ERNIE-Image release announcement — Baidu / PaddlePaddle 8B DiT on Hugging Face (Adina Yakup)

TL;DR

Adina Yakup (Hugging Face) signal-boosts Baidu / PaddlePaddle releasing the ERNIE-Image series on Hugging Face: an 8B DiT for text-to-image and an ERNIE-Image-Turbo variant, both under Apache 2.0. The tweet highlights strong text rendering for posters and UI-style images, plus structured outputs (comics, multi-panel scenes). This is the generative half of the ERNIE-Image collection whose aesthetics-scoring sibling, ERNIE-Image-Aes, was filed earlier; the collection now spans generation (ERNIE-Image, ERNIE-Image-Turbo), aesthetic scoring (ERNIE-Image-Aes), and a 1K-image benchmark (ERIA-1K). No technical report at announcement time — the model cards and demo space are the only documentation.

Key claims

ERNIE-Image is released as an 8B DiT for text-to-image generation under Apache 2.0, with a Turbo variant alongside it [tweet body].
Headline capabilities called out: strong text rendering for posters and UI-style images, plus structured outputs for comics and multi-panel scenes [tweet body].
The release lives in Baidu’s ERNIE-Image collection on Hugging Face alongside ERNIE-Image-Aes (already filed) and the ERIA-1K-Benchmark [HF collection].
No accompanying paper or technical report is linked at announcement; only the model collection and a HF demo space are provided [tweet body].

Method

The tweet itself is a release pointer; no architectural details, training data, or evaluation numbers are provided. From the HF collection, two text-to-image checkpoints are released — baidu/ERNIE-Image and baidu/ERNIE-Image-Turbo — both as text-to-image pipelines. The “8B DiT” descriptor places it in the same parameter neighborhood as Qwen-Image 2.0 and the FLUX.2 [klein] family, and the text-rendering / multi-panel emphasis matches a trend (Qwen-Image, Recraft V4) of treating typography and structured layouts as first-class generative targets.

Results

None reported in the tweet or model card excerpts available at filing. The Turbo variant has higher download counts (11.6k) than the base ERNIE-Image (10k) at the time of fetching, suggesting community interest in the faster path.

Why it’s interesting

Adds another datapoint to Open foundation-model releases: an 8B Apache-2.0 DiT from a major Chinese lab, sibling to the previously filed ERNIE-Image-Aes: Robust Image Aesthetics Scoring with Balanced Category Generalization aesthetics scorer — the same lab is now shipping the generation model and its scoring/curation infrastructure as a coordinated package, which is the same coordinated-release pattern Qwen and NVIDIA have been doing (separate model + dataset + benchmark, all under one banner). The text-rendering and structured-output emphasis puts it in the same competitive band as Qwen-Image-2.0 (Qwen-Image 2.0) and Introducing Recraft V4: Design Taste Meets Image Generation (Recraft V4) — both of which similarly treat poster/UI/typography as the headline capability. Worth tracking whether a technical report follows, since the tweet alone exposes no training-data or benchmark detail.