Fix different first gen with Approx NN previews

The loading of the model for approx nn live previews can change the internal state of PyTorch, resulting in a different image. This can be avoided by preloading the approx nn model in advance.
This commit is contained in:
brkirch 2023-01-23 22:49:20 -05:00
parent 5c1cb9263f
commit f64af77adc

View file

@ -13,7 +13,7 @@ from skimage import exposure
from typing import Any, Dict, List, Optional
import modules.sd_hijack
from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste, script_callbacks, extra_networks
from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste, script_callbacks, extra_networks, sd_vae_approx
from modules.sd_hijack import model_hijack
from modules.shared import opts, cmd_opts, state
import modules.shared as shared
@ -568,6 +568,10 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
with devices.autocast():
p.init(p.all_prompts, p.all_seeds, p.all_subseeds)
if shared.opts.live_previews_enable and sd_samplers.approximation_indexes.get(shared.opts.show_progress_type, 0) == 1:
# preload approx nn model before sampling for a more deterministic result
sd_vae_approx.model()
if not p.disable_extra_networks:
extra_networks.activate(p, extra_network_data)