Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -3,26 +3,29 @@ from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
|
3 |
from diffusers import StableDiffusionPipeline
|
4 |
from sentence_transformers import SentenceTransformer, util
|
5 |
import torch
|
|
|
6 |
|
7 |
# --- Load models ---
|
8 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
9 |
|
10 |
# Text-to-text model
|
11 |
-
text_model_name = "google/flan-t5-
|
12 |
text_tokenizer = AutoTokenizer.from_pretrained(text_model_name)
|
13 |
-
text_model = AutoModelForSeq2SeqLM.from_pretrained(text_model_name)
|
14 |
|
15 |
# Text-to-image model
|
16 |
image_model_id = "runwayml/stable-diffusion-v1-5"
|
17 |
image_pipe = StableDiffusionPipeline.from_pretrained(
|
18 |
-
image_model_id,
|
|
|
|
|
19 |
)
|
20 |
image_pipe = image_pipe.to(device)
|
21 |
|
22 |
# Sentence similarity model
|
23 |
embedder = SentenceTransformer('all-MiniLM-L6-v2')
|
24 |
|
25 |
-
#
|
26 |
image_triggers = [
|
27 |
"generate an image of",
|
28 |
"draw a",
|
@@ -33,30 +36,30 @@ image_triggers = [
|
|
33 |
"sketch",
|
34 |
]
|
35 |
|
36 |
-
# ---
|
37 |
def multimodal_agent(prompt):
|
38 |
-
# Step 1:
|
39 |
prompt_embedding = embedder.encode(prompt, convert_to_tensor=True)
|
40 |
trigger_embeddings = embedder.encode(image_triggers, convert_to_tensor=True)
|
41 |
-
|
42 |
cosine_scores = util.pytorch_cos_sim(prompt_embedding, trigger_embeddings)
|
43 |
max_score = torch.max(cosine_scores).item()
|
44 |
|
45 |
-
# Step 2:
|
46 |
if max_score > 0.65:
|
47 |
# Generate image
|
48 |
-
|
49 |
-
|
|
|
50 |
else:
|
51 |
# Generate text
|
52 |
-
inputs = text_tokenizer(prompt, return_tensors="pt")
|
53 |
outputs = text_model.generate(**inputs, max_new_tokens=100)
|
54 |
text = text_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
55 |
-
return text, None
|
56 |
|
57 |
-
# ---
|
58 |
with gr.Blocks() as demo:
|
59 |
-
gr.Markdown("# 🤖 Smart Multimodal AI Agent\
|
60 |
|
61 |
input_box = gr.Textbox(label="Enter your prompt")
|
62 |
output_text = gr.Textbox(label="Text Output")
|
|
|
3 |
from diffusers import StableDiffusionPipeline
|
4 |
from sentence_transformers import SentenceTransformer, util
|
5 |
import torch
|
6 |
+
import contextlib
|
7 |
|
8 |
# --- Load models ---
|
9 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
10 |
|
11 |
# Text-to-text model
|
12 |
+
text_model_name = "google/flan-t5-large"
|
13 |
text_tokenizer = AutoTokenizer.from_pretrained(text_model_name)
|
14 |
+
text_model = AutoModelForSeq2SeqLM.from_pretrained(text_model_name).to(device)
|
15 |
|
16 |
# Text-to-image model
|
17 |
image_model_id = "runwayml/stable-diffusion-v1-5"
|
18 |
image_pipe = StableDiffusionPipeline.from_pretrained(
|
19 |
+
image_model_id,
|
20 |
+
torch_dtype=torch.float16 if device == "cuda" else torch.float32,
|
21 |
+
safety_checker=None # Optional for debugging
|
22 |
)
|
23 |
image_pipe = image_pipe.to(device)
|
24 |
|
25 |
# Sentence similarity model
|
26 |
embedder = SentenceTransformer('all-MiniLM-L6-v2')
|
27 |
|
28 |
+
# Image-like trigger phrases
|
29 |
image_triggers = [
|
30 |
"generate an image of",
|
31 |
"draw a",
|
|
|
36 |
"sketch",
|
37 |
]
|
38 |
|
39 |
+
# --- Core logic ---
|
40 |
def multimodal_agent(prompt):
|
41 |
+
# Step 1: Semantic similarity to image triggers
|
42 |
prompt_embedding = embedder.encode(prompt, convert_to_tensor=True)
|
43 |
trigger_embeddings = embedder.encode(image_triggers, convert_to_tensor=True)
|
|
|
44 |
cosine_scores = util.pytorch_cos_sim(prompt_embedding, trigger_embeddings)
|
45 |
max_score = torch.max(cosine_scores).item()
|
46 |
|
47 |
+
# Step 2: Decision branch
|
48 |
if max_score > 0.65:
|
49 |
# Generate image
|
50 |
+
with torch.autocast("cuda") if device == "cuda" else contextlib.nullcontext():
|
51 |
+
image = image_pipe(prompt).images[0]
|
52 |
+
return None, image
|
53 |
else:
|
54 |
# Generate text
|
55 |
+
inputs = text_tokenizer(prompt, return_tensors="pt").to(device)
|
56 |
outputs = text_model.generate(**inputs, max_new_tokens=100)
|
57 |
text = text_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
58 |
+
return text, None
|
59 |
|
60 |
+
# --- UI ---
|
61 |
with gr.Blocks() as demo:
|
62 |
+
gr.Markdown("# 🤖 Smart Multimodal AI Agent\nGive a prompt — It decides text vs image automatically!")
|
63 |
|
64 |
input_box = gr.Textbox(label="Enter your prompt")
|
65 |
output_text = gr.Textbox(label="Text Output")
|