AI workflows
Construct reusable character stills with Runway References
Runway Gen-4 Image References builds still images from selected reference assets. Save and name the assets, use their names to assign roles, and approve the resulting composition before animating it. Image construction and video generation are separate stages, with different models and different things to inspect.
In this guide
- Name inputsSave identity and scene references with names that explain their roles.
- Compose a stillSelect the active references and describe each asset’s role with @ names.
- Approve the imageInspect both people, the setting and the intended starting pose.
- Animate separatelyUse an accepted still in the documented video stage and review the motion.
Start in Gen-4 Image rather than treating References as a video model
The verified References procedure belongs to Gen-4 Image and Gen-4 Image References. The separate current video guide describes Gen-4.5. These names identify different stages here: first construct a still, then decide whether to use that approved still as the input for a moving shot. Do not treat a saved reference as a finished video performance.
The image stage solves composition questions. Can both intended partners occupy a sunlit memory kitchen? Is the breakfast table visible? Does the framing leave both faces and a raised cup readable? Answer these before adding motion. If the still already assigns the wrong face to the seated person, the motion stage begins with an unresolved image problem.
The workflow below was checked against official documentation on October 15, 2026. The older model-specific guide and newer general reference guide use different labels for saving reference details. Follow the control shown for your selected model rather than assuming every screenshot has the current wording.
Assign each active reference a different job
Plan the reference set before loading it. The Gen-4 Image References guide documents up to three active references for an image generation. For a two-person composition, a practical authored allocation is one reference for each partner and one for the kitchen. This is a planning choice within that model-specific limit, not a universal rule for every Runway model.
Give each reference a job you can explain. The first guides one partner’s appearance, the second guides the other partner and the third guides the kitchen setting. If a room reference contains an unrelated person, note that ambiguity before using it; your prompt should not silently ask one asset to stand for both a room and an identity.
Keep the original files separate from the reference library. An accepted generated room view or alternate character angle may become useful later, but it remains generated material. Record that origin so it cannot gradually replace the original person as your only comparison source.
| Active reference | Assigned job | Approval question |
|---|---|---|
| @mira | Partner beside the breakfast table | Does the face match the intended person? |
| @theo | Partner holding a cup | Is the second person independently recognizable? |
| @memorykitchen | Breakfast table and kitchen character | Does the composition support the planned memory scene? |
Load references and save names for later sessions
Open Tool mode, choose Image and select Gen-4 Image or Gen-4 Image Turbo as documented. Enable References. Load a reference from Assets or drag a file into the prompting canvas. Check the file you actually loaded before giving it a role; similar filenames and thumbnails can hide a wrong selection.
Save the reference with a clear name for reuse. The Gen-4 Image References guide describes tagging to save, while the newer general guide describes Reference details. The important distinction is between a temporary unnamed upload and a named saved reference. The documentation says unnamed uploads are temporary to a session, while saved named references can be reused across sessions.
Record the name beside the original filename. Use distinct names that do not change meaning between shots. For example, memorykitchen identifies a setting; survivor identifies a story role that might move between people. A person's stable name is usually less ambiguous for an identity asset than a role that can change during the film.
Check the reference media supported by the selected model. A general guide covering image, video and audio references does not make those types interchangeable in every model. This procedure uses image references for still construction.
Write the image prompt around asset roles
Select the references you intend to use and mention their @ names in the composition request. State which character occupies which role and how the setting should be used. The prompt should describe a still you can inspect, rather than asking the image stage to perform the whole embrace over time.
The proposed example below creates a quiet breakfast-memory composition. It is original editorial material, not a tested image. Substitute your actual saved names. If the accepted kitchen reference places the table elsewhere, resolve that composition choice rather than describing furniture that contradicts the reference.
Proposed reference-based still
Use @memorykitchen for the sunlit breakfast setting. Place @mira beside the table with one hand resting near a plate. Place @theo at the adjacent side, visibly gripping a cup in one hand below shoulder height. Show both faces and keep the cup clear of both faces and Mira’s hands. Make a quiet memory still before the cup is raised.
Inspect the generated still before saving it as another reference
A correct @ name in the request is not evidence that the generated face is correct. Compare each person independently with the original source. Then inspect their relationship in the composition: who stands beside the plate, who holds the cup and whether any clothing cue has transferred to the other person.
Inspect the hands and shoulders needed for the next action. A future cup lift needs a plausible visible arm. A slight lean across the table needs room to move. If the still fails those conditions, revise the image stage before animation. Motion prompting should not become an indirect attempt to fix an already incorrect opening pose.
The documentation allows an accepted generated output to be saved as a new reference when building a shot set. Use that after approval, and record what the new reference represents. It may preserve the room composition or an approved character appearance. Do not silently treat it as a new original photograph of the person.
- Compare each generated face with its intended source independently.
- Verify the plate-side and cup-holding roles against the prompt.
- Inspect the arm and body space needed by the next action.
- Label saved generated outputs by their origin and intended use.
Build related stills with deliberate differences
For a small shot set, decide which facts remain stable and which change. The kitchen wide view and a closer table view may share the people and room while using different framing. A later ruined scene requires a deliberate story-state change. Document that change rather than assuming the same reference set should produce both states without direction.
Change one compositional question at a time while the identity references remain stable. If both a person and the room look wrong, start with the error that prevents you from judging the other. Save the accepted still with a descriptive filename and the reference names used. This creates a usable handoff to motion without requiring a complex asset system.
If you use the documented sketch workflow, remember that exported sketches cannot be edited in place under that procedure. Keep the editable planning material and label exported copies so you know which version produced an accepted composition.
Animate an approved still in a separate stage
Once a still passes, follow the current Gen-4.5 video guide for the available image-to-video operation. The image supplies the initial visual scene; the motion request describes what should change. For the kitchen setup, a small cup lift can be evaluated before attempting a larger shared action.
Review the resulting video against the accepted still and the action brief. The fact that both characters passed the image-stage check does not guarantee that their faces remain stable during motion. Keep the image approval and moving-shot approval as separate records.
Zombie Couple currently offers a local two-photo casting preview and reference-film playback; generation and checkout are disabled. This independent Runway workflow is based on official documentation reads, without a paid render or signed-in UI test. Saved references guide generation; they are not a promise of exact identity or unchanged wardrobe.