Blender previs gives Seedance planned camera references

• A screenplay can become a rough Blender animation before any final video generation, then serve as Seedance reference footage, as rainisto's v0.002 test describes.
• The working file retains planned cuts, camera angles, and lens choices, according to rainisto's Blender follow-up.
• The control problem is deliberate: Seedance needs room to improve crude blocking without rewriting what the blocking video establishes, while rainisto's failure clip shows how badly a guidance system can drift when it forgets its instructions.

Runway’s Seedance 2.0 guide describes a model that accepts video alongside text, image, and audio references. Blender’s markers documentation includes camera-bound markers for specific frames, and BytePlus’s ModelArk documentation lists Seedance tutorials and prompt guides. Rainisto’s experiment joins those pieces into a lightweight previs pipeline: decide the shots in 3D, then hand the model moving evidence of those decisions.

Rainisto’s sequence is compact:

1. Start with a screenplay.
2. Generate a rough Blender animation.
3. Use that animation as video reference for Seedance.

The goal is to establish shots, camera angles, cuts, and lens selection before generation, then preserve them in the Blender timeline. The roughness is part of the design. Previs is being used as a controllable camera-and-edit plan, not as the final look.

A later update from rainisto says the camera-angle changes are saved in Blender. That makes the generated reference a record of editorial choices rather than a single image that merely suggests composition.

Blender’s marker system can denote camera changes and bind a camera to a selected frame. Rainisto did not specify the implementation, but the native timeline mechanics fit the kind of shot record shown in the test.

Rainisto frames the handoff as a narrow control window: the video model should improve the “crappy blocking” but must not completely reinterpret it. The reference is therefore both a constraint and an invitation to add finish.

A separate rainisto guidance-workflow clip captures the opposite outcome, where the LLM forgets the prompts it is supposed to use and invents its own path. In this workflow, that failure would mean the camera plan survives in Blender but not in the generated shot.

The same creator uses an intentionally different visual style to separate content guidance from final-image styling. In rainisto's note on environment references, detailed hand-drawn sketches supply the environment while their stylistic distance is meant to stop the model from copying the sketch’s lighting or finish.

That gives the reference stack distinct jobs:

• Blender video carries cuts, lensing, and camera movement.
• Hand-drawn environment art carries place and layout.
• Final generation supplies the target rendering style.

Techhalla’s Seedance 2.5 prompt shows the textual counterpart to previs. Instead of tagging each supplied image, the creator names the character turnarounds, slingshot, and treehouse as strict visual references, then lays out a 30-second slingshot battle in time-coded beats.

• 0 to 3 seconds: standoff, rapid extreme close-ups, then silence before the action.
• 3 to 7.5 seconds: load, release, projectile orbit, and a dive behind cover.
• 7.5 to 16.5 seconds: counterfire, tracking shots, repositioning on a tire swing, and a second launch.
• 16.5 to 21 seconds: 1.5-second chaos cuts, then a hero-angle final shot.
• 21 to 30 seconds: an overshot projectile, shattered window, and a freeze on the neighbor’s reaction.

The prompt specifies moves such as push-ins, whip pans, low angles, crane shots, speed ramps, and freeze-frames. Rainisto’s Blender pass could carry those same beats as motion and editing rather than prose alone.

For a separate Seedance piece, techhalla's reply described two ways to handle a composite: Photoshop Harmonize for layer lighting, or passing the composite through Nano Banana 2 or Pro with a relighting instruction. The same reply says techhalla would skip composites when using Seedance’s omniRef mode.

That places another decision beside Rainisto’s previs: whether individual elements arrive as assembled plates that need lighting reconciliation, or as references the video model resolves together.

The Blender method uses 3D to plan a video. DannyLimanseta’s horde-shooter prototype uses generated 3D assets at runtime: Grok 4.6 handled nearly everything in a ThreeJS game prototype, including UI, VFX, and SFX, while Meshy supplied the character models and rigs. A follow-up points directly to Meshy as the asset source.

Limanseta said the ThreeJS and WebGPU build held thousands of enemies on screen, though depth of field and other shaders pushed its limits. It is a different endpoint for AI-generated 3D, where the asset has to survive a realtime renderer rather than serve as a camera reference.