Dev runs Qwen-Image 2.1 fully on a 12 GB RTX 3060, loses honest 20-prompt bake-off to gpt-image-2
Sg1000deping · reddit · 2026-09-24
A developer released a free Windows desktop app and CLI (qil, MIT-licensed) that runs Qwen-Image 2.1 entirely in VRAM on a 12 GB card (RTX 3060), generating 1024² images in 35 seconds.
How it fits in 12 GB: the text encoder (8.9 GB), DiT (6.9 GB) and VAE (0.6 GB) can't coexist in VRAM, so each stage is loaded wholesale when it runs while others wait in RAM, using Comfy-Org's int8 "convrot" weights. One gotcha: CUDA 12.8 PyTorch builds silently disable ComfyUI's optimized CUDA backend — cu130 fixes it.
Measured: 32–35 s at 1024²/20 steps, 11.4 GB peak VRAM, 7 min for native 2K at 40 steps.
Honest 20-prompt comparison vs gpt-image-2 (no cherry-picking): Qwen 1 win, 8 ties, GPT 11 wins. Editing is on par (7 ties + 1 win in 8 tasks), but GPT clearly leads at text-to-image — Qwen follows long prompts less closely, garbles small text, and has weaker world knowledge. All prompts, images and verdicts are published; the repo builds with one command.
More from Infra
- Anthropic paid $5B for short-term compute in Q3 as spot prices run ~2x long-term deals — tengyanAI · 2026-11-11
- UK Sovereign AI invests in Fractile to build next-gen frontier AI chips — charlieharris01 · 2026-11-11
- What happened when I moved vector retrieval into Snowflake SQL instead of orchestrating RAG outside — dima806_dima · 2026-11-11
- AI chip startup Fractile raises $720M Series C for DRAM-on-processor design — Tom_Westgarth15 · 2026-11-11
- Mobile AI Factory: agentic pipeline builds on-device SLMs that beat 5-shot frontier models — kalyan_kpl · 2026-11-11
- Vivo's Starlink direct-to-cell service hits 10M users in two weeks across Brazil — XFreeze · 2026-11-11