Game dev's 4-view character workflow: IP-Adapter at 0.6 overrides prompts, lower weights break consistency
coonstaantiin · reddit · 2026-09-04
A developer making chibi sprites for a mobile idle game shares an SDXL pipeline: Blender base mesh, depth maps from four 35° camera angles via ControlNet Depth, plus IP-Adapter style transfer at 0.6. At 0.6 the reference image overrides the prompt (copying pants color and flaws across views); lowering it breaks cross-view identity. They're weighing attention masks, alternative IP-Adapter modes, or a style LoRA.
More from Multimodal
- Garry Tan Impressed by Grok Image Generation, Sets Lobster-Costume Portrait as Avatar — garrytan · 2026-09-04
- Meta's muse spark 1.3 recreates Lies of P mechanical heart in 3D from screenshots — alexandr_wang · 2026-09-04
- Sam Altman's favorite GPT-6 video critiqued for missing the iconic 'Put That There' pointing interaction — tianshi_li · 2026-09-04
- Higgsfield launches Genjutsu, full-frame motion control for AI video generation — LexiLove · 2026-09-04
- Principia benchmark exposes major physics reasoning gaps in video generation models — Varun Varma Thozhiyoor · 2026-09-04
- LLaDA-Image: 6B diffusion model with Muon optimizer hits open-source SOTA image generation — inclusionAI · 2026-09-04