Title: Struggling with face consistency in AI festival photo generator — moved past Nano Banana 2, what API is giving you 90%+ face accuracy?*

via

StableDiffusion

Building a product where users upload their photo, pick a festival-themed template (think Holi, Diwali, Coachella-style setups), and get a stylized version of themselves in that scene. The prompt is pre-written on the backend — users just select a template and upload. Currently using Nano Banana 2 for face-swap / face-consistent image generation and hitting around 70–80% face accuracy — which is honestly too low for a consumer-facing product. Faces are either slightly off, lighting doesn't blend

More from StableDiffusion

Vibe coded and made a Knights of New Order like free open sourced tool for proof-checking deepbooru tags

StableDiffusion·3

The greatest amateur photo realism I have ever achieved.

StableDiffusion·3

FastSDCPU release v1.0.0-beta.301

StableDiffusion·3

Side-by-side comparison of Qwen-Image, ERNIE Base/Turbo, and FLUX.2 Dev across 8 custom styles (single RTX 5090)

StableDiffusion·7

Bridged Compositing Example

StableDiffusion·8

I released a new LTX-focused update for Deno Custom Nodes for ComfyUI.

StableDiffusion·7