Structural Anchoring: Using 3D Blockouts, Phonetic Pangrams, and Three-Panel Sheets for Total Scene Control
Filmmakers are moving from unpredictable text-only prompting to structural anchors—3D environments, phonetic voice data, and simplified character sheets—to lock in consistency across Seedance 2.5 and ElevenLabs workflows.
Top learnings
Use a simplified three-panel character sheet featuring a face closeup, front outfit, and rear body to prevent model confusion and facial drift.
Build low-poly 3D environments in Blender to provide exact camera and motion references, ensuring spatial consistency that text prompts cannot match.
Generate a phonetic pangram in ElevenLabs to provide a complete vocal data set, allowing video models to maintain consistent accents and tones during lip-syncing.
Why are structural anchors replacing text-only prompting?
AI filmmaking is shifting from guessing with text prompts to anchoring with structural data. Text prompts often lead to drift, where characters or voices change between shots. To solve this, creators are using 3D blockouts—simple 3D scenes that define camera movement—and phonetic pangrams, which are sentences containing every sound in a language. These anchors provide the AI with a concrete map of how a scene should look and sound, reducing the reliance on the model's random noise generation. Dan Kieft observes that 3D renders serve as superior references because they provide exact camera and object motion that text cannot describe. Similarly, Rourke demonstrates that providing a complete phonetic data set allows video models to maintain consistent accents and attitudes across different generations.
Read source excerpts 2
scenes because you've built out the entire thing in 3D. The second reason why I use Blender is because it gives better references than just prompting. Because you have this
Dan Kieft
that uses every sound in the English language. So, we want a sentence that frenetically pronounces every single sound. And this is called a phonetic panggram. So, this is an
Rourke
How do visual and vocal consistency approaches differ?
Creators differ in how they manage these anchors. Jian argues that less is more for visual consistency, recommending a simplified three-panel character sheet with only one prominent face to avoid confusing the model. He notes that multiple face angles in a single sheet often lead to facial drift. In contrast, Rourke focuses on vocal anchoring, using Claude to translate a character's backstory into technical voice prompts for ElevenLabs. While Jian uses specific physical scales, such as defining a coffee mug as exactly 9 cm tall, to maintain object size, Rourke uses phonetic data to ensure the AI understands the full range of a character's accent and tone. Both agree that providing the model with a single, high-quality reference point is more effective than providing multiple, conflicting data points.
Read source excerpts 2
camera angles multiple face views of the same character. It turns out this actually confuses the AI video models. It's better to have one prominent face in your character
CyberJungle
I've also spent some time building a free Claude skill that you can plug straight into your own Claude, which actually has all of 11 Lab's own best practices when it comes to
Rourke
What is the Anchor Pipeline for multi-shot consistency?
A practical workflow begins with building a low-poly 3D scene in Blender. Using the Higgsfield MCP (Multi-Control Plugin), directors can use natural language to position cameras and objects. This 3D render is then exported as an MP4 to serve as a motion reference. Simultaneously, a character sheet is generated—ideally using a model like SeaDream 5.5—and a phonetic pangram is recorded in ElevenLabs 11v4. Finally, all three assets (motion reference, character sheet, and audio) are fed into Seedance 2.5 to generate the final, consistent shot. ReelStack suggests adopting this structural anchor pipeline for narrative projects where character identity and spatial logic are non-negotiable. By locking the environment in 3D and the voice in phonetic data, you insulate the project against the inherent randomness of diffusion models.
Read source excerpts 2
Hicksville plugin for Blender. This allows you to directly prompt inside of Blender through Hicksville, and you can control everything through prompting. For this, you're
Dan Kieft
because we also of course want to keep the location consistent. And after that, in the very beginning of my seed dance 2.5 prompt, I always start by mapping the references
CyberJungle
What are the hardware limits and generation trade-offs?
These workflows have clear technical limits. High-detail 3D scenes can crash standard PCs, requiring creators to stick to low-poly prompts during the blockout phase to save compute power. Furthermore, diffusion-based video models like Minimax H3 still struggle with fine details like text or map lines because they generate images from random noise. While tools like Content Factory bypass this by using coded graphics to ensure text accuracy, they lack the cinematic flexibility of full diffusion models. Creators must also manage credit costs, as 1080p hero renders are significantly more expensive than 2K image assets. Rourke notes that even with high-quality character sheets, generating in 4K is often prohibitively expensive, making 2K the current production standard for asset creation.
Read source excerpts 2
his apartment. Important thing to note, I'm saying a low poly scene here in Blender. If we don't do that, your computer might get fried because it might be too detailed. Okay,
Dan Kieft
Miniax H3, which is a video generation model. A video model starts every shot from random noise and slowly turns that into a picture. So, small details like the lines on a map
Youri van Hofwegen
Key moments to explore
Optional deep divesWant to see a technique in action? Jump into the source videos. These AI-extracted timestamps may be approximate.
- 02:55 ↗Blender ensures perfect consistency across scenes because the entire environment is constructed in a 3D space.Dan Kieft · Claude + Blender Is The Best Way To Make AI Videos (Full Course)
- 00:58 ↗A five-step process for character voice consistency involves character design, voice prompting with Claude, building the voice in ElevenLabs, scene generation in Seedance 2.5, and environmental integration.Rourke · Seedance 2.5 Tutorial: Consistent Voices With Elevenlabs V4
- 01:15 ↗Less is more for consistency: use one optimized three-panel sheet instead of complex multi-angle sheets.CyberJungle · The Newest Method to Make Realistic Consistent AI Characters With Seedance 2.5
- 01:27 ↗Seedance 2.5 supports 1080p resolution and 16:9 aspect ratio for generations up to 30 seconds.Youri van Hofwegen · How to Make YouTube Explainer Videos With AI (No Editing Required)
- 03:04 ↗3D renders serve as superior references for AI video generators compared to text prompts by providing exact camera and object motion.Dan Kieft · Claude + Blender Is The Best Way To Make AI Videos (Full Course)
- 02:19 ↗Character sheets featuring multiple angles are essential for locking in visual consistency across various AI generations.Rourke · Seedance 2.5 Tutorial: Consistent Voices With Elevenlabs V4
Put it into practice
Plan camera movement with Blender →Practical steps and checks before committing to production. Get the practical weekly briefing →The useful ideas in one email. Subscribe to keep learning. Need a filmmaker for your project? →Tell us what you want to make and submit a project brief.Go to the source
4 videos