Text-to-image foundation model family (1B, 2B, 5B) using Semantic-First Diffusion with separate semantic and texture streams.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
SeFi-Image (also referred to as CFI image) is a text-to-image foundation model from the SeFi-Team, described in an arXiv preprint (2606.22568, 2026). It applies Semantic-First Diffusion (SFD), which separates high-level structure from texture details. The semantic stream denoises ahead of the texture stream by a small timestep offset, so texture generation follows an already-formed structural trajectory.
Generation runs in three stages: semantic initialization, asynchronous joint denoising with semantics leading, and texture completion before decoding. The model comes in 1B, 2B and 5B variants. The project page says the 5B model was trained with 125K A800 GPU hours and reports results on GenEval, DPG-Bench, LongTextBench, CVTG-2K and OneIG (English and Chinese). Showcases cover scenes, text-rich layouts such as posters and menus, anime characters, varied styles and portraits.
The project page links no code, weights, demo or license at the time of the provided material, so availability and licensing are unconfirmed. A YouTube reviewer described it as a small open-source generator that is fast and beats larger open models on benchmarks, though its quality trails the leading models.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Small open-source image generator in 1B to 5B variants that separates layout and texture streams. Outperforms larger open models on benchmarks and is fast, though quality trails the leading models.
Closest neighbours by meaning, not just by category.
Model · Princeton
Fully open 3B text-to-image diffusion model from Princeton with released weights, training code, and data pipeline.
Model
One- and two-step image generation model distilled from Z-Image-Turbo using text-stream adversarial distillation.
Model
Open-weights 6B text-to-image model family from Alibaba's Tongyi-MAI lab, with a fast 8-step Turbo variant.
Model
Rank-asymmetric flow method for pixel-space image generation, incl. AsymFLUX.2 klein finetuned from FLUX.2 klein 9B.
Model
Pixel-space text-to-image diffusion model transferred from latent diffusion models, without a VAE, supporting 4K output.
Model · Bria AI
Open-weights text-to-image model from Bria AI, fine-tuned from FIBO, supporting structured JSON prompts and image editing.
