Technical
The Reference Image Problem: Why AI Video Keeps Ignoring What You Give It
ยท RenderBob team
Public Adobe Community posts show Firefly video ignoring a reference photo and a detailed prompt. A 2025 paper traces that kind of drift to how models encode fine detail.
A reference still on the left and a generated frame on the right that no longer matches the subject.
You give a video model a reference image and a plain instruction. The clip comes back about something else. Adobe Community has the complaint in public. One poster wrote that Firefly ignored a detailed prompt and the uploaded reference photo. Another thread describes sequential clips where the same prompt, the same seed, and the previous frame still fail to hold the subject: the tool swaps in a different model and the shot drifts.
These examples are public Adobe Community posts about Firefly. They document a failure pattern. They are not a claim that Adobe is the only vendor that hits it.
The paper names a mechanical cause
The Consistency Critic paper (arXiv 2511.20614, submitted 25 November 2025) is about images, and the mechanism matches what the video threads are complaining about. It traces inconsistent fine detail to VAE encode and decode, and to shallow-layer information that decoder-only models drop. Text, logos, and the small identifying parts of a reference are where that loss shows. The next article covers the correction pass the authors propose.
You can pay for a result you cannot use
Drift and a failed job are different faults, and both show up on the same bill. A Community poster reported three Firefly attempts that failed with no output and still cost 400 credits each: 1,200 credits, nothing to cut with. Adobe's own community explainer says credits taken by a failed generation are supposed to return automatically. Posters keep writing in because they did not.
A timeline editor should treat the first result as uncommitted. Show it beside the reference, make a fresh attempt cheap, and leave it off the locked cut until a person has checked that it matches. The following two articles cover a research correction pass and the habits editors are using while that pass is not in the products they actually ship.
More from the blog
- Why Apple Silicon's Unified Memory Is the Real Story in On-Device AI Editing
On Apple Silicon the CPU, GPU, and Neural Engine share one memory pool, so an editor can decode, analyse, and preview a clip without copying it between separate memories.
- Consistency Critic: An Academic Fix for the Reference-Drift Problem
ImageCritic, from a November 2025 paper, corrects a generated image against its reference after the fact. Public weights and a ComfyUI node exist. Commercial video APIs do not run it for you.