edit

app.py

@@ -4,18 +4,17 @@ import spaces
 import os
 import numpy as np
 from PIL import Image
-
 from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
 from omegaconf import OmegaConf
+
+from image_datasets.dataset import image_resize
 from src.flux.util import load_ae, load_clip, load_flow_model2, load_t5, tensor_to_pil_image
 from src.flux.xflux_pipeline import XFluxSampler
-from image_datasets.dataset import image_resize
-
-# ===== No CUDA/model initialization globally =====
 args = OmegaConf.load("inference_configs/inference.yaml")
 is_schnell = args.model_name == "flux-schnell"
-
+'/home/user/app/assets/0_camera_zoom/20486354.png'
+'/home/user/app/assets/0_camera_zoom/20486354.png'
 # sampler = None
 device = torch.device("cuda")
 dtype = torch.bfloat16
@@ -23,36 +22,39 @@ dit = load_flow_model2(args.model_name, device="cpu").to(device, dtype=dtype)
 vae = load_ae(args.model_name, device="cpu").to(device, dtype=dtype)
 t5 = load_t5(device="cpu", max_length=256 if is_schnell else 512).to(device, dtype=dtype)
 clip = load_clip("cpu").to(device, dtype=dtype)
-
-vae.requires_grad_(False)
-t5.requires_grad_(False)
-clip.requires_grad_(False)
-
-model_path = hf_hub_download(
-    repo_id="Boese0601/ByteMorpher",
-    filename="dit.safetensors",
-    use_auth_token=os.getenv("HF_TOKEN")
-)
-state_dict = load_file(model_path)
-dit.load_state_dict(state_dict)
-dit.eval()
-dit.to(device, dtype=dtype)
-
-sampler = XFluxSampler(
-    clip=clip,
-    t5=t5,
-    ae=vae,
-    model=dit,
-    device=device,
-    ip_loaded=False,
-    spatial_condition=False,
-    clip_image_processor=None,
-    image_encoder=None,
-    improj=None
-)
 #test push
 @spaces.GPU
 def generate(image: Image.Image, edit_prompt: str):
+
+    
+    
+
+    vae.requires_grad_(False)
+    t5.requires_grad_(False)
+    clip.requires_grad_(False)
+
+    model_path = hf_hub_download(
+        repo_id="Boese0601/ByteMorpher",
+        filename="dit.safetensors",
+        use_auth_token=os.getenv("HF_TOKEN")
+    )
+    state_dict = load_file(model_path)
+    dit.load_state_dict(state_dict)
+    dit.eval()
+    dit.to(device, dtype=dtype)
+
+    sampler = XFluxSampler(
+        clip=clip,
+        t5=t5,
+        ae=vae,
+        model=dit,
+        device=device,
+        ip_loaded=False,
+        spatial_condition=False,
+        clip_image_processor=None,
+        image_encoder=None,
+        improj=None
+    )
     # global sampler
     # device = torch.device("cuda")
     # dtype = torch.bfloat16
@@ -95,18 +97,17 @@ def generate(image: Image.Image, edit_prompt: str):
     img = torch.from_numpy((np.array(img) / 127.5) - 1)
     img = img.permute(2, 0, 1).unsqueeze(0).to(device, dtype=dtype)
 
-    with torch.no_grad():
-        result = sampler(
-            prompt=edit_prompt,
-            width=args.sample_width,
-            height=args.sample_height,
-            num_steps=args.sample_steps,
-            image_prompt=None,
-            true_gs=args.cfg_scale,
-            seed=args.seed,
-            ip_scale=args.ip_scale if args.use_ip else 1.0,
-            source_image=img if args.use_spatial_condition else None,
-        )
+    result = sampler(
+        prompt=edit_prompt,
+        width=args.sample_width,
+        height=args.sample_height,
+        num_steps=args.sample_steps,
+        image_prompt=None,
+        true_gs=args.cfg_scale,
+        seed=args.seed,
+        ip_scale=args.ip_scale if args.use_ip else 1.0,
+        source_image=img if args.use_spatial_condition else None,
+    )
     return tensor_to_pil_image(result)
 
 def get_samples():

assets/0_camera_zoom/20486354.json

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7917201faf043e935ea4ddd94c7e570fe5ca51f8bed66ee4d4dabe480f8390b5
-size 1128