Should an 8GB-VRAM Flux LoRA + WAN 2.2 Workflow Retire for Newer Models?

wreck_of_u · reddit · 2026-08-20

A creator shares their low-cost pipeline: building a 20-image 512x512 dataset with hand-written captions, training a Flux1 character LoRA on rented vast.ai (or Replicate's fast trainer), then running ComfyUI on an old headless rig (Ryzen 1700 / 32GB RAM / RTX 2070 8GB) with the full Flux1-Dev safetensor to generate 1024px images at 1 min each; favorite outputs become 7-second WAN 2.2 clips with inference on paid Replicate.

They enjoy the "leave it alone" workflow but ask whether newer models still support character LoRAs with high output variation — noting z-image-turbo feels too "same same" under identical prompts — and whether models like Krea 2 can run on 8GB VRAM.

Original post →

More from Multimodal

Multimodal channel →