RTX 3060 shootout of five image models: Qwen Image 2.1 trails ERNIE and Krea 2
SirMick · reddit · 2026-09-28
The author tested Qwen Image 2.1 against ERNIE-Image-Turbo, Krea 2 Turbo, HiDream-O1-Image and Z-Image-Turbo locally in ComfyUI on an RTX 3060 12GB, using identical prompts plus harder tests for characters, hands, object relations and scene geometry.
- ERNIE: most consistently strong image quality
- Krea 2: best balance of quality and speed
- HiDream: striking compositions, less reliable fine detail
- Z-Image: fastest iteration
- Qwen Image 2.1: interesting global compositions but recurring structural issues with hands, straps, object placement and local geometry
They also tried the base model at 2048×2048 (more detail, same structural flaws) and the Pruna 8-step workflow (much faster, noticeable quality drop). Puzzled by the model's recent popularity, the author asks the community for better sampler/scheduler, step count, CFG, quantization or LoRA setups, especially from 12GB GPU users.
More from Multimodal
- How to Recreate the Viral AI Halloween Door Video — motionbynick · 2026-09-28
- Style Atlas: a gallery of 1,337 art styles with copyable prompts for AI image generation — mhmazur · 2026-09-28
- Qwen Image character sheet workflow: one reference image to a full design sheet — LeebLaab · 2026-09-28
- Moiré Pattern Interference prompt: overlapping fine lines that fake motion and depth — LudovicCreator · 2026-09-28
- Fighting MiniMax H3 in ComfyUI: static camera, clean background, full-body room — Kassiber · 2026-09-28
- A production-line workflow for pro motion design with Opus 5.5: storyboards first, render last — sven_ai · 2026-09-28