Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,854 Bytes
703e263 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 |
from ..models import ModelManager, FluxDiT, FluxTextEncoder1, FluxTextEncoder2, FluxVAEDecoder, FluxVAEEncoder
from ..prompters import FluxPrompter
from ..schedulers import FlowMatchScheduler
from .base import BasePipeline
import torch
from tqdm import tqdm
class FluxImagePipeline(BasePipeline):
def __init__(self, device="cuda", torch_dtype=torch.float16):
super().__init__(device=device, torch_dtype=torch_dtype)
self.scheduler = FlowMatchScheduler()
self.prompter = FluxPrompter()
# models
self.text_encoder_1: FluxTextEncoder1 = None
self.text_encoder_2: FluxTextEncoder2 = None
self.dit: FluxDiT = None
self.vae_decoder: FluxVAEDecoder = None
self.vae_encoder: FluxVAEEncoder = None
def denoising_model(self):
return self.dit
def fetch_models(self, model_manager: ModelManager, prompt_refiner_classes=[]):
self.text_encoder_1 = model_manager.fetch_model("flux_text_encoder_1")
self.text_encoder_2 = model_manager.fetch_model("flux_text_encoder_2")
self.dit = model_manager.fetch_model("flux_dit")
self.vae_decoder = model_manager.fetch_model("flux_vae_decoder")
self.vae_encoder = model_manager.fetch_model("flux_vae_encoder")
self.prompter.fetch_models(self.text_encoder_1, self.text_encoder_2)
self.prompter.load_prompt_refiners(model_manager, prompt_refiner_classes)
@staticmethod
def from_model_manager(model_manager: ModelManager, prompt_refiner_classes=[]):
pipe = FluxImagePipeline(
device=model_manager.device,
torch_dtype=model_manager.torch_dtype,
)
pipe.fetch_models(model_manager, prompt_refiner_classes)
return pipe
def encode_image(self, image, tiled=False, tile_size=64, tile_stride=32):
latents = self.vae_encoder(image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
return latents
def decode_image(self, latent, tiled=False, tile_size=64, tile_stride=32):
image = self.vae_decoder(latent.to(self.device), tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
image = self.vae_output_to_image(image)
return image
def encode_prompt(self, prompt, positive=True):
prompt_emb, pooled_prompt_emb, text_ids = self.prompter.encode_prompt(
prompt, device=self.device, positive=positive
)
return {"prompt_emb": prompt_emb, "pooled_prompt_emb": pooled_prompt_emb, "text_ids": text_ids}
def prepare_extra_input(self, latents=None, guidance=0.0):
latent_image_ids = self.dit.prepare_image_ids(latents)
guidance = torch.Tensor([guidance] * latents.shape[0]).to(device=latents.device, dtype=latents.dtype)
return {"image_ids": latent_image_ids, "guidance": guidance}
@torch.no_grad()
def __call__(
self,
prompt,
local_prompts=[],
masks=[],
mask_scales=[],
negative_prompt="",
cfg_scale=1.0,
embedded_guidance=0.0,
input_image=None,
denoising_strength=1.0,
height=1024,
width=1024,
num_inference_steps=30,
tiled=False,
tile_size=128,
tile_stride=64,
progress_bar_cmd=tqdm,
progress_bar_st=None,
):
# Tiler parameters
tiler_kwargs = {"tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride}
# Prepare scheduler
self.scheduler.set_timesteps(num_inference_steps, denoising_strength)
# Prepare latent tensors
if input_image is not None:
image = self.preprocess_image(input_image).to(device=self.device, dtype=self.torch_dtype)
latents = self.encode_image(image, **tiler_kwargs)
noise = torch.randn((1, 16, height//8, width//8), device=self.device, dtype=self.torch_dtype)
latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0])
else:
latents = torch.randn((1, 16, height//8, width//8), device=self.device, dtype=self.torch_dtype)
# Encode prompts
prompt_emb_posi = self.encode_prompt(prompt, positive=True)
if cfg_scale != 1.0:
prompt_emb_nega = self.encode_prompt(negative_prompt, positive=False)
prompt_emb_locals = [self.encode_prompt(prompt_local) for prompt_local in local_prompts]
# Extra input
extra_input = self.prepare_extra_input(latents, guidance=embedded_guidance)
# Denoise
for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
timestep = timestep.unsqueeze(0).to(self.device)
# Classifier-free guidance
inference_callback = lambda prompt_emb_posi: self.dit(
latents, timestep=timestep, **prompt_emb_posi, **tiler_kwargs, **extra_input
)
noise_pred_posi = self.control_noise_via_local_prompts(prompt_emb_posi, prompt_emb_locals, masks, mask_scales, inference_callback)
if cfg_scale != 1.0:
noise_pred_nega = self.dit(
latents, timestep=timestep, **prompt_emb_nega, **tiler_kwargs, **extra_input
)
noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
else:
noise_pred = noise_pred_posi
# Iterate
latents = self.scheduler.step(noise_pred, self.scheduler.timesteps[progress_id], latents)
# UI
if progress_bar_st is not None:
progress_bar_st.progress(progress_id / len(self.scheduler.timesteps))
# Decode image
image = self.decode_image(latents, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
return image
|