QwenImage2.1 Viggle LoRA tested: 6 steps is the sweet spot, 40 steps unnecessary
reeight · reddit · 2026-09-23
Detailed step-count comparison for QwenImage2.1 with the official 1MB Viggle turbo LoRA (open-sourced on Hugging Face): the LoRA is usable at 6 steps but worse than baseline at 4; the plain model is fully usable between 12 and 25 steps at CFG 2.1 — 40 steps is unnecessary. Extras: CFG >1 gives a free quality boost, Qi2.1 defaults to Chinese-looking faces, and it's in the same ballpark as Zimage and Krea2.
Related event: QwenImage2.1 Viggle Turbo LoRA: 6 Steps Work, 4 Fall Apart(3 posts)→
More from Multimodal
- Generating accurate 3D building models from 150 planning PDFs with new Gemini model — fofrAI · 2026-11-11
- Stanford's DyCraft Uses Agentic Decomposition to Cut Physical Errors in Long-Horizon 4D Scene Generation — StanfordUniversity · 2026-11-11
- Midjourney SREF pick: --2523103199 for loose pencil-sketch portraits — tisch_eins · 2026-11-11
- H3 Image Tips: 16MP via ComfyUI Workflow Plus Detailed Prompts Yields Best Results — gatortux · 2026-11-11
- Gemini-generated endless loop of quantum cryostats self-assembling — hapticdata · 2026-11-11
- Dad brings his daughter's crayon drawings to life with AI for $1.75, and it's unsettling — Illustrious_Ad_4292 · 2026-11-11