MAJORFailure Reference

Runway ML Hallucinated Text — Pre-Generation Risk Reference

Technical Classification

Glyph Synthesis & Semantic Adherence Failure

Glyph Synthesis & Semantic Adherence Failure occurs when the diffusion model attempts to render in-scene text (signs, labels, screens, t-shirts, subtitles) but produces shapes that resemble letters without forming actual readable characters. Common manifestations: text that looks like Cyrillic-but-isn't, mirrored or flipped glyphs, letterforms that change frame-to-frame within the same word, and substitution of the prompted text with unrelated characters. Text generation is among the highest-fail categories in video diffusion models because each frame must independently produce coherent typography while maintaining temporal consistency.

How to identify this failure

✕Storefront sign reads as garbled non-Latin characters
✕Same word morphs into different letters frame-to-frame
✕T-shirt or product label text is unreadable
✕Subtitles show wrong words or scrambled glyphs
✕Letters mirror, flip, or stretch unnaturally

Real generation examples

Prompt used

"A storefront with a neon sign that reads OPEN 24 HOURS"

Failure observed @ 0:00 - 0:05

Sign reads "QPFN 2H HOIIRS" with shifting glyphs frame-to-frame

Prompt used

"Person wearing a t-shirt that says HELLO WORLD"

Failure observed @ 0:02

T-shirt text renders as "HFLLU WURIY" with mirrored second L

Documentation strength

If you need to escalate

HIGH — Hallucinated text is an explicit prompt-adherence failure and is one of the easier categories to get acknowledged if the prompt specified the exact text.

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 Glyph Synthesis & Semantic Adherence Failure 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 Glyph Synthesis & Semantic Adherence Failure 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 "Glyph Synthesis & Semantic Adherence Failure". 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

How does Runway support handle for garbled text in a generation?

Yes, especially when your prompt specified the exact text that was supposed to render. Use the technical term "Glyph Synthesis & Semantic Adherence Failure" and quote both the prompt-specified text and what was actually rendered, with timestamps.

Why does Runway hallucinate letters?

Video diffusion models are trained primarily on natural imagery — text in their training set is sparse and inconsistent. Each frame is denoised semi-independently, so even when the model produces letter-like shapes, they rarely form coherent words and rarely stay stable across frames.

How do I get Runway to render text correctly?

Short text (1-3 letters) and large display formats have a slightly higher success rate. Long text is currently unreliable on Gen-4. AVA's L1 scanner flags any prompt containing quoted text as high-text-failure-risk so you can adjust before spending credits.

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