CRITICALFailure Reference

Luma Dream Machine Prompt Adherence Failure — Refund Guide

Technical Classification

Text Conditioning Drift

Text Conditioning Drift on Luma Dream Machine occurs when the text-encoder conditioning fails to steer the diffusion process toward the prompt-specified subject, action, or scene. Unlike total collapse, Luma drift typically produces output that contains some prompt elements but discards others — a partial-attendance failure where the model commits to a different interpretation than the input requested. Most common on multi-element prompts with rare vocabulary or specific style descriptors.

How to identify this failure

✕Wrong subject (cat instead of dog, etc.) despite explicit prompt
✕Action described in prompt is absent or replaced with generic motion
✕Scene environment is wrong (indoor instead of outdoor, etc.)
✕Style descriptors (anime, watercolor, etc.) are ignored
✕Multi-subject prompts collapse to single-subject output

Real generation examples

Prompt used

"Watercolor-style golden retriever running on a beach at sunset"

Failure observed @ full duration

Output is photorealistic, the dog is standing not running, scene is in a park not beach

Prompt used

"Two children flying a kite in an open meadow, hand-drawn animation style"

Failure observed @ full duration

Output shows one child standing still in a forest, photorealistic

Documentation strength

If you need to escalate

HIGH — Prompt-adherence failure on paid output is a clear feature defect. Luma support honours refund tickets that document specific divergence.

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 Drift 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 Drift 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 Drift". 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 prompt-adherence failures?

Yes — submit the prompt and the output side-by-side, listing each prompt element the model discarded. Cite "Text Conditioning Drift". Luma support recognises this as a generation defect when documented specifically.

Why does Luma ignore parts of my prompt?

Luma's text encoder maps multi-clause prompts to a single conditioning vector. When clauses conflict in style (e.g. "watercolor + photorealistic"), describe rare subjects, or contain too many independent elements, the vector collapses toward the model's training prior — discarding the unusual elements.

How do I structure Luma prompts to maximize adherence?

Use single-clause prompts with subject + action + style in that order. Avoid contradictory descriptors. Test rare-vocabulary subjects with a reference image when possible. AVA pre-flights Luma prompts for drift 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