One reference image, no LoRA: keeping a character consistent across 8 scenes with SnapRef
PoleRacers · reddit · 2026-11-20
The author tested Anima's SnapRef v2a in character mode, using a single head-and-hat crop as the only reference — no LoRA training — to generate 8 scenes of the same OC (white hair, red eyes, pointed ears, purple lipstick). Full settings shared: 1024x1536, ersde simple, 36 steps, CFG 4, ref guidance 4.5, weight 1, Detailer LoRA 0.25, @wlop style tag.
What held: face, hair, ears and hat across different lighting, profile, back view, and even a flat cel style; the reference's outfit never leaked in.
What didn't: "hat off" worked 2 of 3 times; full-body shots soften the face; an open-mouth laugh turned into screaming in all 3 seeds (a softer smile fixed it); painterly styles drift rounder while flat style kept her sharpest.
Author concludes a single reference goes further than expected and asks how SnapRef compares to the v2 ref adapter.
More from Multimodal
- VLM turns car photos into sprite-stacking 3D assets with a single plain prompt — CSProfKGD · 2026-11-20
- Meta unveils AVM: autoregressive view model streams 2D views as you steer a virtual camera — JonathonLuiten · 2026-11-20
- Kitten TTS 2 Mini Trends on Hugging Face: Lightweight Multilingual Open TTS — KittenML · 2026-11-20
- World Snapshot Models unify 1,000+ visual tasks as generation — World-Snapshot · 2026-11-20
- Character LoRA merges pollute art style? Illustrious user asks for a fix — Maximum_Cause4444 · 2026-11-20
- InfiniGen: a Windows local AI image/video studio wrapping ComfyUI, with AI shot planning — methodangel · 2026-11-20