Flux.2 Klein 9B identity drift: ReferenceLatent breaks down in action scenes
RecordEmbarrassed787 · reddit · 2026-09-07
A filmmaker building a ComfyUI pipeline with Flux.2 Klein 9B via the fal.ai API (reference image → 1MP upscale → VAE Encode → ReferenceLatent conditioning → sampler) reports that character identity dissolves whenever prompts introduce combat scenes or dramatic lighting — native ReferenceLatent can't anchor facial features. With a local RTX 5090 Laptop 16GB, they're weighing face masking before VAE encode, ReferenceLatentPlus, or abandoning the API for a local Flux-PuLID/FaceDetailer pass. No custom LoRA yet.
Related event: Flux.2 Klein 9B Faces Identity Drift in Cinematic Pipelines(2 posts)→
More from Multimodal
- Sheet Music to MIDI Pipeline Offers a More Structured Space for Music Models Than Audio — felpix_ · 2026-09-07
- Seen2Scene open-sourced: flow matching trained on real partial scans for 3D scene completion, ECCV 2026 — angelaqdai · 2026-09-07
- TAIKEN: a 4-minute AI action film made with MiniMax H3 and ChatGPT — bennash · 2026-09-07
- Midjourney watercolour-style generations showcased — davidhoang · 2026-09-07
- Wuxia "Netherworld Divine Palm" AI video made with open-source MiniMax H3, prompt shared — Low_Loquat_4035 · 2026-09-07
- ComfyUI-Easy-Use adds Pass or None node for conditional value routing in workflows — altoiddealer · 2026-09-07