Estimate too high

[jonesi13]

Yeah, it's not really accurate.

8GB estimate for single image.

[space-nuko]

Does it correct itself if you adjust the sliders first? It shows in the graph the reserved peak should be only about 1 GB with HR enabled.

[jonesi13]

It does scale up and down, but the numbers are always high.

[space-nuko]

Could you provide the stats.json that should be in the extension's folder? Those are the values the extension uses for calculations, could help with debugging

[jonesi13]

Sure.

stats.txt

[douglasrizzo]

I don't believe the estimates are necessarily always high, but there is definitely something weird going on.

In the screenshots below, I only change the batch size value. My VRAM estimate reaches a peak with batch size 8, then actually starts to go down if I increase batch size even further, and becomes negative after a batch size of 15.

stats.txt

[douglasrizzo]

I found something similar when changing image resolution (hires fix off, batch size = 1). I kept width fixed at 512 and changed only height,

VRAM estimate decreases when I change height from 64 to 240, by as much as 1 GB. Then the estimates go up until I reach a height of 1680, when it starts to go down again.

[angryptonX]

So either I don't understand the function or it's just confusing because the estimate applies more to LoRa generation for me. With LoRa is at 16 batches end, more than this and I get OOM but with Text2Image I can easily set a batch size of 60 and i am at 20GB VRAM. But the estimate says:

Estimated VRAM usage: 659175.81 MB / 24576 MB (2682.19%) (3960 MB system + 595650.74 MB used) <<< without Hires

Estimated VRAM usage: 620408.34 MB / 24576 MB (2524.45%) (3960 MB system + 560407.58 MB used) <<< with Hires

Estimated VRAM usage: 12488.11 MB / 24576 MB (50.81%) (3960 MB system + 7752.83 MB used) <<< Hires with Batchsize 1 (one)