\n| \u8de8\u6a21\u6001\u7279\u5f81\u5bf9\u9f50\u5c42<\/td>\n | \u7b80\u5355\u5355\u5c42\u7ebf\u6027\u6295\u5f71（Linear Projection）<\/td>\n | \u4e24\u5c42\u6df1\u5ea6 MLP Adapter + \u7a7a\u95f4\u4e0b\u91c7\u6837\u6c60\u5316 (Spatial Pooling)<\/td>\n | \u5c06\u89c6\u89c9\u7279\u5f81\u5411 LLM \u6587\u672c\u8bed\u4e49\u7a7a\u95f4\u7684\u6295\u5f71\u7cbe\u5ea6\u62c9\u6ee1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n\u4e8c\u3001LLaVA AnyRes \u52a8\u6001\u7f51\u683c\u5207\u7247\u4e0e\u5168\u5c40\u53cc\u6d41\u7279\u5f81\u878d\u5408\u62d3\u6251<\/h3>\n[\u539f\u59cb\u8d85\u9ad8\u6e05\u8f93\u5165\u539f\u56fe: 1344 x 672 \u50cf\u7d20 (\u5bbd\u9ad8\u6bd4 2:1)]
\n |
\n +——————+——————+
\n | |
\n v (\u901a\u8def 1: \u63d0\u53d6\u5168\u5c40\u5b8f\u89c2\u8bed\u4e49) v (\u901a\u8def 2: \u52a8\u6001\u7f51\u683c\u5207\u7247\u63d0\u53d6\u5fae\u89c2\u7ec6\u8282)
\n[\u4e0b\u91c7\u6837\u4e3a 336×336 \u5168\u5c40\u7f29\u7565\u56fe] [\u4f9d\u636e\u5bbd\u9ad8\u6bd4\u5212\u5206\u4e3a 2 \u4e2a 336×336 \u9ad8\u6e05\u5207\u7247]
\n | |
\n v v
\n[ViT \u7f16\u7801: \u4ea7\u51fa\u5168\u5c40\u7279\u5f81 $F_{\\\\text{global}}$] [ViT \u72ec\u7acb\u7f16\u7801: \u4ea7\u51fa\u5207\u7247\u7279\u5f81 $F_{\\\\text{crop1}}, F_{\\\\text{crop2}}$]
\n | |
\n +——————+——————+
\n |
\n v (\u5728\u7279\u5f81\u7ef4\u5ea6\u6309\u7a7a\u95f4\u7f51\u683c\u62d3\u6251\u8fdb\u884c\u62fc\u63a5\u62fc\u63a5，\u6ce8\u5165\u6362\u884c\u7b26 Token)
\n+——————————————————————————-+
\n| 🌟 \u7a7a\u95f4\u6c60\u5316\u538b\u7f29\u4e0e MLP Adapter \u6a21\u6001\u5bf9\u9f50\u6295\u5f71 (Spatial Pooling & MLP Projector): |
\n| 1. \u5bf9\u6bcf\u4e2a 2×2 \u89c6\u89c9 Token \u7f51\u683c\u6267\u884c\u5377\u79ef\/\u6c60\u5316\u4e0b\u91c7\u6837 (Token \u6570\u91cf\u538b\u7f29 75%!) |
\n| 2. \u901a\u8fc7 2 \u5c42 MLP \u6fc0\u6d3b\u6295\u5f71\u5c42: $\\\\mathbf{H}_V = \\\\text{GELU}(W_1 \\\\mathbf{F}_V) W_2$|
\n| \u5c06\u89c6\u89c9\u9690\u85cf\u7ef4\u5ea6 (1024) \u6295\u5f71\u6620\u5c04\u5230 LLM \u6587\u672c\u5d4c\u5165\u7a7a\u95f4\u7ef4\u5ea6 (4096) |
\n+——————————————————————————-+
\n |
\n v
\n[\u88c5\u914d\u4e3a\u89c6\u89c9 Token \u5e8f\u5217，\u4e0e\u7528\u6237\u6587\u672c Prompt `<image>\\\\n\u8bf7\u5206\u6790\u56fe\u4e2d\u8d22\u62a5\u6570\u636e` \u6df7\u5408\u9001\u5165 LLM \u89e3\u7801!]<\/p>\n
\n\u4e09\u3001\u52a8\u6001\u7f51\u683c\u5019\u9009\u641c\u7d22（Grid Selection）\u6570\u5b66\u7b97\u6cd5<\/h3>\n\u5f53\u4efb\u610f\u5c3a\u5bf8 $W \\\\times H$ \u7684\u539f\u59cb\u56fe\u7247\u8f93\u5165\u65f6，\u7cfb\u7edf\u9884\u8bbe\u4e86\u4e00\u7ec4\u6807\u51c6\u7684\u5019\u9009\u7f51\u683c\u914d\u7f6e $\\\\mathcal{G}$（\u5982 ${ (1,1), (1,2), (2,1), (2,2), (1,3), (3,1) }$，\u6bcf\u4e2a\u683c\u5b50\u7269\u7406\u5c3a\u5bf8\u4e3a $S \\\\times S = 336 \\\\times 336$）\u3002<\/p>\n 1. \u5bfb\u627e\u6709\u6548\u5206\u8fa8\u7387\u635f\u5931\u6700\u5c0f\u7684\u6700\u4f73\u7f51\u683c<\/h4>\n\u7cfb\u7edf\u904d\u5386\u6240\u6709\u5019\u9009\u7f51\u683c $(m, n) \\\\in \\\\mathcal{G}$，\u8ba1\u7b97\u5c06\u539f\u56fe\u7f29\u653e\u5230 $mS \\\\times nS$ \u65f6\u7684\u7f29\u653e\u5c3a\u5ea6（Scale Factor）\u4e0e\u6709\u6548\u50cf\u7d20\u5229\u7528\u7387：<\/p>\n $$(m^, n^) = \\\\arg\\\\min_{(m, n) \\\\in \\\\mathcal{G}} \\\\left| \\\\frac{W}{H} – \\\\frac{m}{n} \\\\right| \\\\quad \\\\text{\u4e14} \\\\quad \\\\text{Area}(m, n) \\\\approx \\\\text{Area}(W, H)$$<\/p>\n \n- \u7269\u7406\u610f\u4e49：\u7b97\u6cd5\u81ea\u52a8\u9009\u62e9\u6700\u8d34\u5408\u539f\u56fe\u7269\u7406\u6bd4\u4f8b\u3001\u4e14\u56fe\u50cf\u53d8\u5f62\u5931\u771f\u6700\u5c0f\u7684\u5207\u7247\u7f51\u683c，\u5f7b\u5e95\u6d88\u9664\u4e86\u4f20\u7edf\u53d8\u5f62\u7f29\u653e\u9020\u6210\u7684\u7269\u4f53\u62c9\u4f38\u4e0e\u7578\u53d8！<\/li>\n<\/ul>\n
\n\u56db\u3001\u751f\u4ea7\u7ea7 PyTorch \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\u89c6\u89c9\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218\u4ee3\u7801<\/h3>\n\u4e0b\u9762\u7684\u4ee3\u7801\u5c55\u793a\u4e86\u5728 PyTorch \u4e2d\u5982\u4f55\u6784\u5efa \u52a8\u6001\u56fe\u50cf\u5207\u7247\u5668（AnyRes Slicer）\u3001ViT \u7279\u5f81\u63d0\u53d6\u6a21\u62df\u5668\u3001\u7a7a\u95f4\u6c60\u5316\u538b\u7f29\u5668\u4e0e MLP \u6a21\u6001\u6295\u5f71\u5c42\u3002<\/p>\n """
\nllava_anyres_vision_projector.py
\nLLaVA AnyRes \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\u7f51\u683c\u5207\u7247\u4e0e\u7a7a\u95f4\u6c60\u5316\u6a21\u6001\u5bf9\u9f50\u7f51\u7edc\u5b9e\u6218
\n"""<\/p>\n import math
\nimport torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F
\nfrom typing import List, Tuple<\/p>\n class AnyResImageSlicer:
\n """\u52a8\u6001\u9ad8\u5206\u8fa8\u7387\u56fe\u50cf\u5207\u7247\u5668：\u5c06\u4efb\u610f\u5c3a\u5bf8\u56fe\u7247\u5207\u5206\u4e3a\u6700\u4f73\u7f51\u683c + \u5168\u5c40\u7f29\u7565\u56fe"""<\/p>\n def __init__(self, patch_size: int = 336):
\n self.patch_size = patch_size
\n # \u9884\u8bbe\u5019\u9009\u7f51\u683c\u62d3\u6251 (\u9ad8\u5ea6\u5207\u7247\u6570, \u5bbd\u5ea6\u5207\u7247\u6570)
\n self.candidate_grids = [(1, 1), (1, 2), (2, 1), (2, 2), (1, 3), (3, 1)]<\/p>\n def select_best_grid(self, original_width: int, original_height: int) -> Tuple[int, int]:
\n """\u9009\u62e9\u4e0e\u539f\u56fe\u5bbd\u9ad8\u6bd4\u6700\u5339\u914d\u7684\u5019\u9009\u7f51\u683c"""
\n orig_aspect_ratio = original_width \/ original_height
\n best_grid = (1, 1)
\n min_diff = float("inf")<\/p>\n for grid_h, grid_w in self.candidate_grids:
\n grid_aspect_ratio = grid_w \/ grid_h
\n diff = abs(orig_aspect_ratio – grid_aspect_ratio)
\n if diff < min_diff:
\n min_diff = diff
\n best_grid = (grid_h, grid_w)<\/p>\n return best_grid<\/p>\n def slice_image_tensor(self, image_tensor: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor], Tuple[int, int]]:
\n """
\n \u8f93\u5165: [3, H, W] \u539f\u59cb\u56fe\u7247\u5f20\u91cf
\n \u8fd4\u56de: (\u5168\u5c40\u7f29\u7565\u56fe [3, 336, 336], \u5c40\u90e8\u9ad8\u6e05\u5207\u7247\u5217\u8868, \u9009\u4e2d\u7684\u7f51\u683c\u5c3a\u5bf8)
\n """
\n _, h, w = image_tensor.shape
\n grid_h, grid_w = self.select_best_grid(w, h)<\/p>\n # 1. \u901a\u8def 1: \u751f\u6210\u5168\u5c40\u7f29\u7565\u56fe
\n global_thumb = F.interpolate(
\n image_tensor.unsqueeze(0), size=(self.patch_size, self.patch_size), mode="bilinear", align_corners=False
\n ).squeeze(0)<\/p>\n # 2. \u901a\u8def 2: \u5c06\u539f\u56fe\u7f29\u653e\u5230\u76ee\u6807\u7f51\u683c\u603b\u5c3a\u5bf8\u5e76\u5207\u7247
\n target_h = grid_h * self.patch_size
\n target_w = grid_w * self.patch_size
\n resized_full = F.interpolate(
\n image_tensor.unsqueeze(0), size=(target_h, target_w), mode="bilinear", align_corners=False
\n ).squeeze(0)<\/p>\n patches = []
\n for i in range(grid_h):
\n for j in range(grid_w):
\n patch = resized_full[
\n :,
\n i * self.patch_size : (i + 1) * self.patch_size,
\n j * self.patch_size : (j + 1) * self.patch_size,
\n ]
\n patches.append(patch)<\/p>\n return global_thumb, patches, (grid_h, grid_w)<\/p>\n class MultiModalVisionProjector(nn.Module):
\n """
\n \u591a\u6a21\u6001\u8de8\u6a21\u6001\u5bf9\u9f50\u7f51\u7edc：
\n \u5305\u542b\u7a7a\u95f4\u6c60\u5316（Spatial Pooling \u538b\u7f29 Token）\u4e0e 2 \u5c42 MLP \u6295\u5f71\u5c42
\n """<\/p>\n def __init__(self, vision_dim: int = 1024, llm_dim: int = 4096):
\n super().__init__()
\n # 2×2 \u7a7a\u95f4\u6c60\u5316\u4e0b\u91c7\u6837 (\u5c06\u76f8\u90bb 4 \u4e2a\u89c6\u89c9 Token \u878d\u5408\u6210 1 \u4e2a，\u964d\u4f4e 75% \u5f00\u9500!)
\n self.pooling = nn.AvgPool2d(kernel_size=2, stride=2)
\n # MLP \u6295\u5f71\u5c42
\n self.mlp_adapter = nn.Sequential(
\n nn.Linear(vision_dim, llm_dim),
\n nn.GELU(),
\n nn.Linear(llm_dim, llm_dim)
\n )<\/p>\n def forward(self, vision_features: torch.Tensor, grid_shape: Tuple[int, int]) -> torch.Tensor:
\n """
\n vision_features: [NumPatches, SeqLenPerPatch, VisionDim]
\n \u8fd4\u56de: \u5bf9\u9f50\u5230 LLM \u8bed\u4e49\u7a7a\u95f4\u4e14\u7ecf\u8fc7\u538b\u7f29\u7684\u89c6\u89c9 Token: [TotalVisionTokens, LLM_Dim]
\n """
\n # \u5047\u8bbe\u5355\u5207\u7247\u5728 ViT \u8f93\u51fa\u4e3a 24×24 = 576 \u4e2a Token (\u5bf9\u5e94 336×336 \u56fe\u7247)
\n num_patches, seq_len, v_dim = vision_features.shape
\n side_len = int(math.sqrt(seq_len))<\/p>\n # \u8f6c\u6362\u4e3a\u7a7a\u95f4 2D \u7279\u5f81\u56fe: [NumPatches, VisionDim, 24, 24]
\n feat_2d = vision_features.permute(0, 2, 1).view(num_patches, v_dim, side_len, side_len)<\/p>\n # 🌟 \u7a7a\u95f4\u6c60\u5316\u538b\u7f29: 24×24 \u2794 12×12 (\u5355\u4e2a\u5207\u7247 Token \u6570\u4ece 576 \u538b\u7f29\u81f3 144!)
\n pooled_2d = self.pooling(feat_2d)
\n _, _, p_h, p_w = pooled_2d.shape<\/p>\n # \u5c55\u5e73\u56de\u5e8f\u5217: [NumPatches, 144, VisionDim]
\n pooled_seq = pooled_2d.view(num_patches, v_dim, p_h * p_w).permute(0, 2, 1)<\/p>\n # 🌟 \u901a\u8fc7 MLP Adapter \u6295\u5f71\u6620\u5c04\u5230 LLM \u5d4c\u5165\u7a7a\u95f4 (1024 \u2794 4096)
\n llm_aligned_tokens = self.mlp_adapter(pooled_seq)<\/p>\n # \u5408\u5e76\u6240\u6709\u5207\u7247\u4e0e\u5168\u5c40\u7279\u5f81\u4e3a\u4e00\u4e2a\u6241\u5e73\u7684 Token \u6d41
\n final_tokens = llm_aligned_tokens.reshape(-1, llm_aligned_tokens.shape[-1])
\n return final_tokens<\/p>\n if __name__ == "__main__":
\n torch.manual_seed(2026)
\n print("=================================================================")
\n print("🔬 \u918d\u9190\u5b9e\u9a8c\u5ba4：\u591a\u6a21\u6001\u5927\u6a21\u578b AnyRes \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\u4e0e\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218")
\n print("=================================================================\\\\n")<\/p>\n slicer = AnyResImageSlicer(patch_size=336)
\n projector = MultiModalVisionProjector(vision_dim=1024, llm_dim=4096)<\/p>\n # 1. \u6a21\u62df\u8f93\u5165\u4e00\u5f20 1344 x 672 (\u5bbd\u9ad8\u6bd4 2:1) \u7684\u8d85\u9ad8\u6e05\u957f\u56fe
\n mock_raw_image = torch.randn(3, 672, 1344)
\n print(f"1. [\u539f\u59cb\u56fe\u50cf\u8f93\u5165]: \u5c3a\u5bf8\u4e3a {mock_raw_image.shape[-1]} x {mock_raw_image.shape[-2]} \u50cf\u7d20 (\u9ad8\u5206\u8fa8\u7387)")<\/p>\n # 2. \u52a8\u6001\u5207\u7247
\n thumb, patches, selected_grid = slicer.slice_image_tensor(mock_raw_image)
\n print(f"2. [AnyRes \u81ea\u52a8\u5207\u7247]:")
\n print(f" – \u6700\u4f73\u5339\u914d\u7f51\u683c: {selected_grid[0]} \u884c x {selected_grid[1]} \u5217 (\u5171 {len(patches)} \u4e2a\u9ad8\u6e05\u5207\u7247)")
\n print(f" – \u5168\u5c40\u7f29\u7565\u56fe\u5c3a\u5bf8: {thumb.shape}")
\n print(f" – \u6bcf\u4e2a\u9ad8\u6e05\u5207\u7247\u5c3a\u5bf8: {patches[0].shape}")<\/p>\n # 3. \u6a21\u62df ViT \u7279\u5f81\u63d0\u53d6 (\u5168\u5c40\u56fe + 2 \u4e2a\u5207\u7247 = 3 \u4e2a Patch)
\n total_crops = [thumb] + patches # \u5f62\u72b6: 3 \u4e2a [3, 336, 336]
\n # \u6a21\u62df ViT \u4e3a\u6bcf\u4e2a\u5207\u7247\u8f93\u51fa 24×24 = 576 \u4e2a 1024 \u7ef4\u7279\u5f81
\n mock_vit_output = torch.randn(len(total_crops), 576, 1024)<\/p>\n # 4. \u7a7a\u95f4\u6c60\u5316\u538b\u7f29\u4e0e\u6a21\u6001\u5bf9\u9f50\u6295\u5f71
\n aligned_llm_tokens = projector(mock_vit_output, selected_grid)<\/p>\n print(f"\\\\n3. [\u8de8\u6a21\u6001\u7279\u5f81\u5bf9\u9f50\u4e0e\u7a7a\u95f4\u538b\u7f29\u7ed3\u679c]:")
\n print(f" – \u539f\u59cb\u672a\u7ecf\u538b\u7f29\u7684\u89c6\u89c9 Token \u6570: {len(total_crops) * 576} \u4e2a")
\n print(f" – \u7a7a\u95f4\u6c60\u5316\u540e\u9001\u5165 LLM \u7684\u89c6\u89c9 Token \u6570: {aligned_llm_tokens.shape[0]} \u4e2a (\u538b\u7f29\u7387 75%!)")
\n print(f" – \u6700\u7ec8\u89c6\u89c9 Token \u5d4c\u5165\u7ef4\u5ea6: {aligned_llm_tokens.shape[1]} (\u5b8c\u7f8e\u5bf9\u9f50 LLM \u7684 4096 \u7ef4\u8bcd\u8868\u7a7a\u95f4)")<\/p>\n print("\\\\n💡 \u9a8c\u8bc1\u7ed3\u8bba：AnyRes \u6210\u529f\u517c\u987e\u4e86\u9ad8\u5206\u8fa8\u7387\u7ec6\u8282\u6355\u83b7\u4e0e\u89c6\u89c9 Token \u6570\u91cf\u7684\u5927\u5e45\u7cbe\u7b80！")
\n print("=================================================================")<\/p>\n
\n\u4e94\u3001\u591a\u6a21\u6001\u7cfb\u7edf\u751f\u4ea7\u7ea7\u843d\u5730\u907f\u5751\u4e0e\u8c03\u4f18\u7ea2\u7ebf<\/h3>\n\u5728\u751f\u4ea7\u4e2d\u90e8\u7f72\u4e0e\u5fae\u8c03\u9ad8\u5206\u8fa8\u7387\u591a\u6a21\u6001\u5927\u6a21\u578b\u65f6，\u5fc5\u987b\u4e25\u683c\u628a\u63a7\u4ee5\u4e0b\u56db\u9879\u5de5\u7a0b\u7ea2\u7ebf：<\/p>\n - \u5207\u7247\u4e4b\u95f4\u5fc5\u987b\u663e\u5f0f\u6ce8\u5165\u201c\u56fe\u50cf\u6362\u884c\u7b26 Token（Image Newline Token）\u201d：\u5f53\u5c06 2D \u7f51\u683c\u5207\u7247\u5c55\u5e73\u6210 1D Token \u5e8f\u5217\u9001\u5165 LLM \u65f6，\u5fc5\u987b\u5728\u6bcf\u4e00\u884c\u7684\u672b\u5c3e\u663e\u5f0f\u63d2\u5165\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u7279\u6b8a\u6362\u884c\u6807\u8bb0（\u5982 <image_newline>），\u5426\u5219 LLM \u4f1a\u4e27\u5931\u5bf9\u7a7a\u95f4 2D \u51e0\u4f55\u90bb\u63a5\u5173\u7cfb\u7684\u611f\u77e5，\u5bfc\u81f4\u56fe\u8868\u5750\u6807\u5b9a\u4f4d\u5b8c\u5168\u6df7\u4e71；<\/li>\n
- \u9650\u5236\u6700\u5927\u5207\u7247\u7f51\u683c\u6570（\u63a8\u8350\u4e0a\u9650\u4e3a $2 \\\\times 2$ \u6216 $3 \\\\times 3$）：\u5bf9\u4e8e\u6781\u7aef\u8d85\u957f\u56fe（\u5982 $10000 \\\\times 500$ \u7684\u957f\u957f\u7f51\u9875\u622a\u56fe），\u82e5\u751f\u6210 30 \u4e2a\u5207\u7247，\u5373\u4f7f\u6c60\u5316\u4e5f\u4f1a\u751f\u6210\u6570\u5343\u4e2a Token\u3002\u5fc5\u987b\u8bbe\u5b9a\u6700\u5927\u5207\u7247\u4e0a\u9650（Max Patches $\\\\le 6$），\u8d85\u51fa\u90e8\u5206\u81ea\u52a8\u8fdb\u884c\u7b49\u6bd4\u5e73\u6ed1\u9884\u4e0b\u91c7\u6837；<\/li>\n
- \u5fae\u8c03\u521d\u671f\u5fc5\u987b\u51bb\u7ed3 ViT \u89c6\u89c9\u7f16\u7801\u5668：\u5728\u8de8\u6a21\u6001\u5bf9\u9f50\u5fae\u8c03（Stage 1）\u9636\u6bb5，\u7edd\u5bf9\u7981\u6b62\u5f00\u542f ViT \u5168\u53c2\u6570\u5fae\u8c03！\u4ec5\u5fae\u8c03 MLP Projector \u6743\u91cd，\u5f85\u5bf9\u9f50\u635f\u5931\u5145\u5206\u6536\u655b\u540e，\u518d\u5728\u6307\u4ee4\u5fae\u8c03（Stage 2）\u9636\u6bb5\u4ee5\u6781\u5c0f\u5b66\u4e60\u7387（\u5982 $2\\\\text{e-}6$）\u8fdb\u884c\u7aef\u5230\u7aef\u8054\u5408\u5fae\u8c03，\u9632\u6b62\u7834\u574f ViT \u9884\u8bad\u7ec3\u7684\u5e95\u5c42\u89c6\u89c9\u8868\u5f81\u3002<\/li>\n
\u901a\u8fc7\u6784\u5efa\u57fa\u4e8e AnyRes \u52a8\u6001\u7f51\u683c\u5207\u7247\u7684\u5fae\u89c2\u7ec6\u8282\u611f\u77e5\u901a\u8def\u4e0e\u7a7a\u95f4\u6c60\u5316\u538b\u7f29\u7684\u8de8\u6a21\u6001\u6295\u5f71\u5c42，\u591a\u6a21\u6001\u5927\u6a21\u578b\u67b6\u6784\u80fd\u591f\u5f7b\u5e95\u6323\u8131\u4f20\u7edf\u56fa\u5b9a\u5206\u8fa8\u7387\u7684\u7269\u7406\u67b7\u9501，\u5728\u5de5\u4e1a\u7ea7\u590d\u6742\u89c6\u89c9\u8ba4\u77e5\u4e0e\u9ad8\u7cbe\u5ea6\u5bc6\u96c6\u6587\u6863\u89e3\u6790\u4e2d\u5c55\u73b0\u51fa\u4ee4\u4eba\u60ca\u53f9\u7684\u6d1e\u5bdf\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406：\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387（AnyRes）\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218\u5728\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b（Vision-Language Models – VLM，\u5982 GPT-4o\u3001Claude-3.5-Sonnet\u3001LLaVA \u7cfb\u5217）\u5168\u9762\u6e17\u900f\u5de5\u4e1a\u8d28\u68c0\u3001\u533b\u7597\u5f71\u50cf\u5224\u8bfb\u3001\u5bc6\u96c6\u6587\u6863 OCR \u4e0e\u590d\u6742\u56fe\u8868\u5206\u6790\u7684\u4eca\u5929，\u8bb8\u591a\u5f00\u53d1\u56e2\u961f\u5728\u5c1d\u8bd5\u672c\u5730\u5fae\u8c03\u6216\u63a8\u7406\u591a\u6a21\u6001\u5927\u6a21\u578b\u65f6&#<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50],"topic":[],"class_list":["post-103196","post","type-post","status-publish","format-standard","hentry","category-server","tag-50"],"yoast_head":"\n \u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/103196.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406：\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387（AnyRes）\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218\u5728\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b（Vision-Language Models - VLM，\u5982 GPT-4o\u3001Claude-3.5-Sonnet\u3001LLaVA \u7cfb\u5217）\u5168\u9762\u6e17\u900f\u5de5\u4e1a\u8d28\u68c0\u3001\u533b\u7597\u5f71\u50cf\u5224\u8bfb\u3001\u5bc6\u96c6\u6587\u6863 OCR \u4e0e\u590d\u6742\u56fe\u8868\u5206\u6790\u7684\u4eca\u5929，\u8bb8\u591a\u5f00\u53d1\u56e2\u961f\u5728\u5c1d\u8bd5\u672c\u5730\u5fae\u8c03\u6216\u63a8\u7406\u591a\u6a21\u6001\u5927\u6a21\u578b\u65f6&#\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/103196.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-09T18:09:04+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260909180903-6aa1a0bf75252.webp\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"5 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/103196.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/103196.html\",\"name\":\"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-09T18:09:04+00:00\",\"dateModified\":\"2026-09-09T18:09:04+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/103196.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/103196.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/103196.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/103196.html","og_locale":"zh_CN","og_type":"article","og_title":"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406：\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387（AnyRes）\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218\u5728\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b（Vision-Language Models - VLM，\u5982 GPT-4o\u3001Claude-3.5-Sonnet\u3001LLaVA \u7cfb\u5217）\u5168\u9762\u6e17\u900f\u5de5\u4e1a\u8d28\u68c0\u3001\u533b\u7597\u5f71\u50cf\u5224\u8bfb\u3001\u5bc6\u96c6\u6587\u6863 OCR \u4e0e\u590d\u6742\u56fe\u8868\u5206\u6790\u7684\u4eca\u5929，\u8bb8\u591a\u5f00\u53d1\u56e2\u961f\u5728\u5c1d\u8bd5\u672c\u5730\u5fae\u8c03\u6216\u63a8\u7406\u591a\u6a21\u6001\u5927\u6a21\u578b\u65f6&#","og_url":"https:\/\/www.wsisp.com\/helps\/103196.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-09T18:09:04+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260909180903-6aa1a0bf75252.webp"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"5 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/103196.html","url":"https:\/\/www.wsisp.com\/helps\/103196.html","name":"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-09T18:09:04+00:00","dateModified":"2026-09-09T18:09:04+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/103196.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/103196.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/103196.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u591a\u6a21\u6001\u89c6\u89c9\u5927\u6a21\u578b\u5e95\u5c42\u673a\u7406\uff1a\u4ece ViT Patch \u6295\u5f71\u5230 LLaVA \u52a8\u6001\u9ad8\u5206\u8fa8\u7387\uff08AnyRes\uff09\u7279\u5f81\u5bf9\u9f50\u5b9e\u6218"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/103196","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=103196"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/103196\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=103196"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=103196"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=103196"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=103196"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}} |
![\"\u5c01\u9762\u4fe1\u606f\u56fe\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260909180903-6aa1a0bf75252.webp\")
<\/p>\n