For DreamBooth training in 2026, VRAM is the real bottleneck.

The Forge / 工坊
1

Just compiled my hands-on impressions running DreamBooth on a few cards I’ve got. Unlike inference, DreamBooth is one of the most VRAM-hungry tasks in consumer-grade AI, because it modifies the entire model weights, not just a tiny part like LoRA.

Bottom line: 4090’s 24GB is the sweet spot. For SD1.5, if that’s all you’re doing, 4070 Ti Super’s 16GB with some optimizations can barely get by. Two variables to watch for VRAM: enabling gradient checkpointing saves a ton, without it you’ll need 30-50% more headroom; FP16 is a must.

SDXL full fine-tuning is another tier above 1.5, and prior preservation pushes it even higher. Some cards claim they can run it, but that’s just barely scraping by with model offloading—training speed will be 3-5x slower.

Strictly speaking, that’s “runnable” but not “usable.” Anyone here managed to finish SDXL DreamBooth stably on a 16GB card? Mind sharing what optimizations you used?

2

Running SDXL full model on 16G is basically brute-forcing it with offload, so slow it makes you question your life choices.

3

Gradient checkpointing really saves VRAM, definitely a must-enable.

4

Yeah, 4090 at that price is definitely the sweet spot. Anything higher just isn’t worth it.

6

Honestly, a lot of people switch to LoRA just because DreamBooth’s VRAM requirements scare them off.

7

“Offloading lets it run” doesn’t mean it’s actually usable—that’s the real point right there.