test image, from sd

no old size, yes inference steps, warm up on two sizes

segformer fp16 timing is 80-100ms?

seg preferred device vs inpaint, and compile everything

segformer fp16

parameterize inpaint size

parameterize inpaint size

parameterize inpaint size

no compiling for now

segformer fp32 for now

onnxrt

what even are the backends

torch_tensorrt

use lcm for inpainting instead of inpainting model with lcm lora and deep cache etc, parameterize num inference steps and random seed

DeepCache doesnt compile lol

DeepCache?

torch compile

16 steps no lora

384, 8 step lora

512

512

12 steps no lora

8 steps

fp32 for segformer

fp32 for now

fastai unet to segformer

mps

initializer?

more steps

more steps

more steps

more steps

mask_image

print

print

print

new reqs

maybe no pil image conversions?

kwargslol

from_pretrained

inpainting

types

types

types? f8

types

Ensure passing in a PIL image

unused prompt to inpaint

256 to 512 resolution redaction

not doing it live

not doing it live