Mattia Moretti
dc4993e9bb
Replace the unified inference node with dedicated nodes for Flux.1/2, Qwen, and Z models to provide model-specific parameters and better UX. Refactor the API client to simplify image model detection using strict prefixes (black-forest-labs/, Qwen/, Tongyi-MAI/) and remove unused caching logic. Update __init__.py to register all new nodes and print a concise load summary.
70 lines
1.8 KiB
Python
70 lines
1.8 KiB
Python
"""
|
|
utils.py — Shared utility functions for SiliconFlow nodes.
|
|
"""
|
|
|
|
import io
|
|
import base64
|
|
import numpy as np
|
|
|
|
try:
|
|
import torch
|
|
HAS_TORCH = True
|
|
except ImportError:
|
|
HAS_TORCH = False
|
|
|
|
try:
|
|
from PIL import Image
|
|
HAS_PIL = True
|
|
except ImportError:
|
|
HAS_PIL = False
|
|
|
|
|
|
def tensor_to_base64(tensor) -> str:
|
|
"""
|
|
Converts a ComfyUI tensor (B,H,W,C) float32 [0,1] to a base64 PNG string.
|
|
If there are multiple batches, uses the first frame.
|
|
"""
|
|
if tensor is None:
|
|
return None
|
|
|
|
if HAS_TORCH and isinstance(tensor, torch.Tensor):
|
|
img_array = tensor[0].cpu().numpy()
|
|
else:
|
|
img_array = np.array(tensor)[0]
|
|
|
|
img_array = (img_array * 255).clip(0, 255).astype(np.uint8)
|
|
|
|
if HAS_PIL:
|
|
pil_img = Image.fromarray(img_array, mode="RGB")
|
|
buf = io.BytesIO()
|
|
pil_img.save(buf, format="PNG")
|
|
return base64.b64encode(buf.getvalue()).decode("utf-8")
|
|
else:
|
|
raise RuntimeError("[SiliconFlow] PIL not available for image encoding.")
|
|
|
|
|
|
def bytes_to_tensor(image_bytes_list: list[bytes]):
|
|
"""
|
|
Converts a list of PNG/JPEG image bytes to a single ComfyUI tensor (B,H,W,C) float32 [0,1].
|
|
"""
|
|
tensors = []
|
|
for image_bytes in image_bytes_list:
|
|
if HAS_PIL:
|
|
pil_img = Image.open(io.BytesIO(image_bytes)).convert("RGB")
|
|
arr = np.array(pil_img, dtype=np.float32) / 255.0
|
|
else:
|
|
raise RuntimeError("[SiliconFlow] PIL not available for image decoding.")
|
|
|
|
if HAS_TORCH:
|
|
tensors.append(torch.from_numpy(arr))
|
|
else:
|
|
tensors.append(arr)
|
|
|
|
if not tensors:
|
|
return None
|
|
|
|
if HAS_TORCH:
|
|
return torch.stack(tensors) # (B, H, W, C)
|
|
else:
|
|
return np.stack(tensors) # (B, H, W, C)
|