Reddit user shares Qwen 2.1 realism workflow: CFG 3-3.5 plus dual LoRAs beat default settings
Any_Tea_3499 · reddit · 2026-10-08
- The author argues many dismiss Qwen 2.1 because they only try the default Comfy workflow; the model performs poorly at CFG 1 and needs CFG 33.5 to shine
- Setup: Lenovo LoRA at 0.50.7 strength for realism (optional), Viggle Turbo LoRA at low 0.5 strength with 16 steps; V1 works best with ResMultistep + BongTangent, while V3 requires Euler
- Generating at 2MP; character LoRAs trained on Qwen 2.1 with AI Toolkit
- Known weakness: occasional background detail corruption. Full workflow file and Viggle LoRA download link shared.
More from Multimodal
- Creator goes from 60 followers to viral AI anime short film in three months using Vidu — CurieuxExplorer · 2026-10-08
- Roblox founder: GOTY used to take studios, now a prompt and a weekend — CurieuxExplorer · 2026-10-08
- ComfyUI PR may break custom nodes: Krea2 garbled output, MinimaxH3 OOMs — VegetableTicketO2 · 2026-10-08
- Midjourney style prompt paired with opus 5.5 animation demo — sergeantsref · 2026-10-08
- AI virtual singer Desiree's breakout track remade with Suno v6 in slopcore style — AIandDesign · 2026-10-08
- The LoopHole Project opens submissions for collaborative AI film, deadline October 21 — LudovicCreator · 2026-10-08