news

WeChat's WeLM-Image advertises a 4.98-second image, and its own table lists the same model at 13.17 seconds

The technical report's headline latency is one column of a seven-column table whose far end runs 2.6x slower on the distilled model and 100.18 seconds on the undistilled one.

By Priya Shenoy · 8 November 2026

WeLM-Image's abstract sells a single speed number: a 1024x1024 image in 4.98 seconds on one NVIDIA H20, using eight guidance-free steps. The report's own efficiency table lists seven numbers for that configuration, not one. The 4.98-second entry is the third column, and the far end of the same row, the editing case with the most input images, reads 13.17 seconds. The undistilled 50-step model runs the same spread from 32.89 seconds to 100.18 seconds. Every figure comes from the authors, and only the most favourable one made the summary.

The model itself is a unified image foundation model covering text-to-image generation, instruction-guided editing and multi-reference synthesis, described as designed for single-GPU deployment in WeChat's large-scale consumer applications. Text, reference images and target latents are processed as one interleaved sequence rather than through separate pipelines.

31.7B parameters stored, about 5.4B activated per latent token

The backbone is a dual-path Mixture-of-Transformers: shared attention, but independent understanding and generation parameters, which the authors frame as a way to limit optimization interference between the two jobs. Each path carries its own Mixture-of-Experts feed-forward layers. That gives 31.7B stored parameters with roughly 5.4B generation parameters activated per target latent token. Training is bilingual and progressive, reaching native 1.5K resolution, and the fast variant, WeLM-Image-Turbo, comes from MeanFlow-based few-step distillation down to eight steps.

Nine benchmarks, every score run by the authors

The evaluation section reports on GenEval, PRISM-Bench, OneIG-Bench, TIIF-Bench, DPG-Bench, Qwen-Image-Bench, WeGenBench, LongText-Bench and CVTG-2K. The headline row for the full model reads 0.851 on GenEval, 0.534 on OneIG, 7.555 on PRISM, 0.931 on LongText, 54.155 on Qwen-Image-Bench and 0.948 on WeGenBench. The Turbo variant lands close behind on most of them, at 0.836, 0.520, 7.482 and 0.948 respectively.

The more interesting line is where Turbo wins. On Qwen-Image-Bench the eight-step distilled model scores 55.992 against the 50-step model's 54.155, and on LongText-Bench it posts 0.933 against 0.931. A six-fold cut in sampling steps producing a higher score on two of six reported boards is either a genuine property of the distillation or a sign that those two boards do not separate these two models. The report does not say which, and no confidence intervals are attached to any of the figures.

Text rendering carries a lot of weight in that lineup, with two of the nine benchmarks (LongText-Bench and CVTG-2K) measuring nothing else. It is also one of the categories we score directly on the models we track.

GenEval is doing work that other 2026 image papers say it cannot do

Two of the benchmarks on WeLM-Image's list have been publicly questioned this year by other image-model teams. The i1 report notes that GenEval has been found misaligned with human judgment and poorly correlated with human-perceived capability, and that the common practice of fine-tuning on BLIP3o-60K inflates GenEval scores without moving other benchmarks, which is why it reports GenEval "only for completeness." Other recent reports attach similar disclaimers to DPG-Bench. WeLM-Image reports both without that caveat, and claims rank positions on GenEval sub-scores (second-highest on single-object generation, third on attribute binding among open-source models) that inherit whatever noise those critiques describe.

What the report does not contain

There is no weights release, no API price, no availability date for use outside WeChat, and no third-party arena placement. All nine benchmark results are the lab's own runs, which puts WeLM-Image in the same position as FLUX 3 Image Klein, whose first-place claims were also scored in-house. Until an independent board rates it, the only externally checkable figure here is the architecture count, and the most-cited speed figure is one point on a curve the report itself publishes in full.

Sources: arXiv, WeLM-Image technical report, Hugging Face, i1 report.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.