People keep asking what GPU to buy for training LoRAs, but it really depends on what you’re training. SDXL and Flux.1 LoRA fine-tuning, DreamBooth for character consistency, and full model fine-tuning are on completely different levels hardware-wise. For LoRA, 16G VRAM is the sweet spot. The 4090 is the fastest consumer card—a 1500-step Flux.
1 LoRA finishes in about 20-30 minutes, so iteration feels great. On a tight budget, go for the 4060Ti 16G—it gives you plenty of VRAM headroom. The 3060 12G can handle lightweight training but quickly runs out of room. Flux itself is especially memory-hungry; with 8G you’d need aggressive gradient checkpointing just to barely make it work, and it’ll blow up as soon as batch size increases.
The 4070Ti Super 16G is tight for Flux training—you can run it with checkpointing, but it’s 50-80% slower than 24G cards. Full model fine-tuning on consumer cards? Forget it, that’s a whole different world.