A visual bible needs more than one reference image

The shot

Production guidance has long recommended building a reference set, sometimes called a visual bible, of character, wardrobe, and location images before a sequence goes into production. Two current video-generation products document a matching interface feature. Runway's Gen-4 Image References guide states that up to three reference images can be used for a single generation, and its companion article on using reference media extends the same idea to uploaded video and audio, each cited by an at-mention inside a prompt. Luma Labs documents the same concept for Dream Machine in its Character and Object Consistency guide, and states in its Ray3.14 FAQ that the current Ray3.14 model "lacks character consistency and the reference feature" outright, as of the FAQ's 18 September 2025 publication and this note's 16 September 2026 retrieval.

What the documents show

Runway's own promotional claim is that Gen-4 References "excels at generating consistent characters across different lighting conditions, locations, and treatments, all from just a single reference image" -- a vendor capability claim, not an independently measured result, and it describes still-image generation rather than a video sequence. Luma's reference guide states the opposite kind of thing plainly: giving the system one image and asking for an unseen angle "has to invent what it cannot see. That's when drift happens." Luma's documented fix is not a single toggle but a workflow: a Master Reference Assets pack of multiple clean angles -- front, rear, top, three-quarter, side, full body -- plus a written description of what must not change. This is the vendor's own admission that a single reference is insufficient, offered as a practical checklist rather than a solved feature.

The workflow

Per Runway's documentation, a reference image, video, or audio clip is uploaded or dragged in, optionally tagged and saved for reuse across sessions and shared with a workspace, then called into a prompt by name; images specifically are capped at three per generation. Per Luma's guide, an operator instead assembles several angle images, has the system describe each into a structured text document, locks a list of features that must not change, and re-anchors every new variation to that saved pack rather than to the most recent output. Both are documented, current interface steps, not third-party accounts of use.

What the tool does not change

Neither vendor's documentation claims the reference feature removes the need for a person to decide which features are load-bearing to a character's identity, to catch drift before it compounds across a sequence, or to correct it once it appears. Luma's own guide frames correction as an ongoing manual step, not an automated guarantee.

• How many reference angles does the shot list actually require before generation starts?
• Has anyone reviewed the vendor's own stated limitations for the specific model version in use?
• Who is responsible for re-anchoring a generation once drift is spotted mid-sequence?

Read together, the two vendors agree more than their marketing copy suggests: a reference image narrows a generation, and neither company's own documentation describes it as locking one.