Higgsfield tests AI animation timing and camera moves in Blender

• Passport Rush combined traditional animation with AI for a car chase that ran to 18 scenes in three days, according to Higgsfield AI's project thread.
• Difficult choreography went through a gray Blender previs before image references supplied the final look. Higgsfield AI's taxi-stunt breakdown uses a two-wheel squeeze between trucks as the example.
• The workflow links hand-drawn boards to 3D animation. Higgsfield AI's storyboard clip shows the first and last stages, while its pipeline card names Soul 2.0, Photoshop, Blender, Claude Opus 5.5, and Seedance 2.5.
• Higgsfield is also turning Blender operations into Claude skills: its crash-simulation demo and its scene-builder demo both end with editable scenes rather than a finished video alone.

A full 28-minute breakdown includes a continuity fix where artists painted missing headphones into a framing screenshot, and the next generation retained both the accessory and the camera angle. Higgsfield's official changelog says its Blender plug-in includes Scene Builder, 3D Model, Character Animation, Image and Video, and Camera tools. The strangest production failure was a comic pause after a broken handle, which proved harder than the chase stunts.

Higgsfield AI says in its Passport Rush thread that it began with storyboards, then used Soul 2.0 for character, prop, and location designs.

The team says in the same thread that every asset needing cross-shot consistency received its own sheet, with colors and painted detail adjusted by hand to bring 3D characters into the watercolor setting.

Higgsfield AI's storyboard-to-animation clip frames the pipeline as a hand-drawn board becoming a finished 3D AI shot.

The taxi stunt was planned in Blender because the intended movement was specific: tip onto two wheels, pass between trucks, and land. Higgsfield AI explains in its stunt breakdown that a gray preview carried movement and camera decisions, while image references handled the final appearance.

Higgsfield AI's three-step pipeline clip labels the handoff as Blender previs, Blender blockout, then final Higgsfield render. The same raw-3D-to-stylized-video handoff appears in another Blender render demo.

Their first watercolor taxi failed in motion, looking like a stretched PNG. Higgsfield AI reported in its taxi comparison that it replaced the moving watercolor object with a cleaner 3D car and left the painted treatment in the background.

Higgsfield AI gave the team one reusable Claude structure: scene and style, the role of each reference, timed action, then constraints. Its workflow card places that prompt step between Blender camera work and Seedance 2.5 video generation.

The full production chain was:

1. Shot preparation: handmade storyboards.
2. Character sheets: Higgsfield Soul 2.0.
3. Quick fixes: Photoshop.
4. Movement and camera work: Blender.
5. Scene prompts: Claude Opus 5.5.
6. Final video generation: Higgsfield Seedance 2.5.

A framing reference can override a character sheet. Higgsfield AI reported in its production breakdown that it painted headphones back into the camera screenshot after a generation dropped them, then regenerated with both the intended angle and accessory intact.

AIwithSynthia's action-sequence follow-up centers on a 30-second Seedance 2.5 campus chase described as a miniature production brief, not a short mood prompt.

The prompt separates the brief into:

• Character descriptions and wardrobe.
• A chain of locations, from campus corridor to stairwell, parking garage, convenience store, and exit.
• Individual fight actions and transitions between them.
• Audio direction.
• A cel-shaded 3D visual target plus a long negative-constraint list.

Higgsfield's current changelog describes a plug-in installed by dragging its downloaded ZIP onto an open Blender window, then opening the Higgsfield panel and logging in.

Three public demos describe Claude-operated scene tasks:

• /destruction-studio: Higgsfield's crash-simulation demo says Opus 5.5 simulates crumpled bodywork, glass, and debris in Blender, then returns an editable scene for camera changes and further destruction work.
• Exploded View: Higgsfield's product-breakdown demo turns a Blender model into an animated separation of its parts.
• /scene-builder: Higgsfield's scene-builder demo has Opus 5.5 arrange assets, pose characters, and set a camera.

Higgsfield paired Claude Opus 5.5 with Blender in its agentic 3D-modeling showcase. A separate building-collapse demo depicts a building reconstructed from one photo in Blender, then put through a collapse simulation.

The production's cut shot was not a stunt. Higgsfield AI reported in its seventh lesson that it could not get the tiny reaction pause after a girl snapped off a handle, either through prompts or Blender previews, so the moment never made the sequence.

The Higgsfield AI animation channel is where the company says more production breakdowns are coming.