[Release] Visage-Image - a small text-to-image model for face generation
retained text
https://preview.redd.it/0wzb4mbs8gxh1.png?width=1280&format=png&auto=webp&s=44d9cbb4ccb9968c0ea54508b88e8d58413f2f79 Hi everyone! I've released my very first text-to-image model for face generation. I collected my own dataset and captioned it with Qwen3.5-4B and a bunch of other face-analysis models. I didn't like the results I got with the FLUX.2 VAE, so I decided to train my own narrow-domain autoencoder. It turned out to beat FLUX.2 on iFID on my dataset, and in my opinion, on generation quality too. The Qwen-VAE-2.0 tech report helped hugely with that - getting to FLUX.2 quality was hard. Overall, it's a basic single-stream DiT with a rectified flow objective and a couple of modern tricks from recent papers. The model supports different ages, ethnicities, poses, gaze directions and expressions, though only a limited set of expressions works reliably. You can read about model architectures, training regime and data preparation on github. I'd be happy to hear any feedback, and I hope this release inspires more models! HF: [https://huggingface.co/Sherstnev/Visage-Image](https://huggingface.co/Sherstnev/Visage-Image) GH: [https://github.com/Nikita-Sherstnev/Visage-Image](https://github.com/Nikita-Sherstnev/Visage-Image) ComfyUI custom nodes and workflow: [https://github.com/Nikita-Sherstnev/ComfyUI-Visage](https://github.com/Nikita-Sherstnev/ComfyUI-Visage) (or search "Visage" in ComfyUI Manager) [Expression examples](https://preview.redd.it/397rw5509gxh1.jpeg?width=1504&format=pjpg&auto=webp&s=c111eaa1232977efaaebbcb99f3e386aa1b7267f)
document fetched 2026-10-25 00:29 · canonical https://reddit.com/r/StableDiffusion/comments/1xehgeb/release_visageimage_a_small_texttoimage_model_for