SD1.5 / SDXL / Flux, three generations in one house, how much difference is there in the cost per image?

The Forge / 工坊 Models / LoRA / 模型与 LoRA
1

Now in ComfyUI, three generations of models often coexist, and picking one is really about the economics, not some mystical stuff. SD1.5 has the deepest fine-tuning ecosystem of the bunch, the community’s built up stuff you can’t beat; SDXL is the default backbone for most home users; Flux uses the DiT architecture with 12B parameters, and the image quality is way closer to real photography—hands, text in images, complex lighting, all the stuff SD and SDXL struggle with, it handles way better.

The trade-off is hardware: Flux needs 12 to 24G VRAM, and on the same card, each image is four to ten times slower than SDXL.

Whether it’s worth it depends on how many images you’re generating at once. At US electricity rates, the per-image power cost isn’t that crazy, but if you’re doing batch iterations, being four times slower is a real time cost. Sometimes a $30 monthly cloud subscription ends up being the better deal.

2

Yeah, DiT is a real game-changer.

3

Hands and text—those two eternal pains—Flux is actually really good at them.

4

The real killer is that batch scenes are 4x slower.

5

Honestly, the electricity bill isn’t the real issue—it’s the time that kills you.

6

Honestly, the deep fine-tuning ecosystem for 1.5 is still unmatched. Nobody’s been able to replace it yet.

7

Front row

8

Batch users are really feeling the pain with Flux—it’s expensive to run in bulk. For mass output, SDXL with an acceleration LoRA is way more cost-effective; the quality is good enough as long as it gets the job done.