Spaces:
Running
on
Zero
Running
on
Zero
Create inference_coz_single.py
Browse files- inference_coz_single.py +258 -0
inference_coz_single.py
ADDED
@@ -0,0 +1,258 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import tempfile
|
3 |
+
import uuid
|
4 |
+
import torch
|
5 |
+
from PIL import Image
|
6 |
+
from torchvision import transforms
|
7 |
+
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
8 |
+
from qwen_vl_utils import process_vision_info
|
9 |
+
from osediff_sd3 import OSEDiff_SD3_TEST, SD3Euler
|
10 |
+
|
11 |
+
# -------------------------------------------------------------------
|
12 |
+
# Helper: Resize & center-crop to a fixed square
|
13 |
+
# -------------------------------------------------------------------
|
14 |
+
def resize_and_center_crop(img: Image.Image, size: int) -> Image.Image:
|
15 |
+
w, h = img.size
|
16 |
+
scale = size / min(w, h)
|
17 |
+
new_w, new_h = int(w * scale), int(h * scale)
|
18 |
+
img = img.resize((new_w, new_h), Image.LANCZOS)
|
19 |
+
left = (new_w - size) // 2
|
20 |
+
top = (new_h - size) // 2
|
21 |
+
return img.crop((left, top, left + size, top + size))
|
22 |
+
|
23 |
+
|
24 |
+
# -------------------------------------------------------------------
|
25 |
+
# Helper: Generate a single VLM prompt for recursive_multiscale
|
26 |
+
# -------------------------------------------------------------------
|
27 |
+
def _generate_vlm_prompt(
|
28 |
+
vlm_model,
|
29 |
+
vlm_processor,
|
30 |
+
process_vision_info,
|
31 |
+
prev_image_path: str,
|
32 |
+
zoomed_image_path: str,
|
33 |
+
device: str = "cuda"
|
34 |
+
) -> str:
|
35 |
+
"""
|
36 |
+
Given two image file paths:
|
37 |
+
- prev_image_path: the “full” image at the previous recursion.
|
38 |
+
- zoomed_image_path: the cropped+resized (zoom) image for this step.
|
39 |
+
This builds a single “recursive_multiscale” prompt via Qwen2.5-VL.
|
40 |
+
Returns a string like “cat on sofa, pet, indoor, living room”, etc.
|
41 |
+
"""
|
42 |
+
# (1) Define the system message for recursive_multiscale:
|
43 |
+
message_text = (
|
44 |
+
"The second image is a zoom-in of the first image. "
|
45 |
+
"Based on this knowledge, what is in the second image? "
|
46 |
+
"Give me a set of words."
|
47 |
+
)
|
48 |
+
|
49 |
+
# (2) Build the two-image “chat” payload:
|
50 |
+
messages = [
|
51 |
+
{"role": "system", "content": message_text},
|
52 |
+
{
|
53 |
+
"role": "user",
|
54 |
+
"content": [
|
55 |
+
{"type": "image", "image": prev_image_path},
|
56 |
+
{"type": "image", "image": zoomed_image_path},
|
57 |
+
],
|
58 |
+
},
|
59 |
+
]
|
60 |
+
|
61 |
+
# (3) Wrap through the VL processor to get “inputs”:
|
62 |
+
text = vlm_processor.apply_chat_template(
|
63 |
+
messages, tokenize=False, add_generation_prompt=True
|
64 |
+
)
|
65 |
+
image_inputs, video_inputs = process_vision_info(messages)
|
66 |
+
inputs = vlm_processor(
|
67 |
+
text=[text],
|
68 |
+
images=image_inputs,
|
69 |
+
videos=video_inputs,
|
70 |
+
padding=True,
|
71 |
+
return_tensors="pt",
|
72 |
+
).to(device)
|
73 |
+
|
74 |
+
# (4) Generate tokens→decode
|
75 |
+
generated = vlm_model.generate(**inputs, max_new_tokens=128)
|
76 |
+
# strip off the prompt tokens from each generated sequence:
|
77 |
+
trimmed = [
|
78 |
+
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated)
|
79 |
+
]
|
80 |
+
out_text = vlm_processor.batch_decode(
|
81 |
+
trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
82 |
+
)[0]
|
83 |
+
|
84 |
+
# (5) Return exactly the bare words (no extra “,” if no additional user prompt)
|
85 |
+
return out_text.strip()
|
86 |
+
|
87 |
+
|
88 |
+
# -------------------------------------------------------------------
|
89 |
+
# Main Function: recursive_multiscale_sr
|
90 |
+
# -------------------------------------------------------------------
|
91 |
+
def recursive_multiscale_sr(
|
92 |
+
input_png_path: str,
|
93 |
+
upscale: int,
|
94 |
+
) -> list[Image.Image]:
|
95 |
+
"""
|
96 |
+
Perform exactly four recursive_multiscale super-resolution steps on a single PNG.
|
97 |
+
- input_png_path: path to a single .png file on disk.
|
98 |
+
- upscale: integer up-scale factor per recursion (e.g. 4).
|
99 |
+
Returns a list of 4 PIL.Image objects, corresponding to each SR output
|
100 |
+
at recursion steps 1, 2, 3, 4 (in that order).
|
101 |
+
|
102 |
+
All other parameters (model checkpoints, prompt model, process size, etc.)
|
103 |
+
are hard-coded exactly as in your command-line example.
|
104 |
+
"""
|
105 |
+
###############################
|
106 |
+
# 1. Fixed hyper-parameters
|
107 |
+
###############################
|
108 |
+
device = "cuda"
|
109 |
+
process_size = 512 # same as args.process_size
|
110 |
+
rec_num = 4 # fixed to 4 recursions
|
111 |
+
# model checkpoint paths (hard-coded to your example)
|
112 |
+
LORA_PATH = "ckpt/SR_LoRA/model_20001.pkl"
|
113 |
+
VAE_PATH = "ckpt/SR_VAE/vae_encoder_20001.pt"
|
114 |
+
SD3_MODEL = "stabilityai/stable-diffusion-3-medium-diffusers"
|
115 |
+
# VLM model name (hard-coded)
|
116 |
+
VLM_NAME = "Qwen/Qwen2.5-VL-3B-Instruct"
|
117 |
+
|
118 |
+
###############################
|
119 |
+
# 2. Build a dummy “args” namespace
|
120 |
+
# to satisfy OSEDiff_SD3_TEST constructor.
|
121 |
+
###############################
|
122 |
+
class _Args:
|
123 |
+
pass
|
124 |
+
|
125 |
+
args = _Args()
|
126 |
+
args.upscale = upscale
|
127 |
+
args.lora_path = LORA_PATH
|
128 |
+
args.vae_path = VAE_PATH
|
129 |
+
args.pretrained_model_name_or_path = SD3_MODEL
|
130 |
+
args.merge_and_unload_lora = False
|
131 |
+
args.lora_rank = 4
|
132 |
+
args.vae_decoder_tiled_size = 224
|
133 |
+
args.vae_encoder_tiled_size = 1024
|
134 |
+
args.latent_tiled_size = 96
|
135 |
+
args.latent_tiled_overlap = 32
|
136 |
+
args.mixed_precision = "fp16"
|
137 |
+
args.efficient_memory = False
|
138 |
+
# (other flags are not used by OSEDiff_SD3_TEST, so we skip them)
|
139 |
+
|
140 |
+
###############################
|
141 |
+
# 3. Load the SD3 SR model (non-efficient)
|
142 |
+
###############################
|
143 |
+
# 3.1 Instantiate the underlying SD3-Euler UNet/VAE/text encoders
|
144 |
+
sd3 = SD3Euler()
|
145 |
+
# move all text encoders+transformer+VAE to CUDA:
|
146 |
+
sd3.text_enc_1.to(device)
|
147 |
+
sd3.text_enc_2.to(device)
|
148 |
+
sd3.text_enc_3.to(device)
|
149 |
+
sd3.transformer.to(device, dtype=torch.float32)
|
150 |
+
sd3.vae.to(device, dtype=torch.float32)
|
151 |
+
# freeze
|
152 |
+
for p in (
|
153 |
+
sd3.text_enc_1,
|
154 |
+
sd3.text_enc_2,
|
155 |
+
sd3.text_enc_3,
|
156 |
+
sd3.transformer,
|
157 |
+
sd3.vae,
|
158 |
+
):
|
159 |
+
p.requires_grad_(False)
|
160 |
+
|
161 |
+
# 3.2 Wrap in OSEDiff_SD3_TEST helper:
|
162 |
+
model_test = OSEDiff_SD3_TEST(args, sd3)
|
163 |
+
# (by default, “model_test(...)” takes (lq_tensor, prompt=str) and returns a list[tensor])
|
164 |
+
|
165 |
+
###############################
|
166 |
+
# 4. Load the VLM (Qwen2.5-VL)
|
167 |
+
###############################
|
168 |
+
vlm_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
169 |
+
VLM_NAME,
|
170 |
+
torch_dtype="auto",
|
171 |
+
device_map="auto" # immediately dispatches layers onto available GPUs
|
172 |
+
)
|
173 |
+
vlm_processor = AutoProcessor.from_pretrained(VLM_NAME)
|
174 |
+
|
175 |
+
###############################
|
176 |
+
# 5. Pre-allocate a Temporary Directory
|
177 |
+
# to hold intermediate JPEG/PNG files
|
178 |
+
###############################
|
179 |
+
unique_id = uuid.uuid4().hex
|
180 |
+
prefix = f"recms_{unique_id}_"
|
181 |
+
|
182 |
+
with tempfile.TemporaryDirectory(prefix=prefix) as td:
|
183 |
+
# (we’ll write “prev.png” and “zoom.png” at each step)
|
184 |
+
|
185 |
+
###############################
|
186 |
+
# 6. Prepare the very first “full” image
|
187 |
+
###############################
|
188 |
+
# 6.1 Load + center crop → first_image is (512×512) PIL on CPU
|
189 |
+
img0 = Image.open(input_png_path).convert("RGB")
|
190 |
+
img0 = resize_and_center_crop(img0, process_size)
|
191 |
+
|
192 |
+
# 6.2 Save it once so VLM can read it as “prev.png”
|
193 |
+
prev_path = os.path.join(td, "step0_prev.png")
|
194 |
+
img0.save(prev_path)
|
195 |
+
|
196 |
+
# We will maintain a list of PIL outputs here:
|
197 |
+
sr_pil_list: list[Image.Image] = []
|
198 |
+
prompt_list = []
|
199 |
+
|
200 |
+
###############################
|
201 |
+
# 7. Recursion loop (exactly 4 times)
|
202 |
+
###############################
|
203 |
+
for rec in range(rec_num):
|
204 |
+
# (A) Crop + upsample the “prev” image to obtain this step’s input → zoomed
|
205 |
+
prev_pil = Image.open(prev_path).convert("RGB")
|
206 |
+
w, h = prev_pil.size # should be (512×512) each time
|
207 |
+
new_w, new_h = w // upscale, h // upscale # e.g. 128×128 for upscale=4
|
208 |
+
# center-crop region:
|
209 |
+
left = (w - new_w) // 2
|
210 |
+
top = (h - new_h) // 2
|
211 |
+
right = left + new_w
|
212 |
+
bottom = top + new_h
|
213 |
+
cropped = prev_pil.crop((left, top, right, bottom))
|
214 |
+
|
215 |
+
# (B) Resize that crop back up to (512×512) via BICUBIC → zoomed
|
216 |
+
zoomed = cropped.resize((w, h), Image.BICUBIC)
|
217 |
+
zoom_path = os.path.join(td, f"step{rec+1}_zoom.png")
|
218 |
+
zoomed.save(zoom_path)
|
219 |
+
|
220 |
+
# (C) Generate a recursive_multiscale VLM “tag” prompt
|
221 |
+
prompt_tag = _generate_vlm_prompt(
|
222 |
+
vlm_model=vlm_model,
|
223 |
+
vlm_processor=vlm_processor,
|
224 |
+
process_vision_info=process_vision_info,
|
225 |
+
prev_image_path=prev_path,
|
226 |
+
zoomed_image_path=zoom_path,
|
227 |
+
device=device,
|
228 |
+
)
|
229 |
+
# (By default, no extra user prompt is appended.)
|
230 |
+
|
231 |
+
# (D) Prepare the low-res tensor for SR: convert zoomed→Tensor→[0,1]→[−1,1]
|
232 |
+
to_tensor = transforms.ToTensor()
|
233 |
+
lq = to_tensor(zoomed).unsqueeze(0).to(device) # shape (1,3,512,512)
|
234 |
+
lq = (lq * 2.0) - 1.0
|
235 |
+
|
236 |
+
# (E) Do SR inference:
|
237 |
+
with torch.no_grad():
|
238 |
+
out_tensor = model_test(lq, prompt=prompt_tag)[0] # (3,512,512) on CPU or GPU
|
239 |
+
out_tensor = out_tensor.clamp(-1.0, 1.0).cpu()
|
240 |
+
# back to PIL in [0,1]:
|
241 |
+
out_pil = transforms.ToPILImage()((out_tensor * 0.5) + 0.5)
|
242 |
+
|
243 |
+
# (F) Save this step’s SR output as “prev.png” for next iteration:
|
244 |
+
out_path = os.path.join(td, f"step{rec+1}_sr.png")
|
245 |
+
out_pil.save(out_path)
|
246 |
+
prev_path = out_path
|
247 |
+
|
248 |
+
# (G) Append the PIL to our list:
|
249 |
+
sr_pil_list.append(out_pil)
|
250 |
+
prompt_list.append(prompt_tag)
|
251 |
+
|
252 |
+
# end for(rec)
|
253 |
+
|
254 |
+
###############################
|
255 |
+
# 8. Return the four SR‐PILs
|
256 |
+
###############################
|
257 |
+
# The list sr_pil_list = [ SR1, SR2, SR3, SR4 ] in order.
|
258 |
+
return sr_pil_list, prompt_list
|