Skip to content

GPT Astra 6 + Blender + Higgsfield Builds The Craziest Cinematic Scenes!

2026-09-12 · long tutorial / working session · Adil · recording, 15:36 · watch on YouTube

Topic: Video generation & shot direction (also: Prompting & AI assistants, Pre-production & story, Assets, references & consistency) · mode: Tutorial · tags: prompts shared, fiction, one creator's experience

In this long tutorial Adil directs GPT-6 Astra from one chat to block five scenes in 3D, make character and location references and render them with Seedance 2.5 through Higgsfield MCP. Blocking is a rough 3D version of the scene in simple shapes that fixes camera and timing before generating; he says it saves credits on rerolls and needs no high-end PC or advanced Blender skills. The scenes run from a seamless loop and a shot through a mirror to a late-70s casino built in 3D Jutsu, a concert with 66 dancers and a one-take vacuum commercial.

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: five scenes built without making the 3D setups by hand

00:00:37 — The infinite loop: the finished video played

00:01:19 — What blocking is; connecting Astra, Blender and Higgsfield

  • ★★★ Adil defines blocking as a rough 3D version of the scene made of simple shapes to plan the camera and action; he uses it to check how a loop's ending connects to its beginning before generating. — 00:01:19 · previsualization, Blender
  • ★★ To install the Higgsfield Blender Plugin: on Higgsfield open Plugins, choose Blender and download it; in Blender drag the plugin in and log in, and the Higgsfield toolbar appears in the viewport. — 00:01:19 · Blender, Higgsfield AI
  • ★★ To connect GPT-6 Astra, Blender and Higgsfield, Adil opens the chat, clicks Plugins, finds Higgsfield and enters the Higgsfield /use Blender command (as heard). — 00:01:46 · GPT-6 Astra, Blender, Higgsfield AI
  • ★★ Adil shows GPT-6 Astra using computer-use skills to check the Blender viewport live and building the whole loop setup on its own. — 00:01:46 · GPT-6 Astra, Blender
  • ★★ Adil briefs GPT-6 Astra with the story and asks for an editable camera path with a bit of handheld drift. — 00:01:46 · GPT-6 Astra

00:02:09 — Building and refining the infinite loop: nested miniature ending, base image, Seedance 2.5 render

  • ★★ To give the infinite loop a stronger hook, Adil keeps the room and opening camera move and replaces the ending with a pullback revealing the room on a table inside a bigger copy of itself, which drops back into the opening shot. — 00:02:52 · GPT-6 Astra
  • ★★ With the nested-miniature ending, Adil says the loop's first and last frames match although the camera is in a different spot, and setting this up by hand would have been a headache. — 00:03:21
  • ★★★ For a seamless loop Adil makes one base image for both start and end: GPT-6 Astra sends the blocking's opening frame to Nano Banana as a layout reference, prompted as an old mansion study. — 00:03:47 · Nano Banana, GPT-6 Astra
  • ★★ Adil uses the location image as the reference for both the opening and closing frames, and exports the Blender blocking as a video to serve as the motion reference. — 00:04:16 · Blender
  • ★★ Adil generates the cinematic loop through Higgsfield MCP with Seedance 2.5 at 30 seconds. — 00:04:16 · Higgsfield MCP, Seedance 2.5
  • ★ For the Seedance 2.5 loop Adil sets 16:9 and one continuous shot, and generates two versions to choose the better one. — 00:04:37 · Seedance 2.5
  • ★★ Adil says the loop prompts were nothing complex: GPT-6 Astra did the heavy lifting of the blocking and Higgsfield turned it into the final cinematic shot. — 00:05:17 · GPT-6 Astra, Higgsfield AI
  • ★★ Adil says he will put the blocking files in the description so viewers can test them. — 00:05:17

00:05:39 — The impossible mirror shot: mirrored room, detective character sheets, two 15-second versions

  • ★★★ For a 15-second shot where the camera backs through a mirror, Adil has GPT-6 Astra build a mirrored copy of the room with both characters moving in sync; the mirror is just an opening the camera passes through, and Astra makes the character and location through Higgsfield MCP. — 00:05:47 · GPT-6 Astra, Higgsfield MCP
  • ★ GPT-6 Astra confirms in the chat that layout and timing are locked and sets up the Blender timeline markers itself. — 00:06:37 · GPT-6 Astra, Blender
  • ★★ For the mirror shot Adil has GPT-6 Astra make four detective character sheets through Higgsfield, picks one, attaches it with the blocking and asks for two 15-second versions, each one continuous shot. — 00:06:37 · GPT-6 Astra, character sheet
  • ★★★ Adil says Seedance 2.5 turns the blocking into a finished shot in the same chat, replacing texturing, shading, lighting, rigging, rendering and compositing by hand. — 00:07:25 · Seedance 2.5
  • ★★★ In 3D Jutsu on Higgsfield a scene is built and adjusted in the browser and the blocking sent to Seedance for the video; from a reference video, 3D Jutsu extracts the blocking and the camera move. — 00:07:47 · Seedance
  • ★★ Adil says prompting one camera through four locations in a 30-second continuous shot with text alone would be a mess where everything falls apart. — 00:07:47

00:07:48 — 3D Jutsu: four locations in one continuous shot from the street into a late-70s casino

  • ★★ In 3D Jutsu Adil chooses GPT-6 Astra and asks for one continuous 30-second shot from the street to the casino, with separate paths for the camera and the characters so he can adjust them. — 00:08:17 · GPT-6 Astra
  • ★★ Adil says Higgsfield's Discord community has almost 200,000 users who help with questions. — 00:08:17 · Discord
  • ★★ Adil revises the 3D Jutsu layout with ceilings and an entrance door and a stronger opening: start behind the car window, the car pulls away and the camera follows the couple inside. — 00:08:44
  • ★ In the casino blocking, colours mark roles (black the man, white the woman, yellow the kitchen workers) and paths show where each moves. — 00:09:13
  • ★★★ For the late-70s casino shot Adil gives character sheets only for the couple and reference images for the street and gaming hall, describes cooks, staff and guests in the prompt, and has GPT-6 Astra write the Seedance 2.5 prompt. — 00:09:13 · character sheet, GPT-6 Astra, Seedance 2.5
  • ★★★ Adil says planning the camera journey in 3D saves credits on endless rerolls and a lot of time by locking the exact camera move, and needs no high-end PC or advanced Blender skills. — 00:09:54 · Blender
  • ★★ For a crowded concert Adil uses blocking to plan formations and check them from each angle, asking GPT-6 Astra for a 30-second Blender sequence already cut: close-ups, wides, overheads, slow dolly moves and quick push-ins. — 00:10:37 · GPT-6 Astra, Blender
  • ★★ In the concert blocking the singer stands at the centre within three rings of dancers and 10 columns; dancers are boxes and the singer another shape so the roles are easy to tell apart. — 00:10:37

00:10:39 — Concert: one singer, 66 dancers, cut sequence with music generated with the video

  • ★★★ Adil lets the blocking boxes fix only where dancers stand and face, and writes the choreography in the prompt so the model keeps movement fluid instead of copying the stiff blocking. — 00:11:17
  • ★★ Adil has GPT-6 Astra make the singer's character sheet from his own photo, and Soul Cinema create the desert location with mirrored columns. — 00:11:17 · character sheet, GPT-6 Astra, Soul Cinema
  • ★★ For the concert Adil puts the music and vocal directions in the same Seedance 2.5 request so the track is generated with the video. — 00:12:00 · Seedance 2.5
  • ★★ Adil says the dancers stayed in formation across the complex camera moves of the Seedance 2.5 concert, which he credits to the pre-planned blocking. — 00:12:42 · Seedance 2.5

00:12:54 — The one-take vacuum commercial played

00:13:32 — Building the vacuum commercial: chase blocking, references, Seedance 2.5; community and Discord

  • ★★ Adil's 30-second vacuum commercial is one continuous take across several locations; shifting camera sizes and focal lengths on the fly gives macro shots and transitions between scenes. — 00:13:32
  • ★★ In the vacuum-ad chase blocked by GPT-6 Astra, the 10-centimetre hero is sucked through the hose into the dustbin, then the camera pulls back through its clear plastic wall to reveal the product. — 00:13:58 · GPT-6 Astra, Blender
  • ★★★ For the vacuum ad Adil has GPT-6 Astra make references for the hero (from his photo), the woman (striped pajamas and slippers) and the vacuum with its transparent dust compartment, then generate with Seedance 2.5 from the blocking and all references. — 00:13:58 · GPT-6 Astra, Seedance 2.5, character sheet
  • ★★ Adil says if a client wants changes to the vacuum ad, the blocking is ready to revise. — 00:14:46
  • ★★ Adil says viewers made impressive work from the blocking shared after the previous Blender video. — 00:14:46 · Blender
  • ★★ Adil asks viewers to take the master file from the description, try the scenes and share their versions in Discord. — 00:14:46 · Discord
  • ★★★ Adil sums up: each scene had something hard to get from text alone, so camera and timing were worked out in 3D first and GPT-6 Astra brought them to life through Higgsfield MCP. — 00:14:46 · GPT-6 Astra, Higgsfield MCP
  • ★★ Adil directed all five scenes from a single chat while GPT-6 Astra handled Blender and Higgsfield, without him building anything by hand. — 00:14:46 · GPT-6 Astra, Blender

00:15:05 — Recap and sign-off

Threads in this video

Where the stories that moved here went next

Mentioned

Central: 3D Jutsu · Blender · GPT-6 Astra · Higgsfield MCP · previsualization · Seedance 2.5

Also: character sheet · Discord · Higgsfield Blender Plugin · Nano Banana · Soul Cinema

Also centred on 3D Jutsu: Changelog: Introducing 3D Jutsu · Higgsfield 3D Jutsu: Build and Animate 3D Scenes from a Prompt

Replies to

← Previous: Higgsfield Effects 2.0 in ChatGPT · AI Video Effects · Next → GPT Image 2.5 Sunburst · Higgsfield x OpenAI