I've been testing Luma Dream Machine for character consistency in longer clips, specifically trying to maintain a recognizable face across a full 10-second generation. The out-of-the-box results were... variable. After a couple dozen credits of experimentation, I've landed on a workflow that gets me about 80% usable consistency, which feels solid for the current state of the model.
My core finding is that you need to treat it like a two-stage process: first lock the character, then animate it. Here's my step-by-step:
* **Stage 1: The Character Anchor**
* Generate a high-quality, well-lit 3-4 second clip of your character with minimal movement. A slow pan or a subtle expression change works best.
* Use a detailed prompt for the face: think "a woman with sharp green eyes, a small mole on her left cheekbone, and thin arched eyebrows" not just "a woman".
* If the result is good, download it. This is your anchor clip.
* **Stage 2: Extending with Consistency**
* Upload the anchor clip in the 'Prompt with Video' mode.
* For the new prompt, describe the *action* you want, but **re-state the key facial features from your first prompt**. This seems to reinforce the model's memory.
* Keep the camera motion relatively simple. Aggressive zooms or spins tend to break the facial geometry faster.
* Be prepared to iterate 2-3 times per final clip. Sometimes regenerating with the same anchor and prompt fixes minor drift.
The biggest cost saver here is nailing the anchor clip. Burning credits to get a perfect 10s clip in one go is inefficient. It's cheaper to spend 3-4 credits perfecting a short, stable reference, then use that to generate the longer sequence.
Has anyone else found a different parameter tweak or prompt phrasing that improves consistency? I'm especially curious if using descriptive names (e.g., "a character named 'Mara' who has...") in the prompt has any measurable effect, or if that's just placebo.