🚀 Alibaba's Qwen team dropped Qwen-Image, a 20B-param MMDiT model revolutionizing text-to-image gen! Native in-pixel text rendering for stunning posters, bilingual EN/CN support, excels in photorealistic/anime/styles. Dive into our technical article with full guide to deploy the model on Hyperbolic & Gradio

Architecture: Combines MLLM (Qwen2.5-VL 7B for semantics), VAE (fine-tuned for text-rich recon), & 20B MMDiT (flow matching w/ ODEs, diagonal concat for scalable res). Process: Prompt → feats → denoising → decode. TI2I w/ dual-encoding for edits.

Innovations: Massive data pipeline (billions pairs: Nature 55%, Design 27%, People 13%, Synthetic 5%; EN/CN splits). Curriculum learning for text mastery. MSRoPE (on RoPE) for 2D alignment. Multi-task T2I/TI2I/I2I. SOTA on GenEval, text benches!

Vs. GPT-Image-1: Matches photorealism, crushes bilingual text/multi-line, editing consistency (better fidelity in objects/poses). This is the edge of Open-source vs. API!

GPU infra: ~24GB VRAM est. (20B x 1.2 in BF16). Inference runs smoothly on a single H100. We tested on Hyperbolic's On-Demand Cloud H100 for $1.49/hr, with a simple python script adapted from the official model card for an interactive Gradio UI

Read our full article: Architecture details, innovations, comparison, compute analysis, COMPLETE code & deploy steps on Gradio. Run Qwen-Image yourself and share your image creations with us! Read the full blog: Rent H100s now on Hyperbolic for $1.49/hr:

Read our full article: Architecture details, innovations, comparison, compute analysis, COMPLETE code & deploy steps on Gradio. Run Qwen-Image yourself and share your image creations with us! Read the full blog: Rent H100s now on Hyperbolic for $1.49/hr: