Update HumanSegmentation (#826)
* Change to new mask components on humanSegmentation * Change segformer index * Add segformer_b3_clothes and fashion * Add face_parsing * Fix face_parsing output error images --------- Co-authored-by: yolain <me@yolain.com>
This commit is contained in:
yolain
Submodule ComfyUI-Easy-Use-Frontend updated: 02a985b80c...bafdb96c0b
@@ -6620,7 +6620,7 @@
|
||||
"name": "温度"
|
||||
},
|
||||
"max_tokens": {
|
||||
"name": "最大词令牌数"
|
||||
"name": "最大词元数"
|
||||
},
|
||||
"caption_type": {
|
||||
"name": "提示词类型"
|
||||
@@ -6651,7 +6651,7 @@
|
||||
"name": "温度"
|
||||
},
|
||||
"max_tokens": {
|
||||
"name": "最大词令牌数"
|
||||
"name": "最大词元数"
|
||||
},
|
||||
"caption_type": {
|
||||
"name": "提示词类型"
|
||||
|
||||
@@ -370,6 +370,15 @@ HUMANPARSING_MODELS = {
|
||||
},
|
||||
"human-parts":{
|
||||
"model_url":"https://huggingface.co/Metal3d/deeplabv3p-resnet50-human/resolve/main/deeplabv3p-resnet50-human.onnx",
|
||||
},
|
||||
"segformer_b3_clothes":{
|
||||
"model_name": "sayeed99/segformer-b3-clothes",
|
||||
},
|
||||
"segformer_b3_fashion":{
|
||||
"model_name": "sayeed99/segformer-b3-fashion",
|
||||
},
|
||||
"face_parsing":{
|
||||
"model_name": "jonathandinu/face-parsing"
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
+115
-50
@@ -4,12 +4,13 @@ import torch
|
||||
import numpy as np
|
||||
import comfy.utils
|
||||
import comfy.model_management
|
||||
import shutil
|
||||
from comfy_extras.nodes_compositing import JoinImageWithAlpha
|
||||
from server import PromptServer
|
||||
from nodes import MAX_RESOLUTION, NODE_CLASS_MAPPINGS as ALL_NODE_CLASS_MAPPINGS
|
||||
from PIL import Image, ImageDraw, ImageFilter, ImageOps
|
||||
import torch.nn.functional as F
|
||||
from torchvision.transforms import Resize, CenterCrop, GaussianBlur
|
||||
from torchvision.transforms import Resize, CenterCrop, GaussianBlur, ToPILImage
|
||||
from torchvision.transforms.functional import to_pil_image
|
||||
from ..libs.log import log_node_info
|
||||
from ..libs.utils import AlwaysEqualProxy, ByPassTypeTuple
|
||||
@@ -1277,13 +1278,22 @@ class humanSegmentation:
|
||||
|
||||
@classmethod
|
||||
def INPUT_TYPES(cls):
|
||||
|
||||
return {
|
||||
"required":{
|
||||
"image": ("IMAGE",),
|
||||
"method": (["selfie_multiclass_256x256", "human_parsing_lip", "human_parts (deeplabv3p)"],),
|
||||
"method": (["selfie_multiclass_256x256", "human_parsing_lip", "human_parts (deeplabv3p)", "segformer_b3_clothes", "segformer_b3_fashion", "face_parsing"],),
|
||||
"confidence": ("FLOAT", {"default": 0.4, "min": 0.05, "max": 0.95, "step": 0.01},),
|
||||
"crop_multi": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 10.0, "step": 0.001},),
|
||||
"mask_components":(
|
||||
"EASY_COMBO",{
|
||||
"options": [{'label':'Background','value':0}],
|
||||
"multi_select": {
|
||||
"placeholder": "select mask components",
|
||||
"chip": True,
|
||||
"max_selected_labels": 4,
|
||||
}
|
||||
}
|
||||
)
|
||||
},
|
||||
"hidden": {
|
||||
"prompt": "PROMPT",
|
||||
@@ -1309,12 +1319,7 @@ class humanSegmentation:
|
||||
numpy_image = cv2.cvtColor(numpy_image, cv2.COLOR_BGR2RGB)
|
||||
return mp.Image(image_format=image_format, data=numpy_image)
|
||||
|
||||
def parsing(self, image, confidence, method, crop_multi, prompt=None, my_unique_id=None):
|
||||
mask_components = []
|
||||
if my_unique_id in prompt:
|
||||
if prompt[my_unique_id]["inputs"]['mask_components']:
|
||||
mask_components = prompt[my_unique_id]["inputs"]['mask_components'].split(',')
|
||||
mask_components = list(map(int, mask_components))
|
||||
def parsing(self, image, confidence, method, crop_multi, mask_components, prompt=None, my_unique_id=None):
|
||||
if method == 'selfie_multiclass_256x256':
|
||||
try:
|
||||
import mediapipe as mp
|
||||
@@ -1442,6 +1447,107 @@ class humanSegmentation:
|
||||
output_image = torch.cat(ret_images, dim=0)
|
||||
mask = torch.cat(ret_masks, dim=0)
|
||||
|
||||
elif method in ["segformer_b3_clothes", "segformer_b3_fashion", "face_parsing"]:
|
||||
from transformers import SegformerImageProcessor, AutoModelForSemanticSegmentation
|
||||
|
||||
# 分割
|
||||
def get_segmentation_from_model(tensor_image, model, processor):
|
||||
cloth = tensor2pil(tensor_image)
|
||||
inputs = processor(images=cloth, return_tensors="pt")
|
||||
outputs = model(**inputs)
|
||||
logits = outputs.logits.cpu()
|
||||
upsampled_logits = F.interpolate(logits, size=cloth.size[::-1], mode="bilinear",
|
||||
align_corners=False)
|
||||
pred_seg = upsampled_logits.argmax(dim=1)[0].numpy()
|
||||
return pred_seg, cloth
|
||||
|
||||
|
||||
if method in cache:
|
||||
_, (processor, model) = cache[method][1]
|
||||
else:
|
||||
model_folder_path = os.path.join(folder_paths.models_dir, method)
|
||||
if os.path.exists(model_folder_path):
|
||||
print(f"Start to load existing model...")
|
||||
else:
|
||||
from huggingface_hub import snapshot_download
|
||||
PromptServer.instance.send_sync("easyuse-toast", {"content": f"Model not found locally. Downloading {method}...", "type": 'loading', "duration": 10000})
|
||||
print(f"Model not found locally. Downloading {method}...")
|
||||
model_path_cache = os.path.join(folder_paths.models_dir, "cache-"+method)
|
||||
snapshot_download(
|
||||
repo_id=HUMANPARSING_MODELS[method]['model_name'],
|
||||
local_dir=model_path_cache,
|
||||
local_dir_use_symlinks=False,
|
||||
resume_download=True
|
||||
)
|
||||
shutil.move(model_path_cache, model_folder_path)
|
||||
print(f"Model downloaded to {model_folder_path}...")
|
||||
try:
|
||||
model_folder_path = os.path.normpath(folder_paths.folder_names_and_paths[method][0][0])
|
||||
except:
|
||||
pass
|
||||
|
||||
processor = SegformerImageProcessor.from_pretrained(model_folder_path)
|
||||
model = AutoModelForSemanticSegmentation.from_pretrained(model_folder_path)
|
||||
update_cache(method, 'human_segmentation', (False, (processor, model)))
|
||||
|
||||
ret_images = []
|
||||
ret_masks = []
|
||||
|
||||
if method == "face_parsing":
|
||||
import matplotlib
|
||||
import torchvision.transforms as T
|
||||
transform = ToPILImage()
|
||||
colormap = matplotlib.colormaps['viridis']
|
||||
device = model.device
|
||||
results = []
|
||||
images = []
|
||||
for img in image:
|
||||
size = img.shape[:2]
|
||||
inputs = processor(images=transform(img.permute(2, 0, 1)), return_tensors="pt")
|
||||
inputs = {k: v.to(device) for k, v in inputs.items()}
|
||||
outputs = model(**inputs)
|
||||
logits = outputs.logits
|
||||
upsampled_logits = F.interpolate(
|
||||
logits,
|
||||
size=size,
|
||||
mode="bilinear",
|
||||
align_corners=False)
|
||||
|
||||
pred_seg = upsampled_logits.argmax(dim=1)[0]
|
||||
pred_seg_np = pred_seg.cpu().detach().numpy().astype(np.uint8)
|
||||
results.append(torch.tensor(pred_seg_np))
|
||||
|
||||
results_out = torch.stack(results, dim=0)
|
||||
for img, result_item in zip(image, results_out):
|
||||
mask = torch.zeros(result_item.shape, dtype=torch.uint8)
|
||||
for i in mask_components:
|
||||
mask = mask | torch.where(result_item == i, 1, 0)
|
||||
|
||||
# 将mask转换为numpy数组，并确保数据类型正确
|
||||
mask_np = (mask * 255).numpy().astype(np.uint8)
|
||||
_mask = Image.fromarray(mask_np)
|
||||
|
||||
# 处理图像输出
|
||||
ret_image = RGB2RGBA(tensor2pil(img).convert('RGB'), _mask.convert('L'))
|
||||
ret_images.append(pil2tensor(ret_image))
|
||||
ret_masks.append(image2mask(_mask))
|
||||
|
||||
else:
|
||||
for img in image:
|
||||
pred_seg, cloth = get_segmentation_from_model(img, model, processor)
|
||||
i = torch.unsqueeze(img, 0)
|
||||
i = pil2tensor(tensor2pil(i).convert('RGB'))
|
||||
|
||||
mask = np.isin(pred_seg, mask_components).astype(np.uint8)
|
||||
_mask = Image.fromarray(mask * 255)
|
||||
|
||||
ret_image = RGB2RGBA(tensor2pil(img).convert('RGB'), _mask.convert('L'))
|
||||
ret_images.append(pil2tensor(ret_image))
|
||||
ret_masks.append(image2mask(_mask))
|
||||
|
||||
output_image = torch.cat(ret_images, dim=0)
|
||||
mask = torch.cat(ret_masks, dim=0)
|
||||
|
||||
# use crop
|
||||
bbox = [[0, 0, 0, 0]]
|
||||
if crop_multi > 0.0:
|
||||
@@ -1892,47 +1998,6 @@ class loadImagesForLoop:
|
||||
"result": tuple(["stub", index, image, mask, name] + outputs),
|
||||
"expand": graph.finalize(),
|
||||
}
|
||||
# 姿势编辑器
|
||||
# class poseEditor:
|
||||
# @classmethod
|
||||
# def INPUT_TYPES(self):
|
||||
# temp_dir = folder_paths.get_temp_directory()
|
||||
#
|
||||
# if not os.path.isdir(temp_dir):
|
||||
# os.makedirs(temp_dir)
|
||||
#
|
||||
# temp_dir = folder_paths.get_temp_directory()
|
||||
#
|
||||
# return {"required":
|
||||
# {"image": (sorted(os.listdir(temp_dir)),)},
|
||||
# }
|
||||
#
|
||||
# RETURN_TYPES = ("IMAGE",)
|
||||
# FUNCTION = "output_pose"
|
||||
#
|
||||
# CATEGORY = "EasyUse/🚫 Deprecated"
|
||||
#
|
||||
# def output_pose(self, image):
|
||||
# image_path = os.path.join(folder_paths.get_temp_directory(), image)
|
||||
# # print(f"Create: {image_path}")
|
||||
#
|
||||
# i = Image.open(image_path)
|
||||
# image = i.convert("RGB")
|
||||
# image = np.array(image).astype(np.float32) / 255.0
|
||||
# image = torch.from_numpy(image)[None,]
|
||||
#
|
||||
# return (image,)
|
||||
#
|
||||
# @classmethod
|
||||
# def IS_CHANGED(self, image):
|
||||
# image_path = os.path.join(
|
||||
# folder_paths.get_temp_directory(), image)
|
||||
# # print(f'Change: {image_path}')
|
||||
#
|
||||
# m = hashlib.sha256()
|
||||
# with open(image_path, 'rb') as f:
|
||||
# m.update(f.read())
|
||||
# return m.digest().hex()
|
||||
|
||||
class makeImageForICRepaint:
|
||||
@classmethod
|
||||
|
||||
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
+1
-1
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
@@ -0,0 +1 @@
|
||||
import{x as e,y as t,j as n,o,n as i,i as l,C as r,R as u,e as a,S as s}from"./vue-CYk3Qk0T.js";function c(n){return!!e()&&(t(n),!0)}const v="undefined"!=typeof window&&"undefined"!=typeof document;"undefined"!=typeof WorkerGlobalScope&&(globalThis,WorkerGlobalScope);const f=e=>null!=e,d=Object.prototype.toString,p=e=>"[object Object]"===d.call(e);function m(e){return Array.isArray(e)?e:[e]}function h(e,t=!0,n){l()?o(e,n):t?e():i(e)}const w=v?window:void 0;function y(e){var t;const n=u(e);return null!=(t=null==n?void 0:n.$el)?t:n}function b(...e){const t=[],o=()=>{t.forEach((e=>e())),t.length=0},i=r((()=>{const t=m(u(e[0])).filter((e=>null!=e));return t.every((e=>"string"!=typeof e))?t:void 0})),l=(s=([e,n,i,l])=>{if(o(),!(null==e?void 0:e.length)||!(null==n?void 0:n.length)||!(null==i?void 0:i.length))return;const r=p(l)?{...l}:l;t.push(...e.flatMap((e=>n.flatMap((t=>i.map((n=>((e,t,n,o)=>(e.addEventListener(t,n,o),()=>e.removeEventListener(t,n,o)))(e,t,n,r))))))))},v={flush:"post"},n((()=>{var t,n;return[null!=(n=null==(t=i.value)?void 0:t.map((e=>y(e))))?n:[w].filter((e=>null!=e)),m(u(i.value?e[1]:e[0])),m(a(i.value?e[2]:e[1])),u(i.value?e[3]:e[2])]}),s,{...v,immediate:!0}));var s,v;return c(o),()=>{l(),o()}}function g(e){const t=function(){const e=s(!1),t=l();return t&&o((()=>{e.value=!0}),t),e}();return r((()=>(t.value,Boolean(e()))))}function O(e,t={}){const{reset:o=!0,windowResize:i=!0,windowScroll:l=!0,immediate:a=!0,updateTiming:v="sync"}=t,d=s(0),p=s(0),O=s(0),S=s(0),j=s(0),x=s(0),z=s(0),A=s(0);function R(){const t=y(e);if(!t)return void(o&&(d.value=0,p.value=0,O.value=0,S.value=0,j.value=0,x.value=0,z.value=0,A.value=0));const n=t.getBoundingClientRect();d.value=n.height,p.value=n.bottom,O.value=n.left,S.value=n.right,j.value=n.top,x.value=n.width,z.value=n.x,A.value=n.y}function E(){"sync"===v?R():"next-frame"===v&&requestAnimationFrame((()=>R()))}return function(e,t,o={}){const{window:i=w,...l}=o;let a;const s=g((()=>i&&"ResizeObserver"in i)),v=()=>{a&&(a.disconnect(),a=void 0)},f=r((()=>{const t=u(e);return Array.isArray(t)?t.map((e=>y(e))):[y(t)]})),d=n(f,(e=>{if(v(),s.value&&i){a=new ResizeObserver(t);for(const t of e)t&&a.observe(t,l)}}),{immediate:!0,flush:"post"}),p=()=>{v(),d()};c(p)}(e,E),n((()=>y(e)),(e=>!e&&E())),function(e,t,o={}){const{window:i=w,...l}=o;let a;const s=g((()=>i&&"MutationObserver"in i)),v=()=>{a&&(a.disconnect(),a=void 0)},d=r((()=>{const t=m(u(e)).map(y).filter(f);return new Set(t)})),p=n((()=>d.value),(e=>{v(),s.value&&e.size&&(a=new MutationObserver(t),e.forEach((e=>a.observe(e,l))))}),{immediate:!0,flush:"post"}),h=()=>{p(),v()};c(h)}(e,E,{attributeFilter:["style","class"]}),l&&b("scroll",E,{capture:!0,passive:!0}),i&&b("resize",E,{passive:!0}),h((()=>{a&&E()})),{height:d,bottom:p,left:O,right:S,top:j,width:x,x:z,y:A,update:E}}export{b as a,O as u};
|
||||
@@ -1 +0,0 @@
|
||||
import{v as e,x as t,i as n,o,n as i,h as l,B as r,Q as u,e as a,R as s}from"./vue-BPkYw06T.js";function v(n){return!!e()&&(t(n),!0)}const c="undefined"!=typeof window&&"undefined"!=typeof document;"undefined"!=typeof WorkerGlobalScope&&(globalThis,WorkerGlobalScope);const f=e=>null!=e,d=Object.prototype.toString,p=e=>"[object Object]"===d.call(e);function m(e){return Array.isArray(e)?e:[e]}function h(e,t=!0,n){l()?o(e,n):t?e():i(e)}const w=c?window:void 0;function b(e){var t;const n=u(e);return null!=(t=null==n?void 0:n.$el)?t:n}function y(...e){const t=[],o=()=>{t.forEach((e=>e())),t.length=0},i=r((()=>{const t=m(u(e[0])).filter((e=>null!=e));return t.every((e=>"string"!=typeof e))?t:void 0})),l=(s=([e,n,i,l])=>{if(o(),!(null==e?void 0:e.length)||!(null==n?void 0:n.length)||!(null==i?void 0:i.length))return;const r=p(l)?{...l}:l;t.push(...e.flatMap((e=>n.flatMap((t=>i.map((n=>((e,t,n,o)=>(e.addEventListener(t,n,o),()=>e.removeEventListener(t,n,o)))(e,t,n,r))))))))},c={flush:"post"},n((()=>{var t,n;return[null!=(n=null==(t=i.value)?void 0:t.map((e=>b(e))))?n:[w].filter((e=>null!=e)),m(u(i.value?e[1]:e[0])),m(a(i.value?e[2]:e[1])),u(i.value?e[3]:e[2])]}),s,{...c,immediate:!0}));var s,c;return v(o),()=>{l(),o()}}function g(e){const t=function(){const e=s(!1),t=l();return t&&o((()=>{e.value=!0}),t),e}();return r((()=>(t.value,Boolean(e()))))}function O(e,t={}){const{reset:o=!0,windowResize:i=!0,windowScroll:l=!0,immediate:a=!0,updateTiming:c="sync"}=t,d=s(0),p=s(0),O=s(0),x=s(0),z=s(0),A=s(0),R=s(0),S=s(0);function j(){const t=b(e);if(!t)return void(o&&(d.value=0,p.value=0,O.value=0,x.value=0,z.value=0,A.value=0,R.value=0,S.value=0));const n=t.getBoundingClientRect();d.value=n.height,p.value=n.bottom,O.value=n.left,x.value=n.right,z.value=n.top,A.value=n.width,R.value=n.x,S.value=n.y}function E(){"sync"===c?j():"next-frame"===c&&requestAnimationFrame((()=>j()))}return function(e,t,o={}){const{window:i=w,...l}=o;let a;const s=g((()=>i&&"ResizeObserver"in i)),c=()=>{a&&(a.disconnect(),a=void 0)},f=r((()=>{const t=u(e);return Array.isArray(t)?t.map((e=>b(e))):[b(t)]})),d=n(f,(e=>{if(c(),s.value&&i){a=new ResizeObserver(t);for(const t of e)t&&a.observe(t,l)}}),{immediate:!0,flush:"post"}),p=()=>{c(),d()};v(p)}(e,E),n((()=>b(e)),(e=>!e&&E())),function(e,t,o={}){const{window:i=w,...l}=o;let a;const s=g((()=>i&&"MutationObserver"in i)),c=()=>{a&&(a.disconnect(),a=void 0)},d=r((()=>{const t=m(u(e)).map(b).filter(f);return new Set(t)})),p=n((()=>d.value),(e=>{c(),s.value&&e.size&&(a=new MutationObserver(t),e.forEach((e=>a.observe(e,l))))}),{immediate:!0,flush:"post"}),h=()=>{p(),c()};v(h)}(e,E,{attributeFilter:["style","class"]}),l&&y("scroll",E,{capture:!0,passive:!0}),i&&y("resize",E,{passive:!0}),h((()=>{a&&E()})),{height:d,bottom:p,left:O,right:x,top:z,width:A,x:R,y:S,update:E}}export{y as a,O as u};
|
||||
File diff suppressed because one or more lines are too long
Reference in New Issue
Block a user