10:38
AI filmmaking
The video argues that the single biggest lever for making an AI-generated short film look 'Hollywood-level' is not the generation models but upfront planning — locking in a full story, characters, and every visual asset before generating a single scene — and it walks through a four-step workflow (plan with Claude, generate locked assets, generate scenes referencing those assets, generate voiceover and assemble) built entirely inside the Higgsfield platform's Cinema Studio 3.5.
Planning out the entire movie before generating anything is presented as the most important step, because most AI films are made 'on the go' with no plan, leading to inconsistent characters and random dialogue.
Step 1: paste a large pre-written prompt into a new Claude conversation, filling in only a short story brief, to have Claude act as 'creative director' and output the full story (title, setting, tone), character profiles (wardrobe, personality), an asset list, and a generation plan.
Step 2: generate every character, prop, and location as a locked asset in Higgsfield's Cinema Studio 3.5 before generating any scene, using GPT Image 2 at 2K resolution and 16:9 aspect ratio.
Characters are built as a single image with a three-panel structure — full-body front, full-body back, and a close-up of the face — so the later video model can keep the character consistent from any angle.
For a character meant to resemble the creator, two extra reference images (a face photo and the suit) are uploaded alongside the character prompt.
Locations (judge's bench, defense bench, prosecution bench) and props (briefcase, suit) are generated the same way as characters.
Step 3: in Cinema Studio's video generation section, select the video model (referred to in the transcript as 'Cidan's 2.0' / 'SeedAndSpark,' i.e. Seedance 2.0) at 1080p and 16:9, and generate each scene by referencing the exact locked character/location images in the prompt for consistency.
Scene prompts come in two formats: a natural-language paragraph prompt, and a more controllable JSON prompt generated by pasting the paragraph prompt into a free third-party website (signup via email) that converts casual language into JSON.
Scenes are generated one at a time by repeating: grab the prepared prompt, paste it in, click generate.
Step 4: in Higgsfield's audio tab, select 11 Labs (ElevenLabs) V3 as the voice engine, pick a voice matching the scene (a confident male news-anchor voice for the opening montage), and paste in the voiceover script that Claude wrote in step one.
The voiceover is downloaded as an MP3 and the final film is assembled in CapCut by dropping the clips in the correct order and adding the voiceover, with no extra editing needed because Cinema Studio 3.5 already handles per-scene polish.
The video's closing pitch is that Higgsfield is presented as the only platform that runs the entire pipeline — images, video, and audio — without needing another subscription or tool.
Four-step AI film production process — The video's overall framework for making a cinematic AI short film: plan the movie, generate locked assets, generate scenes referencing those assets, then generate voiceover and assemble the final cut. Apply: Complete the steps in order — plan the full story and asset list first, lock every character/prop/location image before any scene generation, generate scenes referencing those images, then add voiceover and assemble — rather than generating scenes on the go.
Claude as 'creative director' planning prompt — A large pre-written prompt pasted into a new Claude conversation that makes Claude output a full story (title, setting, tone), character profiles with wardrobe/personality, an asset list, and a generation plan. Apply: Fill in only the story brief (a couple of sentences describing the concept) at the top of the prompt template and send it to Claude to receive the full plan before generating any images or video.
Three-panel character sheet — A single generated image per character containing three views — full-body front, full-body back, and a close-up of the face — used to lock the character's look. Apply: Generate one image per character using this three-panel structure so the video model has consistent front, back, and facial detail to draw on for any scene angle.
GPT Image 2 — An OpenAI image model, selected in Higgsfield's image section, described as built for ultra-detailed, realistic renders suited to cinematic assets. Apply: Select GPT Image 2 with 2K resolution and 16:9 aspect ratio for generating all character, prop, and location assets before any video generation.
Cinema Studio 3.5 — A Higgsfield tool described as tailor-made for creating films with AI, housing image generation, video generation, and audio in one place. Apply: Use it as the single workspace for the entire film pipeline — assets, scenes, and voiceover — instead of switching between separate platforms.
Image-tag referencing for scene consistency — A method where a scene's video prompt references the exact locked character/location image assets generated earlier, so the video model pulls those visuals into the scene. Apply: When writing a scene prompt, reference the saved character/location images directly (e.g., a specific character in a specific location) so the video model renders that scene with consistent appearances.
Paragraph-to-JSON prompt conversion — A two-format approach to scene prompting: a natural-language paragraph prompt, optionally converted into a more detailed JSON prompt for tighter control over the scene. Apply: Use a free website (sign up with an email address) to convert casual paragraph prompts into JSON prompts before pasting them into the video generation tool for every scene.
Seedance 2.0 (video model, referred to in the transcript as 'Cidan's 2.0'/'SeedAndSpark') — A video generation model in Cinema Studio's video section, described as very good with camera angles, detail, and overall video quality for movie-style content. Apply: Select this model with 1080p resolution and 16:9 aspect ratio, then generate each scene using a prompt that references the locked character/location image assets.
11 Labs (ElevenLabs) V3 — A voice generation engine available in Higgsfield's audio tab, described as producing the most natural-sounding AI voices currently available. Apply: Select ElevenLabs V3, choose a voice matching the scene's tone (e.g., a confident male news-anchor voice), and paste in the voiceover script to generate the audio.
CapCut final assembly — Video editing software used to assemble the finished film by ordering the generated clips and adding the downloaded voiceover track. Apply: Drop the generated scene clips into CapCut in the correct order, add the voiceover MP3, and export — no additional editing is described as necessary since Cinema Studio 3.5 already handles per-scene polish.
Higgsfield as an all-in-one production platform — The video's central platform claim: Higgsfield hosts image generation, video generation, and audio generation together, positioned as the only place to run the full pipeline without extra subscriptions or tools. Apply: Run the entire workflow — asset generation, scene generation, and voiceover — inside Higgsfield rather than assembling a stack of separate single-purpose tools.
The video frames AI-film quality problems as a planning failure rather than a generation-quality failure: it claims most bad AI films come from skipping a locked story/character plan, not from weak models.
Character and location generation is treated as building a persistent asset library ('shopping list') to be referenced repeatedly across scenes, rather than generating imagery fresh for each shot — an asset-pipeline mental model applied to AI filmmaking.
Self-insertion into a generated character is claimed to require only two additional reference images (face + wardrobe) layered onto the standard character-sheet prompt, rather than a different generation process.
Prompting is treated as a two-stage translation problem: casual natural-language description first, then conversion into a structured JSON prompt via a dedicated free tool, on the claim that JSON gives the video model more usable detail than prose.
The three-panel (front/back/face close-up) character sheet is offered as the specific mechanism that gives a downstream video model 'everything it needs' to hold a character's appearance consistent across different camera angles.
«Well, after spending the last couple of weeks figuring out the exact workflow, I finally made it work and put together a full Hollywood-style short film completely with AI.»
— 00:37
«the most important step of this entire process that will actually make your films look good is planning out the entire movie.»
— 00:55
«Because most AI films online are getting made on the go. Nothing is really planned out, and that's why most of them actually turn out pretty bad.»
— 01:00
«we're going to build them in character sheets with a three-panel structure.»
— 03:23
«Today Manhattan watches the trial of the decade. A luxury hotel king stands accused of vanishing his biggest rival. ADA Jean wants a conviction, even without a body, but with Mr. Goodman leading the defense, anything can happen.»
— 08:18
«Higgsfield is the only place where you can actually run the entire production pipeline from images to video to audio without ever needing another subscription or tool.»
— 10:22
Reception
Strong positive reception with appreciative, engaged viewers requesting tutorials and expressing gratitude, with no negative feedback.
This is a promotional workflow tutorial for Higgsfield's Cinema Studio 3.5 that packages AI-film production into a repeatable plan-assets-scenes-voice pipeline, with every tool choice and claim of superiority tied to that one platform's current toolset.
10:38