes Adil gives the video model the situation, what the character feels, and the body's reaction: arm extended, fingers spread, jaw dropped, jaw clenching, eye twitching, a tear starting. — 00:07:20

    “provide situational context and anatomical behavior”

  • ★★★ When generating images, Adil says to prompt like a photographer: lens, aperture, grain, film stock, halation, bokeh, and how the scene is lit. — 00:08:35

  • ★★ Adil says words like "photorealistic, cinematic, high quality" at the end of an image prompt do nothing for the image. — 00:09:03
  • ★★ Without camera knowledge, Adil says to ask Claude to suggest cameras and lenses, or to give Claude a still from a favourite film and ask which camera and lens were used. — 00:09:25 · Claude
  • ★★★ Adil calls a system prompt for prompt adherence the most important tip; the 15-person Hell Grind team spent a full week polishing this one skill. — 00:10:15 · Claude, Hell Grind
  • ★★ Adil says he will leave a link to the Hell Grind prompt adherence skill in the description. — 00:10:15 · Hell Grind
  • ★★ The Hell Grind prompt adherence skill sets one main idea, one main action and one main camera strategy per shot. — 00:10:42 · Hell Grind
  • ★★ The skill also has rules for safe language, on-screen text and references, and mode selection rules for text-to-video, image-to-video, reference-to-video and video-to-video that set the output settings. — 00:11:09 · text-to-video, image-to-video
  • ★★★ The skill's frame coordinate system splits a location into X and Y axes and tells the video model where to place characters or objects, for example 20 on X and 30 on Y; Adil says the model does it. — 00:11:37

00:12:05 — Concept artist Joma Jurabaev on Episode 1 of Hell Grind

  • ★★★ Joma Jurabaev: Concept artist Joma Jurabaev says Hell Grind Episode 1 hooked him with its story, characters and conflict and did not feel like an AI movie, though some places could be improved. — 00:13:06 · Hell Grind
  • ★★★ Joma Jurabaev says that more than the tools, the talent, vision and storytelling of the people behind Hell Grind show; beautiful pictures without a story are hard to watch. — 00:13:36 · Hell Grind
  • ★★ Joma Jurabaev says that until now he could make his ideas only as static concept art, and with these tools artists can bring their stories to life. — 00:14:24

00:15:16 — Prompting continued: iterate, single shot vs multi-shot, phonetic dialogue, shot-list updates, sanitising long prompts

  • ★★★ Adil says the first batch almost never works: look at what landed and what did not, adjust the prompt and run again. — 00:15:16 · rebatching
  • ★★ If the first five seconds of a 15-second clip work, Adil tells the prompt to change only the last ten seconds and runs it again. — 00:15:39
  • ★★★ Adil uses multi-shot for action and dialogue scenes that should be dynamic, and single shot for nearly everything else. — 00:15:39 · multi-shot prompt
  • ★★ A single shot is one continuous motion of camera and characters, but Adil still defines what happens at each second for the most control. — 00:16:07
  • ★★★ One multi-shot generation can hold several cuts: the four reactions to the monster were one generation, which kept lighting and positions consistent instead of four two-second clips. — 00:16:36 · multi-shot prompt
  • ★★★ For dialogue, Adil writes complex words phonetically, spelling out how they sound, instead of the word itself. — 00:17:01
  • ★★ Adil says the more detail a video prompt gives (when to cut, which angle, who to cut to), the less room the model has to hallucinate the wrong motion. — 00:17:01
  • ★★ When one moment of a clip works, Adil screenshots it and gives it to Claude as the reference for positioning and lighting, so the next prompts stay consistent. — 00:17:29 · Claude
  • ★★★ With a long shot list, Adil tells Claude to update only what changed, so it rewrites only the current prompt; he says this cuts tokens by 80%. — 00:17:58 · Claude
  • ★★★ Too many iterations overload a prompt with conflicting parts; when it grows too long, Adil tells Claude to optimise and sanitise it, dropping what is not needed. — 00:18:25 · Claude

00:18:44 — Chapter 2, pre-production: 60-30-10 colour, practical light, establishing shots, location anchors, angles and views

  • ★★ For image colours Adil uses the 60-30-10 rule: 60% dominant colour, 30% secondary, 10% accent, which he calls the most visually appealing split. — 00:18:52
  • ★★★ Light from an unseen extra source makes characters look pasted into the set and plasticky; Adil prompts "practical light only": sun, windows and lamps in the frame. — 00:19:17 · AI look
  • ★★★ Adil starts dialogue scenes with a wide or medium establishing shot, so the model keeps every character in the right position in later cuts. — 00:19:17
  • ★★★ Every location needs an anchor, such as a tree or a stack of books, so characters and the camera can be placed relative to it ("place Roko to the left from the tree"). — 00:20:36
  • ★★★ Adil says Seedance 2.0 has a hard time with depth when a location image is taken head-on, so locations should never be generated from the front. — 00:21:17 · Seedance 2.0
  • ★★ Adil generates location images at three quarters or from a high corner like a CCTV camera. — 00:21:17
  • ★★★ Seedance 2.0 kept grabbing the wrong side of a location shown front and back, so Adil now makes a separate image per view, attaches both and names the view for each cut. — 00:21:45 · Seedance 2.0

00:22:13 — Draw the scene first, beat-by-beat action, face crops, prop sheets, tracking character states

  • ★★★ Adil draws the scene before prompting: Claude makes a top-down map of where the characters stand, which he iterates and turns into a prompt prefix for Seedance. — 00:22:13 · Claude
  • ★★★ Action scenes need a beat-by-beat script: who is in each cut, where they are, which weapon they use, and the choreography, which Claude can write. — 00:22:42 · Claude
  • ★★★ When the wide shot in a character sheet gives a plasticky face that does not look like the character (Rain), Adil cuts that face out of the sheet. — 00:22:42 · character sheet
  • ★★★ For tricky props Adil makes prop sheets: Seedance 2.0 kept making the monster hold its claw like a sword until a sheet showed the hand from outside and inside. — 00:23:26 · Seedance 2.0
  • ★★★ For long films Adil tracks every character's state (injuries, clothing, props, what is held or missing), called script supervising, and locks it before generating; Roko has a sheet per stage. — 00:23:55
  • ★★★ Adil splits a big group into teams of at least two, because alone you stop seeing slop, and appoints no single overseeing director, who would have no time to generate. — 00:24:31

00:24:44 — Chapter 3, teamwork: teams of two, a slop director, a canvas for assets

  • ★★★ Adil advises a "slop director" whose job is to catch AI slop before it ships; on Hell Grind it was Sever (as heard), who checked every generation before it went into the timeline. — 00:25:13 · AI look, Hell Grind
  • ★★★ Adil says to store all assets on a Canvas, so every member of a big team grabs the correct characters and locations in each state. — 00:25:13 · Canvas
  • ★★ Canvas has an assistant, Hixie (as heard), that can be told to optimise the canvas and style it. — 00:25:48 · Canvas

00:26:14 — Chapter 4, editing: frame-by-frame review, duplicated frames, usable seconds, regenerating later

  • ★★ Adil reviews generated clips frame by frame, since one bad frame can kill a whole shot. — 00:26:14
  • ★★★ Though the style prefix asks for 24 fps, some Seedance 2.0 clips stutter at 18 fps or lower; Adil deletes the duplicated frames one by one. — 00:26:14 · Seedance 2.0
  • ★★★ Adil says a failed generation is no waste: one or two seconds of it may be perfect, so cut those out and use them. — 00:26:14
  • ★★ When stuck on a scene, Adil moves on, puts it in the timeline, and returns later with a fresh eye to tweak the prompt. — 00:26:58

00:27:20 — Day 15: the film is locked, the numbers and the final cost

  • ★★★ Adil says the two-week sprint is over and the Hell Grind feature is locked. — 00:27:20 · Hell Grind
  • ★★★ Over 14 days the Hell Grind feature took 108,859 generations (about 7,700 a day, 320 an hour) and 9,540,147 credits, just under the budget, about $400,000 in generation costs. — 00:27:20 · Hell Grind
  • ★★★ Adil puts the total cost of the 90-minute Hell Grind feature, with team, audio and post-production, at half a million dollars, against about $50 million for a comparable live-action VFX film: 1%. — 00:28:06 · Hell Grind

In the description and comments

  • ★★ Higgsfield shares the Hell Grind project canvas publicly. — video description · Canvas, Hell Grind
  • ★★ Higgsfield: Told that the link to the prompt system was missing, Higgsfield says it has fixed the link. — creator comment

Threads in this video

Where the stories that moved here went next

Mentioned

Central: Canvas · Claude · Hell Grind · Hell Grind prompt adherence skill · Seedance 2.0

Also: People: Joma Jurabaev

Works: The Mandalorian

Concepts: AI look · character sheet · image-to-video · multi-shot prompt · rebatching · script supervising · style prefix · text-to-video

Events: Cannes Film Market

Characters: Jax · Lulu · Rain · Roko

Also centred on Hell Grind prompt adherence skill: Hell Grind: Rebuilding Our Biggest Fight Scene from Scratch

Replies to

← Previous: Higgsfield MCP + Claude Just Changed Marketing Forever! · Next → I Created 10+ Beauty Ads in 1 Hour — No Agency Needed