The Cellist
Claude x Blender MCP | 10.2026 | Claude , Higgsfield, Blender
📚 Overview
By connecting Blender directly to AI video models via Claude MCP, this personal project allowed me as a single compositor to execute a 30-second short film - bypassing the steep 3D learning curve while solving generative AI’s notorious lack of camera control.
For The Cellist, I used the Higgsfield plugin to let Claude build my 3D previs directly inside Blender, transforming a normally throwaway block-out into a rigid spatial skeleton for the AI to paint over. Driving the generative video model with this proxy geometry solved the biggest hurdles of standard text-to-video tools by strictly locking in my directorial intent for exact camera moves, focal lengths, and cut points. Because the AI had a precise framing reference to follow, I eliminated the token waste of random AI guesses and jumped straight into look-development—leveraging my traditional compositing skills to stitch together a cinematic film that would normally require a full 3D team and weeks of rendering.
💡The Inspiration
A successful AI film still requires a strong human conceptual anchor—for this project, it was taking a classic photograph and animating the punchline.
The core idea was inspired by one of my favorite photographys by Robert Doisneau, "The Cellist". I loved the quiet humor of that image and asked myself what might happen next. That question became the narrative engine for the film: he protects the instrument with everything he has, only to hail a taxi that brakes through a puddle and splashes them both.
🚀 Process
Step 1 - Building the scene in 3D
Using the Higgsfield MCP Bridge, I connected Claude directly to Blender and guided it, in plain language, to build a rough 3D version of the photograph’s rainy Paris setting. I then worked with Claude to write the script and turn it into a shot list, which Claude used to block out each shot in Blender with its own camera, framing and basic character animation. Because Claude could make changes live in my Blender file and check its own work, I could refine the shots quickly without heavy manual 3D work.
Prompt : Create a rough 3D blockout scene in Blender using simple primitives (low detail is fine). Use metric units. X = left/right, Y = depth (away from camera), Z = up. Ground level is Z = 0. SCENE OVERVIEW A rainy, overcast city overlook at a sidewalk edge. A tall cello case stands upright on the wet pavement. A man in a trench coat stands to its right and holds an open umbrella over the cello case, not over himself. Behind them is a wrought-iron railing at the edge of a terrace, with a misty, washed-out sky beyond. An ornate street lamp post stands at the far right. The mood is quiet, gray, and wet. GROUND - Sidewalk: flat plane, 8 m wide (X from -4 to 4), from Y = 0 to Y = 4, at Z = 0.15. - Curb: long box along the front edge of the sidewalk at Y = 0, 8 m long, 0.15 m tall, 0.2 m deep. - Street: plane in front of the curb (Y = -3 to 0) at Z = 0. Give it a subtle cobblestone look (a displacement or voronoi texture is fine, or leave it flat). - Make the sidewalk material very glossy and dark gray (roughness ~0.1) so it looks rain-soaked with reflections. CELLO CASE (center of the scene) - Location: X = 0, Y = 1.2, standing on the sidewalk. - Total height ~1.35 m. Build it from a stretched, slightly flattened UV sphere for the wide body (about 0.55 m wide, 0.3 m deep, 0.8 m tall), a narrow cylinder or box for the neck (about 0.15 m wide, 0.35 m tall), and a rounded capsule shape on top for the head/scroll section (about 0.2 m wide, 0.25 m tall). - Add two small feet or wheels at the bottom. - Material: near-black, slightly glossy. MAN WITH UMBRELLA (right of the cello) - Location: X = 1.1, Y = 1.2, facing toward the cello (rotated ~30° to the left of camera), height ~1.8 m. - Simple mannequin: legs from two cylinders (dark trousers), a body from a tapered cylinder or cone for a knee-length, belted trench coat (light beige/off-white), head from a sphere, and a flat beret on top (a short, wide, slightly tilted cylinder). - Left arm hangs down with the hand tucked into a coat pocket. - Right arm raised up and out toward the cello, elbow slightly bent, hand at about Z = 1.9 m, X = 0.35. - Umbrella: thin cylinder shaft running from his hand straight up to Z = 2.45 m. Canopy is a shallow cone (radius ~0.55 m, height ~0.18 m, 8 segments to give it ribs), centered roughly over the cello at X = 0.25, with a short tip spike on top. Dark material. RAILING (background) - Wrought-iron fence along Y = 3.5, from X = -4 to X = 3, height ~1.0 m. - Build it from a top rail, a bottom rail, and thin vertical bars spaced ~0.12 m apart, with a thicker post every 1.5 m. - In the middle section (around X = -1.5 to X = 0.5) the railing turns and angles down and away from the camera, as if following a staircase descending behind the terrace. Add a sloped handrail section there. - Material: black metal. LAMP POST (far right) - Location: X = 2.6, Y = 2.2. - Wide, stepped base (stacked cylinders, ~0.5 m diameter, ~1.2 m tall), then a tall tapered column with a few decorative rings, rising to about 6 m so that it extends past the top of the camera frame. - Material: dark, slightly rough metal. BACKGROUND / ATMOSPHERE - No visible horizon detail. Beyond the railing, the world should fade to a bright, foggy off-white. - Optional: a few very faint, low boxes far in the distance (Y = 40+) to suggest a city below, nearly lost in fog. - Add a Volume Scatter to the world or a large volume cube for light fog (low density) so depth fades to white. - No rain particles. The rainy feel comes only from the wet, reflective ground and the fog. LIGHTING - Soft, flat, overcast light. Use a bright light gray world background (strength ~1.5) as the main light. - Add a large area light high above and behind the scene for gentle rim highlights on the wet ground. - No harsh shadows. CAMERA - Location: X = 0.4, Y = -5.5, Z = 1.3, pointing straight toward +Y with a very slight upward tilt. - Focal length ~45 mm. - Portrait aspect ratio: resolution 1080 x 1400. - Frame it so the cello sits just right of center, the man fills the right-center, the lamp post is cut off at the right edge, and the curb runs along the bottom. RENDER / LOOK - Render engine: Eevee (with screen-space reflections on) or Cycles. - In the compositor, convert to black and white and add a warm sepia tint, high contrast, slightly lifted whites, and a light vignette. Keep everything as simple primitives, grouped into collections named Ground, Cello, Man, Railing, LampPost, and Environment.
Shot list: Overview Format: it still says the list assumes 16:9. The film is now vertical 1080 × 1350. Running time and shot count: it says “Eleven shots, about 43 seconds.” The final edit is about 30 seconds, with S03A and S05 cut. “Still to build”: it lists the cab, puddle and splash, which were all built long ago. Act 1 Shot 1: it describes a calm idle. Your later version has a grumpy expression, subtle cold gestures from the first frame, and a few more drops after he looks up. Shot 2: the umbrella opens at 2.7 s after he reaches down. In the final edit, the umbrella is already in his hand and opens about 1.2 s in. Shot 3A: it describes one drop wiped with his thumb, then two pats. It became a few drops gently brushed off with his bare hand, in a low-angle close-up, and he never looks up. The shot was also cut from the final edit. Act 2 Shot 4: it still has the shiver. That was replaced by hunched shoulders, steamy breath and actively looking for a cab, and in the final edit he looks down at the cello. Shot 5: it’s an extreme close-up insert. It became his point of view, with the watch reading almost four o’clock, the stomach growl and the big puddle in the background. It was also cut from the final edit. Shot 6: the cab passes in the far lane with the puddle in the foreground. In the final version it whooshes past close to the camera, he searches actively and looks clearly disappointed, and he stands closer with his arm straight. Shot 7: a head snap and a mild smile. It became a shot that starts already looking down the road, spots the cab, and waves frantically while yelling “TAXI!” with a French accent. Act 3 Shot 8: “mid-hail, smiling.” It’s now frantic waving, then relief. The kind, oblivious driver isn’t mentioned, and neither is the irregular natural puddle. Shot 9: the line is “Oh no.” It’s now a devastated “Ah non...” in French, with his face crumpling. In the final edit he starts with his hand at his forehead. Shot 10: it says “cut to black” and runs 4 s. It now has the iris out and “THE END,” and runs about 5 s. Continuity table The cello: there’s no row or note about the cello staying completely dry until the splash, which became a key rule. Shot 3A: its row still says “wipes the drop, pats the case.” Shot 4: the left-hand note doesn’t match the final version.
Step 2 - Creating the looks
I designed everything on screen as still images in Higgsfield cinema studio: the main character Luc, the taxi driver, the cello case, the umbrella, the taxi, and the street before and during the rain. I auditioned multiple image generations for each element, refining the prompts until the designs felt right: realistic but slightly caricatured characters, true-to-life color, and props that matched the story. The chosen images became my reference set, keeping every shot consistent.
look development of the main character "LUC"
full reference sheet
Step 3 - Creating videos
I generated the live-action shots in Seedance, pairing my reference images with the Blender playblast of each shot so the AI followed my camera angles and timing. Some shots took several rounds: keeping the cello dry until the splash, making raindrops look real, and getting natural, human movement instead of copying the rough 3D animation. Before each regeneration I updated the Blender shot to match my direction, which saved credits.
some of the takes that didn't make it to the final edit
Step 4 - Editing & Adding vintage looks
I edited the final cut from the best takes, removing two shots the story didn’t need. In After Effects I added film grain, bloom, a vignette and color correction to give the film the feel of an old print, in keeping with the era of Doisneau’s photograph.
📍 Overall Takeaway
Integrating Blender MCP into a generative pipeline solves AI video’s fundamental control issues—saving precious tokens while proving that as technical barriers fall, visual storytelling becomes the ultimate differentiator.
Using a 3D previs anchored via Blender MCP granted precise control over camera angles, focal lengths, and scene framing that were virtually impossible to lock down through text prompting alone. Having a rigid spatial reference eliminated the wasteful trial-and-error generation loop, keeping token consumption exceptionally efficient. Beyond the technical pipeline savings, this project reinforced a vital lesson for the modern visual landscape: as generative tools democratize complex 3D layout and rendering, raw execution becomes fast, making foundational storytelling, character pacing, and clear directorial intent more essential than ever.