MAJORFailure Reference

Luma Dream Machine Watermark Bleed — Pre-Generation Risk Reference

Technical Classification

Training-Data Watermark Leak

Luma Dream Machine's training corpus includes substantial stock-footage and licensed broadcast material. Under certain prompt distributions — particularly cinematic establishing shots — the model regurgitates faint watermark patterns from its training set: ghost logos in lower-thirds, faint timecode strips along the top edge, and translucent stock-agency marks across the centre. The user never asked for them and they make the output unusable for commercial work.

How to identify this failure

✕Faint translucent logo in the lower-right corner
✕Timecode strip along the top edge of the frame
✕Stock-agency name appearing as a ghost overlay
✕Watermark visible only in specific lighting conditions
✕Mark drifting in opacity across the clip

Real generation examples

Prompt used

"Aerial shot of a coastal cliff at sunset, cinematic"

Failure observed @ 0:02

Faint translucent "GETTY" logo bled into the lower-right at 0:02–0:05

Prompt used

"Slow drone push toward a city skyline at dusk"

Failure observed @ 0:00

Timecode strip visible along top of frame at 0:00–0:08

Documentation strength

If you need to escalate

VERY HIGH — training-data leakage in commercial output is unambiguously refundable; cite the leak as a deliverability defect.

AVA is a pre-purchase prevention tool, not a post-purchase recovery tool. Platforms generally do not guarantee credit refunds for output-quality failures; goodwill credits are at each platform's discretion. The strength rating reflects how well-formed your support ticket can be, not a promised outcome.

Prevention + documentation steps

01
Score your prompt before you generate
Run your prompt through AVA's pre-flight scoring against the Training-Data Watermark Leak pattern. Green light = generate. Yellow/red = rewrite using the suggested fix before you commit credits.
02
Capture Generation ID + timestamp if it failed anyway
Find the Generation ID in the URL or share link. Note the exact time when the Training-Data Watermark Leak first appears (e.g. "failure first visible at 1.2s"). Timestamped evidence is significantly stronger than a general complaint.
03
Use the correct technical term in your support ticket
Describe this failure as "Training-Data Watermark Leak". This term maps to a recognised internal workflow in the support system and routes the ticket to the right team.
04
Submit via the correct support channel
Runway has no direct email intake. Pro+ plan: open the in-app AI Assistant (help widget bottom-right of app.runwayml.com), describe the failure with the technical term, attach evidence. Free/Standard plan: human support isn't available — your channel is Discord #community-help with @On Call - Moderators.

Frequently asked questions

Does Luma support response credits for watermark bleed?

Yes. Watermark bleed is a deliverability defect that blocks commercial use. Submit the generation ID with a clear screenshot of the leaked mark; Luma support has a documented escalation precedent here.

Why does Luma produce stock-footage watermarks?

Dream Machine's training data includes substantial licensed stock and broadcast footage. The diffusion model can memorise frequent patterns — watermarks among them — and reproduce them when the latent space drifts toward those training samples.

How do I prevent watermark bleed in Luma generations?

Avoid generic cinematic establishing-shot prompts. Add specific, unusual style anchors ("hand-painted look", "anamorphic vintage glass") to push the latent away from common stock-footage neighborhoods. AVA flags high-bleed-risk prompts.

Catch it before you generate

AVA scores this failure mode against your prompt in real time

Free Chrome extension. Analyzes your prompt as you type, flags failure-prone patterns specific to this model, and tells you what to rewrite — before you commit credits to a generation that will fail.

Install free All 105 failure modes

AVA Pro · founders' round

$50 for 6 months of unlimited scoring across all failure modes + personal failure-history dashboard. Locks in $13/mo grandfathered after.

Claim $50 founders