Diffusion and image generation

Why the 4090 wins here

Diffusion models are small enough to fit comfortably in 24 GB, so the extra VRAM on datacentre cards buys you nothing. What matters is raw shader and tensor throughput, and the RTX 4090 delivers 330 TFLOPS of FP16 for $0.44/hr - roughly a fifth the price of an H100 for half the throughput.

Batch and resolution

SDXL at 1024x1024 uses about 12 GB at batch size 4, leaving headroom for ControlNet and LoRA stacks. Flux is heavier: budget 20 GB at batch size 1 in fp16, or use the fp8 variants to halve that.

Throughput reference

• SDXL 1024px, 30 steps: about 1.4 s per image on a 4090
• SD 1.5 512px, 20 steps: about 0.4 s per image on a 4090
• Flux.1-dev 1024px, 28 steps: about 6 s per image on a 4090
• Training a style LoRA on 40 images: 12-20 minutes