Consistent scenes
Build a location or product once in Blender, render every angle and frame pair from it, and keep characters steady with stand ins or rigged models.
Build the set or product once, then render everything from it. Because it is real geometry, every angle, lens and lighting setup comes from one source. The renders become reference images, location plates, first and last frames, or keyframes, and video models follow those much better than text.
Build a set or product once
Get the asset
Let Opus model simple hard surface objects from primitives. Pull everything else from a library: one creator guide warns that having Opus model ordinary props such as chairs wastes tokens and looks fake. BlenderKit (now branded Blendkit) lists 152,407 assets, 73,681 of them free. Or generate the asset (see below). For a product, import the client's CAD file or a photo scan when one exists.
Normalise it
Real dimensions, origin at the base centre, transforms applied, clean material names. Ask Opus to confirm with get_object_detail_summary.
Build locations in stages
Block the space first, then models, then materials. A widely shared organic case built a laundromat from boxes up on a MacBook Air. It used one procedural wear node group (edge wear, dust, grime, drips, fingerprints) on every material at different strengths, made signs as 2D designs rendered top down into textures, and corrected the scene with plain language notes while Claude checked close ups.
Lock a camera kit
Put an empty at the centre of the subject with cameras parented to it: an orbit every 30 or 45 degrees, eye level and high angle, a macro and a top view. Keep the same lens, lights and exposure for every frame.
Render the reference set
Render in EEVEE or Cycles with real materials, at the target aspect ratio, plus one version in neutral light.
Render each shot's frame pair
For every video shot, render the exact first and last frames from the shot camera. Keep the change between them small: the more two frames differ, the more the model invents.
What to make from one scene
| Render | Use it as | Models that take it |
|---|---|---|
| Orbit stills of a product or set | Reference images | Seedance 2.5 (up to 30 images), Kling, Vidu |
| Location plates matched to the blockout angles | Location reference next to a clay video | Seedance 2.5 |
| First and last frame from the shot camera | First and last frames | Veo 3.1, Kling 3.0, Wan 3.0, Seedance, MiniMax H3 |
| A series of keyframes | Keyframes | Luma Ray3.2 (up to 16), Runway Aleph 2.0 (up to 5 on an input video), Kling 4.0 (up to 10, announced, not confirmed shipped) |
Community reports show why this matters. Creators advise always attaching products and logos as images, because a logo described in text renders as fiction. And one creator's location plates, matched to the blockout angles, kept an empty Shibuya street consistent across six angles. These are community results, not measurements, but they agree with the official advice to use keyframe images when you need strict alignment. Limits per model are on Hand off.
Characters: three levels
Start at level 1. Move up only when a shot needs it.
Level 1: stand in plus character sheet
A colour coded capsule or mannequin in Blender. The look comes from a 4 view character sheet made with an image model, and the prompt maps the colour to that image. This is the official Seedance method and the cheapest. Use it for most shots. See Reference images for the sheet.
Level 2: generated mesh, rigged by Opus
Use this when you need the character in specific poses. The pipeline creators repeat: make a turnaround (front, back, three quarter) with GPT Image 2.5 or Nano Banana, turn it into a mesh with an image to 3D tool, then have Opus scale, clean, rig and animate it in Blender. Render the result as references or keyframes.
Organic examples, each from a single creator:
- GPT Image 2.5 four views, Tripo Smart Mesh 2.0 with 8K textures, then Opus 5.5 did textures, expressions, rig and motion in Blender.
- Opus 5.5 with Tripo P2 and the Blender connector made a mesh, an IK rig with weights, clothes and an animation in about 20 minutes.
- Gemini front view, ChatGPT back and 45 degree views, Hunyuan3D, then Blender through MCP.
- Trellis Studio run locally, a polygon reduction add-on, then Auto-Rig Pro.
- On the same workflow, Opus 5.5 made a richer model and better animation than Sonnet 5.5, for about 1.4 times the cost.
Use the rigging prompt on Prompts to have Opus check and fix the mesh first.
Level 3: motion from a real performance
Drive the rig, or the video model directly, with captured motion.
| Tool | What it does | Notes |
|---|---|---|
| Mixamo | Auto rig and animation library | Free with an Adobe ID, bipedal humanoids only, in maintenance mode. Third party guides say its old Blender add-on fails on 4.x (not official) |
| Rokoko Vision | Motion capture from video | Free Starter plan with FBX export, 15 to 30 s of processing a month depending on which Rokoko page you read |
| Meta SAM 3D Body | Human mesh from a single image | Open code. ComfyUI chains it into Blender and Seedance (vendor post) |
| ComfyUI Video2MotionCapture | Video to motion, exported as glb | Shown in one organic post |
| Tencent HY-Motion | Text to motion onto Mixamo rigs | Shown in one organic post |
Or skip the rig: give the performance clip straight to Kling Motion Control, or add it as a second Seedance video reference while the stand in stays still.
3D asset generators
Prices checked on 15 Oct 2026.
| Tool | What it does | Runs on | Licence or price | Watch out |
|---|---|---|---|---|
| Microsoft TRELLIS.2 | Image to 3D, 4B parameters, full PBR, GLB export | Linux tested, NVIDIA GPU with 24 GB or more (A100 and H100 verified) | MIT | About 3 s at 512³, 17 s at 1024³ and 60 s at 1536³, all on an H100; posts quoting "3 seconds" leave that out. GLB transparency is off by default |
| Tencent Hunyuan3D 2.1 | Image to 3D with a PBR texture model | macOS, Windows, Linux; 10 GB VRAM for shape, 21 GB for texture, 29 GB for both | Tencent community licence. Does not apply in the EU, the UK and South Korea. Extra licence above 1M monthly users | A wave of near identical posts hypes the older 2.0 release |
| Tripo P2.0 (Smart Mesh) | Quad dominant meshes with separated parts, up to 25k quad faces | Cloud | Free tier of 2 generations; commercial use on paid plans | Released 21 Sep 2026. Export FBX if quads matter (GLB may triangulate). An independent review found no measured topology data yet |
| Meshy | Text or image to 3D, PBR, auto rig, animation presets | Cloud, with an MCP server that works next to the Blender connector | Paid plan for the API; every follow up is a new task billed in credits | Meshy itself recommends a human review of topology |
| Hyper3D Rodin Gen-2 | Quad or raw meshes, Blender add-on | Cloud | API through resellers | Quad mode from 1k to 200k faces; version dates disagree across sources |
Which to pick: on a normal laptop, use Tripo or Meshy, since both run in the cloud. Use TRELLIS.2 only with a 24 GB NVIDIA card on Linux or a rented cloud GPU. Skip Hunyuan3D if you live or sell in the EU, the UK or South Korea.
Check every generated mesh
Generated meshes need a scale, normals and topology check before rigging. Ask Opus to report non manifold edges, loose parts and polycount through execute_blender_code before you build on them.
Sources
- BlenderKit (Blendkit)
- Microsoft TRELLIS.2
- Hunyuan3D 2.1 and licence
- Meshy MCP guide
- Digital Production: Tripo P2.0 test
- Runware: Hyper3D Rodin Gen-2
- Rokoko pricing
- Meta SAM 3D Body
- Cinevva guide to Mixamo (third party)
- Laundromat build: hayashimon1 (organic)
- Character pipelines: manaimovie, luccacerf, Spectromachina, heecheee, jurlycat (all organic)
- Location plates: gensoFX (organic)