MER-Bench Exposes How AI Models Fail to Rewrite Memes Without Breaking Them
A new benchmark asks image-text models to flip the emotional tone of memes while keeping the joke's layout, characters, and scenario intact.
- MER-Bench introduces Meme Reappraisal: flipping a meme's emotion while keeping scene, characters, and layout intact.
- Task inspired by cognitive reappraisal in psychology, framed as constrained multimodal generation.
- Benchmark annotates source and target emotions, rewritten captions, visual edit specs, and taxonomy labels.
- Evaluation uses an MLLM-as-Judge across generation quality, affect control, structural fidelity, affective alignment.
- Tested systems including BAGEL-7B-MoT, DreamOmni2, FLUX.2-klein-4B, Z-Image-Turbo show large gaps on all three constraints.
- Code, data, and pre-generated outputs available at github.com/one-seven17/MER-Bench; accepted to EMNLP 2026.
Memes expose a difficult weakness in multimodal image editing: the picture and caption form a single semantic unit, so changing either one can destroy the joke. A new benchmark called MER-Bench asks image editors and multimodal generators to turn negatively framed memes into constructive versions while preserving the original scenario, entities, and layout.
The researchers call the task Meme Reappraisal, drawing on the psychological technique of changing an event’s emotional interpretation. Applied to memes, reappraisal resembles emotion-level style transfer under strict visual and textual constraints. The output must express a new emotional frame while remaining recognizably derived from the source.
A small edit with tight constraints
Conventional image editors can repaint scenes, replace objects, or apply visual styles, while multimodal generators can create captioned images from prompts. Meme reappraisal requires several capabilities to work together:
- The caption must remain readable and fit the existing text region.
- The characters, objects, and underlying scenario must survive the edit.
- Single-panel or multi-panel geometry must remain coherent.
- The rewritten text and edited image must express the target emotion together.
A broad instruction such as make this meme positive
can trigger extensive regeneration, producing a different scene or template. MER-Bench measures whether a model can make the narrower emotional change without semantic drift, broken text, or disrupted composition.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
