You write “Sarah’s apartment, living room” in three prompts. Clip 1 is a modern minimalist space. Clip 2 has antique furniture and warm walls. Clip 3 is bright with mid-century chairs. Same character, supposed same room — three different rooms.
Scene consistency across cuts is the hardest multi-clip problem after character consistency. Models invent set dressings on the fly because the same prompt can describe thousands of rooms. To get consistency, you need a visual anchor of the room, not a verbal description.
Common causes
Ordered by hit rate, highest first.
1. Free-form scene description, no anchor
Sarah's apartment, living room → model invents one. Three clips, three inventions.
How to spot it: prompt describes the scene in text but you have no reference image of the established scene.
2. No establishing shot generated first
You’re generating action clips before establishing what the room looks like. Each clip extrapolates differently.
How to spot it: there’s no wide shot of the room. You jumped to close-ups and mediums.
3. Different lighting descriptions across cuts
Clip 1: "warm afternoon light through window"
Clip 2: "cozy interior light"
Clip 3: "morning light"Each lighting prompt produces a different look — even if it’s the same time of day.
How to spot it: lighting words vary across prompts.
4. Different prop / decor descriptions
Clip 1 mentions a green sofa. Clip 2 doesn’t mention a sofa at all. Clip 3 mentions a blue one. Model is free to vary.
How to spot it: scene description text varies across clips.
5. Generations done in different sessions / tools
Same as character continuity: cross-session generations drift. Different tools = different rooms.
How to spot it: clips were generated days apart or in different tools.
6. Tool can’t enforce scene continuity strongly
Some tools (especially text-to-video without image input) have weak scene memory.
How to spot it: even with identical prompts, scene differs. Tool limitation.
Shortest path to fix
Step 1: Generate ONE canonical establishing shot first
Before any action clips:
# Generate wide shot of the room
"wide establishing shot, Sarah's living room: cream-walled modern minimalist
space, blonde wood floor, green velvet 3-seater sofa center-back, brass floor lamp
right of sofa, large window left wall with sheer white curtains, dark wood
coffee table in front of sofa, small abstract painting above sofa,
soft afternoon window light, 5500K daylight, no people, daytime"Save as scene_REFERENCE.png. This is the “set” for all clips.
Step 2: Feed the establishing shot to every action clip
# Image-to-video (best)
- Use scene_REFERENCE.png as start frame for clips set in this room
- Even if the action clip starts with a different camera angle,
the model has seen the room
# Tools without image input
- Paste the verbatim scene description into every prompt
- Same exact words for furniture, decor, lightingStep 3: Lock lighting language across clips
Pick one lighting description and reuse exactly:
# Same wording in every clip
"5500K daylight from window camera left, soft and diffused,
warm but not orange, no shadow on shadow side"Step 4: Use the same scene description across all clips
Copy a single scene block, paste into every clip:
# Scene block (paste verbatim into every clip prompt)
SCENE: Sarah's apartment, living room.
SET: cream walls, blonde wood floor, green velvet 3-seater sofa,
brass floor lamp, large window with sheer white curtains,
dark wood coffee table, small abstract painting above sofa.
LIGHT: 5500K daylight from window camera left, soft and diffused.Step 5: Generate all clips in one session, one tool
# Stick with one tool for the whole project
- Pick Runway OR Kling OR Pika
- Same model version throughout
- Same sampler / settings
- Ideally same sessionStep 6: Color match in post
Even with all the above, expect some drift. Final unification in editing:
# DaVinci Resolve
- Use first clip as reference
- Apply "Color Match" to all other clips
- Manually fine-tune outliers
# Premiere Lumetri
- Lumetri Color → Match → reference clipPrevention
- Storyboard with the establishing shot at the top: render it first, reference it always
- Write a “scene block” once and paste verbatim into every clip prompt
- One tool / one session per scene; never mix
- Always color-match in post even when generation looks consistent