CRITICALFailure Reference

Runway Prompt Ignored — Refund Guide

Technical Classification

Text Conditioning Collapse

Text Conditioning Collapse on Runway Gen-3 occurs when the text-encoder embedding fails to influence the diffusion process, resulting in output that ignores the prompt entirely. The model produces visually plausible motion — but the subject, action, environment, and style bear no relation to the input text. Distinct from partial prompt adherence: this is total collapse, where the output is indistinguishable from an unconditional sample.

How to identify this failure

✕Output subject completely different from prompt subject
✕Action described in prompt is absent
✕Scene / environment unrelated to prompt
✕Style descriptors (cinematic, anime, etc.) ignored
✕Output resembles unconditional generation from the model

Real generation examples

Prompt used

"Red sports car drifting around a mountain hairpin turn, golden hour, cinematic"

Failure observed @ full duration

Output shows a static landscape with no car, no motion, no hairpin road

Prompt used

"Anime-style girl with pink hair drinking tea in a Japanese cafe"

Failure observed @ full duration

Output is a photorealistic street scene with no person, no anime style, no cafe

Documentation strength

If you need to escalate

HIGH — Total prompt-adherence collapse is the strongest refund case. Runway support cannot defend output that ignores the input entirely.

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 Text Conditioning Collapse 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 Text Conditioning Collapse 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 "Text Conditioning Collapse". 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

Will Runway support escalate prompt-ignored generations?

Yes — Runway support recognises total text-conditioning collapse as a defect. Submit the prompt + output side-by-side and cite "Text Conditioning Collapse".

Why does Runway ignore prompts sometimes?

Gen-3 conditions the diffusion process on a text-encoder embedding. Under specific prompt structures (highly stylised, multi-clause, contradictory adjectives, or rare vocabulary), the embedding can collapse to a near-zero vector, causing the diffusion to produce an unconditional sample.

How do I avoid Runway prompt collapse?

Use concrete subject + action + environment + style in that order. Avoid contradictory adjectives ("photoreal cartoon"), rare proper nouns, and overly nested clauses. AVA flags prompt structures with high collapse risk before generation.

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