Skip to content

How to Make AI Films That Don't Look AI (Full Tutorial)

2026-05-09 · long tutorial / working session · Adil, Patrick Kalyn, Higgsfield · recording, 3 h 6 min · watch on YouTube

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

On day 4 of a 14-day push to finish Hell Grind, an 80-minute AI feature for Cannes, Adil builds scenes 21 and 23 and shows every failed take. He starts with a Claude Shotlist builder skill and a style prefix that keeps 15 people consistent, then iterates location and prop stills, because the video takes its lighting and look from the images. The Seedance 2.0 shots are fixed by reordering references, running batches of eight and rebuilding prompts from scratch. The day ends at 4,441,352 credits, about $260,000.

What is said

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

00:00:00 — Day 4 of the 14-day push to finish an 80-minute AI feature for Cannes; what this episode covers

00:01:34 — Script of scenes 21 to 23: Roko back in the old apartment after Lulu is taken

00:02:54 — Shotlist builder skill in Claude, the style prefix and its audio rule, the skill's phases

  • ★★★ To add the Shotlist builder skill to Claude: Customize, create new skills, the plus button, upload a skill, drop in the zip; the directors built it from what they learned on the series. — 00:02:54 · Claude
  • ★★★ A style prefix kept on the Canvas defines style, lighting, colour, composition and audio, so 15 people generating one film stay consistent. — 00:03:21 · Canvas
  • ★★ The lighting line of the Hell Grind style prefix: "natural light only, key light coming from sky and windows only". — 00:03:50 · Hell Grind
  • ★★★ Seedance 2.0 gives only one audio track; with speaking characters, especially in emotional moments, it tends to add music, which is very hard to separate in post-production. — 00:03:50 · Seedance 2.0, native audio
  • ★★★ The style prefix tells Seedance 2.0 to add only environmental sound effects, no music and no subtitles; the music is added when the film is edited. — 00:04:19 · Seedance 2.0, native audio
  • ★★ The Shotlist builder skill splits scenes into 15-second shots with prompts tailored to the video model; after the script, its second phase asks for the characters, locations and props it needs. — 00:04:47 · Claude

00:05:15 — Describing the old apartment to Claude with the city location sheet; Soul Cinema as an alternative

  • ★★ Without a specific look in mind, Adil recommends Soul Cinema: type something like "an old apartment", batch many images and pick the ones you like. — 00:06:59 · Soul Cinema
  • ★★ Nano Banana Pro has a hard time with spatial awareness: given location images it tries to fit them into the picture, so Adil gives the city reference to Claude instead. — 00:06:59 · Nano Banana Pro, Claude

00:08:03 — Collab projects: sharing, permissions, comments and calls on the Hell Grind project

  • ★★ A Collab project can be made public or shared by email or link, with permissions for approval, viewer or collaborator access, visibility in Community and viewer messages. — 00:08:32
  • ★★ Sharing a Collab project instead of downloaded images keeps every asset, input, prompt and the tool used, so the team can go back and tweak any generation. — 00:08:54 · Hell Grind

00:09:35 — Iterating the apartment in Nano Banana Pro and GPT Image 2: layout, plastic look, edits or a fresh prompt, reverse view, combined sheet

  • ★★★ The location image must be right, because most of the lighting and style of the later videos comes from that reference; a weak one makes the video sloppy. — 00:09:35 · location reference sheet
  • ★★ Plasticky textures are very common in Nano Banana Pro images; Adil fixes them through the prompt. — 00:10:06 · Nano Banana Pro, AI look
  • ★★★ Lighting keywords such as "light atmospheric haze" at the end of the prompt change lighting and textures and fight the plasticky look Nano Banana Pro always gives. — 00:11:53 · Nano Banana Pro, AI look
  • ★★ With a detailed prompt Nano Banana Pro gives fairly consistent results, so Adil batches only four images: if something is off, the problem is probably the prompt. — 00:12:18 · Nano Banana Pro, rebatching
  • ★★ For image edits Adil runs both Nano Banana Pro and GPT Image 2, which he says does pretty well with edits. — 00:20:41 · GPT Image 2, Nano Banana Pro
  • ★★ Adil avoids editing the same image again and again, since small details get lost each time; he edits the previous image instead. — 00:22:54
  • ★★ After many successive edits, Nano Banana Pro has a hard time understanding the apartment's spatial layout. (hedged) — 00:24:07 · Nano Banana Pro
  • ★★★ Instead of iterating edits on one image, Adil takes the best version to Claude and asks for a full prompt to regenerate it from scratch, with no reference image. — 00:24:55 · Claude
  • ★★ The final layout of the old-apartment location took Adil 44 iterations. — 00:29:06 · Hell Grind
  • ★★★ Adil combines the front wide view and the reverse view from the hallway into one 16:9 reference image, top and bottom, instead of uploading two images every time. — 00:31:32 · location reference sheet

00:32:01 — VFX artist Patrick Kalyn reacts to Hell Grind and estimates its traditional cost

  • ★★ Patrick Kalyn calls Hell Grind incredible: at times he forgot it was AI-generated, which the animation world calls "the illusion of life", and he felt the emotion. — 00:32:01 · Hell Grind

    “coming from the animation world we call that the illusion of life”

  • ★★★ Patrick Kalyn estimates that Hell Grind, made with a traditional shoot and visual-effects pipeline where almost every other shot is a VFX shot, would cost about 15 to 20 million US dollars. (hedged) — 00:32:42 · Hell Grind

00:33:11 — Props: the Polaroid, the fridge note and a photo wall stitched from single Polaroids

  • ★★★ Nano Banana Pro invented new characters despite the attached character sheets, likely because Claude's prompt described their appearance; the fix is to have Claude describe only the clothing. (hedged) — 00:35:35 · Nano Banana Pro, character sheet, Claude
  • ★★ Adil expects that a whole photo wall generated as one Nano Banana Pro image would drift the characters' faces even with a detailed prompt. (hedged) — 00:41:49 · Nano Banana Pro
  • ★★★ Testing it, Adil finds the one-image photo wall of 12 Polaroids unacceptable for the big screen and generates one Polaroid at a time, then stitches them together. — 00:43:44 · Nano Banana Pro
  • ★★ When the location image will be attached, Adil tells Claude not to describe the apartment in every prompt. — 00:44:47 · Claude
  • ★★ Adil generates each Polaroid of the photo wall from its own Claude prompt and stitches them together in Photoshop. — 00:45:13 · Photoshop, Claude

00:45:38 — Claude maps prompts onto the shot list; reference order, batches of eight, Chinese prompts, failure modes in prompts

  • ★★ Adil recommends editing as you go, with music, to see what is still missing and how the shots work together. — 00:46:36
  • ★★ The Hell Grind shots are generated at 21:9 instead of 16:9 because the film will be shown on big screens. — 00:48:06 · Hell Grind
  • ★★★ Reference order matters in Seedance 2.0: Roko's character sheet goes first because he is described first in the prompt, the apartment second. — 00:48:34 · Seedance 2.0, character sheet
  • ★★★ Adil runs Seedance 2.0 in batches of eight; when several takes make the same mistake, he stops them and fixes the prompt in Claude. — 00:48:34 · Seedance 2.0, Claude, rebatching
  • ★★★ The prompts are in Chinese because at launch Seedance 2.0 worked better with Chinese and capped prompts at 3000 characters; an English word takes five to ten characters, Chinese one to two. — 00:49:03 · Seedance 2.0
  • ★★ Failure-mode warnings such as telling Seedance 2.0 not to drift a small Polaroid work better inside the prompt itself. — 00:50:24 · Seedance 2.0

00:51:37 — Shot 21A: Roko at the door, reviewing takes and editing the prompt

  • ★★ With many takes, Adil skims each video and watches it in full only when the composition looks right. — 00:51:37
  • ★★ Adil suspects Seedance 2.0 struggles with spatial awareness when a character enters a room from outside. (hedged) — 00:52:11 · Seedance 2.0
  • ★★ When all eight takes fail, something is wrong in the prompt and there is no point batching further. — 00:52:46
  • ★★ Claude updates an edited prompt right inside the shot list, which keeps track of every prompt of a big project for later tweaks. — 00:54:02 · Claude
  • ★★ The Hell Grind plan: the first week generates the full 80 minutes, the second week refines key moments. — 00:54:26 · Hell Grind

00:57:57 — Day 2: a 100,000-credit scene scrapped after a script change

  • ★★★ On day two Adil spent nearly 100,000 credits on a two-and-a-half-minute scene, then a script change to a childhood flashback with new location and dialogue scrapped all of it. — 00:57:57 · Hell Grind

00:58:20 — Shot 21A continued: camera closer to the door, a longer boot shot

  • ★★ With the shot list, characters and locations are uploaded once and always tagged in the same order (image one Roko, image two the location), so nothing needs re-tagging. — 01:01:42

01:02:29 — Viewer comments: cost against a traditional shoot, the iteration behind earlier edits

  • ★★★ Adil says a director estimated a traditional shoot at a minimum of 5 million, while it cost the team 70 000: still expensive, but much cheaper than traditional filmmaking. — 01:02:55 · Hell Grind
  • ★★ Adil says his earlier videos showed first-try results thanks to editing; in reality you iterate, change prompts and learn which parts of a prompt to edit. — 01:03:21

01:03:46 — Picking 21A, an L-cut, and shot 21B from behind Roko

  • ★★ In an emotional moment the camera should not overload the viewer: Adil drops the handheld shake and rise and keeps one motion, a push-in. — 01:04:37
  • ★★ When the prompt's motion and action are right and the takes are only slightly off, Adil runs more batches of the same prompt. (hedged) — 01:06:22 · rebatching
  • ★★ Adil plans an L-cut: cut the picture before the door closes and keep the door-closing sound under the next shot. — 01:08:14
  • ★★ For continuity, since the previous shot faced Roko, the next shot places the camera behind him so the two shots do not clash. — 01:09:32
  • ★★ After Adil describes where the hallway is on the reference sheet, Claude adds a spatial layout block describing the room's layout and where the camera should be. — 01:13:52 · Claude

01:15:21 — Viewer comments: why 14 days and 15 people

  • ★★ The film is made in 14 days because someone on the team had the idea a week earlier to make a feature and take it to Cannes, which is why 15 people work on it. — 01:16:08 · Hell Grind
  • ★★★ Fifteen people is very few for traditional filmmaking, but AI production usually works in teams of three to four at most, so Adil calls this a big experiment for the industry. — 01:16:08

01:16:36 — Shot 21B as one continuous shot; a wide establishing shot of the room

  • ★★★ The spatial layout block in the Seedance 2.0 prompt, describing the room and the camera position, put Roko in the right place in the set. — 01:16:36 · Seedance 2.0
  • ★★★ Defining what happens in each few seconds of a prompt forces Seedance 2.0 to cut; for one continuous shot, describe it as one continuous shot. — 01:17:44 · Seedance 2.0

01:31:03 — The team; camera placement of shot 21C over 72 generations; post-production

  • ★★ The 15 Hell Grind team members who generate are professional DOPs and directors; besides them the team has screenwriters, sound designers and editors. — 01:31:03 · Hell Grind
  • ★★ Adil jokes that 90% of AI filmmaking is spending 15 batches fixing one little issue. (overstated) — 01:36:07
  • ★★ Asked to move the camera closer to the fridge, Seedance 2.0 moved the fridge closer to the camera instead. — 01:36:07 · Seedance 2.0

  • ★★★ For characters, Adil describes not the exact actions but the emotions they should feel while doing them (heartbroken, nostalgic). — 01:39:07 · Seedance 2.0

    “I'm not describing the exact actions, but rather the emotions that they should be feeling when they perform those actions.”

  • ★★★ When fixes keep getting added to one prompt, its parts start to clash; Adil has Claude scrap the whole prompt and start from scratch. (hedged) — 01:39:57 · Claude

  • ★★ Instead of anchoring the camera to props at the side of the set, Adil anchors it to the dark hallway, facing it very close. — 01:45:15 · Seedance 2.0
  • ★★ Post-production of an AI film is much like traditional: grade the clips and add missing sound effects; the voices come from the generated footage itself. — 01:46:57
  • ★★ Adil says failed takes are not waste: each one teaches him the prompt for the next. — 01:47:30
  • ★★ Shot 21C, about ten seconds of footage, took Adil 72 generations. — 01:53:24 · Hell Grind

01:54:02 — Shot 21D: a fridge keyframe, five separate references, an NSFW error

  • ★★ An image added after the prompt was pasted is sometimes not tagged as a reference; pasting the prompt again fixes it. — 01:54:24 · Higgsfield AI
  • ★★★ Images are much cheaper, so batch many stills: Seedance 2.0 carries a sloppy look or off lighting from the image into the video, and perfect images save video credits. — 01:59:34 · Seedance 2.0, keyframe
  • ★★ Adil uploads Roko, the location, the fridge, the photo wall and the Polaroid as five separate references so everything stays consistent in an extreme close-up. — 02:00:08 · Seedance 2.0
  • ★★ When Seedance 2.0 batches fail with an NSFW error and all images passed the IP check, Adil assumes the prompt is the cause and has Claude change it. (hedged) — 02:02:21 · Claude, Seedance 2.0
  • ★★ With the changed prompt the NSFW error was gone and the Seedance 2.0 generations worked. — 02:02:50 · Seedance 2.0
  • ★★ Failed generations on Higgsfield are not charged any credits. — 02:02:50 · Higgsfield AI

02:03:42 — Low-frame-rate takes and cutting doubled frames in the edit

  • ★★★ Seedance 2.0 sometimes delivers a lower frame rate, maybe 12 fps, than the 24 fps the prompt asks for, so each frame looks cut or delayed. — 02:03:42 · Seedance 2.0
  • ★★ Against the low frame rate, Adil has Claude add "the video must be 24 fps" to the prompt, keeping the style prefix at the beginning; later takes still lost frames. — 02:04:22 · Seedance 2.0, Claude
  • ★★★ Only when a take is great and its frame rate is the only problem: go through it frame by frame in the editor and cut out every doubled frame. — 02:06:37
  • ★★ Cutting doubled frames also cuts the audio track, so the whole track is lost and the fixed clip suits B-roll. — 02:07:09

02:08:57 — Shot 21D: Roko taking the right Polaroid; editing the fridge image; rebuilding the prompt

  • ★★ Seedance 2.0 could not work out where a Polaroid missing from the fridge keyframe belonged, so Adil adds it to the fridge image himself. — 02:16:24 · Seedance 2.0
  • ★★★ When a prompt grows very long it overwhelms the model; Adil scraps the whole prompt and starts over. (hedged) — 02:23:12 · Claude

02:31:59 — Shot 21E and an extra medium shot; scene 21 cut together

  • ★★ Adil saves both good takes of a shot and sends them to the lead editor, who picks one in the final cut. — 02:31:59 · Hell Grind

02:48:32 — Scene 23: Roko breaks down by the fridge; acting through emotions

  • ★★★ Adil's workflow for almost every shot: the prompts are too long to read and predict, even in English, so he runs a couple of batches to test a prompt and goes from there. — 02:49:17 · Seedance 2.0
  • ★★ An attached fridge reference image may make Seedance 2.0 hold that same frame for the first shot, so Adil removes it. (hedged) — 02:51:43 · Seedance 2.0
  • ★★ Seedance 2.0 does well with emotions when simply told "Roko is very upset", and small details such as "his lower lip trembles" can be added. — 02:52:21 · Seedance 2.0
  • ★★ Rather than cramming a second angle into one prompt, Adil writes a separate prompt for the second shot. — 02:55:43
  • ★★ A Seedance 2.0 take added background music although the prompt specifically asked for none. — 02:56:44 · Seedance 2.0
  • ★★ After the breakdown the camera pulls back, a common cinematography technique of leaving the character alone to feel the emotion. — 02:59:14

03:04:55 — Day 4 totals: credits, generations and cost; the next episode

  • ★★★ In four days the Hell Grind team spent 4,441,352 credits and generated 48,336 images and videos, about $260,000. — 03:04:55 · Hell Grind
  • ★★ Adil is beginning to think the 10 million credits may not be enough for the Hell Grind feature. (hedged) — 03:05:20 · Hell Grind
  • ★★★ During the shoot Adil generated nearly 800 assets, and only 8 made the final cut. — 03:05:20 · Hell Grind
  • ★★ Adil says that by episode 3 of the series the first draft of the full film will be finished. — 03:05:20 · Hell Grind

In the description and comments

  • ★★★ Higgsfield: Adil's day-4 session on the 80-minute AI feature for Cannes cost $6,139 in one sitting: 800 generations, of which 8 shots made the cut. — video description · Hell Grind
  • ★★ Higgsfield says the keyframe matters more than the prompt. — video description · keyframe
  • ★ Higgsfield: The Shotlist builder skill file cannot be viewed through its Dropbox link, but it can be downloaded. — creator comment
  • ★ Higgsfield: Asked to keep posting long videos, Higgsfield says more long ones are coming. — creator comment

Threads in this video

Where the stories that moved here went next

Mentioned

Central: Claude · Hell Grind · Nano Banana Pro · Seedance 2.0 · Shotlist builder · style prefix

Also: Concepts: AI look · character sheet · keyframe · L-cut · location reference sheet · multi-shot prompt · rebatching

Events: Cannes Film Market

Characters: Jax · Lulu · Rain · Roko

Products: Canvas · Collab · GPT Image 2 · Photoshop · Soul Cinema

Taking part: Patrick Kalyn

Also centred on style prefix: 3-Step Workflow To Make Ultra-Realistic AI Ads

Replies to

Replied to by

← Previous: How To Generate AI Videos Straight From Claude with Higgsfield's MCP · Next → Higgsfield MCP + Claude Just Changed Marketing Forever!