Researchers propose a training-free way to reduce repeated work in image generators: merge image tokens only when their predicted content has settled and appears structurally simple. The method, called DaTMe, tries to preserve details that text prompts may not mention, such as background boundaries. 
Why prompt-focused merging can lose detail
Diffusion models build an image through repeated denoising steps. Transformer-based image generators process the image as many small representations called tokens, and repeatedly updating every token costs computation. Token merging reduces that work by combining tokens so later steps handle fewer of them.
The challenge is choosing what to combine. Earlier approaches can protect areas that match the prompt, such as a named object, while overlooking unmentioned structures in the background. If a token still represents a changing edge or a fine boundary, merging it with its neighbor can blur the distinction.
DaTMe checks change and structure
The authors' method looks at two signals. It compares each token's estimated clean-image content across neighboring denoising steps to see whether that part is still changing. It also measures local detail, including edges and texture. Tokens that are changing or structurally detailed are kept as anchors; stable, locally uniform tokens become candidates for merging with nearby tokens.
This lets the method decide from the image-generation process itself rather than from prompt relevance alone. The authors describe DaTMe as training-free: it adds a selection step to existing diffusion models without training new model weights.
Reported speed gains come with quality trade-offs
In the paper's PixArt-alpha test, the researchers used 10,000 image-prompt pairs from the COCO 2014 validation set. At the 0.5 token-merging setting, DaTMe reduced reported generation latency from 1.9417 to 1.6884 seconds per image on a single RTX PRO 6000 Blackwell. Its Fréchet Inception Distance (FID)—a comparison of feature distributions for generated and reference images, where lower scores mean greater similarity under this metric—was 43.42; the uncompressed model scored 42.10. That means the measured speed gain came with some change in this image-quality score, rather than preserving the baseline exactly.
At the same setting, the authors report FID scores of 52.29 for a classifier-free-guidance-based merging method and 94.89 for ToMeSD, compared with 43.42 for DaTMe. Results were not identical across models: on FLUX.1-dev, the 0.5 setting reduced latency from 14.601 to 13.271 seconds, while FID shifted from 30.61 to 31.34. These are the authors' benchmark results, not an independent replication; the paper also tests Stable Diffusion 1.5 and SDXL in supplementary experiments.
The paper says it was accepted at NeurIPS 2026 and links an implementation, but it remains an author-reported method whose speed and image-quality trade-offs depend on the model and compression setting.
Sources
- The DaTMe paper and benchmark tables on arXiv
- The authors' DaTMe code repository
- The paper introducing Fréchet Inception Distance
- Illustrative denoising image: Wikimedia Commons file page and license
Image credit: "Example of Denoising Diffusion models" by MrAlanKoh, Wikimedia Commons, CC BY-SA 4.0. The image illustrates diffusion denoising generally; it is not from the DaTMe study.