+
+
+
+
+
+
+
+ResNet).
++

+
+
+
+
![]() |
+![]() |
+![]() |
+![]() |
+
| Input | +“Clean up my image, | +“Get rid of the grain | +“Remove the strange | +
| + | it’s too fuzzy." | +in my photo" | +spots on my photo" | +
![]() |
+![]() |
+![]() |
+![]() |
+
| “Retouch this image | +“Reduce the motion | +“Please get rid of | +“Reduce the fog in | +
| and improve colors" | +in this shot" | +the raindrops" | +this landmark" | +
![]() |
+![]() |
+![]() |
+![]() |
+
| Input (RealSRSet) | ++ | #1 | +#2 | +
| + | + | + | + |
![]() |
+![]() |
+![]() |
+![]() |
+
| + | + | + | + |
![]() |
+![]() |
+![]() |
+![]() |
+
![]() |
+![]() |
+![]() |
+![]() |
+![]() |
+
![]() |
+![]() |
+![]() |
+![]() |
+![]() |
+
| Input | +AirNet | +PromptIR | ++ | Reference | +
Acc.@q losses. θ:flip(cumsum(flip(⋅))), ϕ:repeat and flattening operation (more in text).import torch
+import torch.nn.functional as F
+import F.pairwise_distance as PD
+def accuracy@q_loss(anchor, pos, q):
+# anchor:B×d; pos:B×d; q:1×1(Acc.@q)
+# B←batch size; d←feature dimension
+acc_q = torch.tensor(0) # Accumulate accuracy
+dist = PD(anchor, pos) # B×B
+for i in range(len(anchor)):
+# Target distance for ith anchor
+tgt_dist = PD(anchor[i], pos[i])# 1×1
+# Relative distance to target
+rel_dist = tgt_dist - PD(anchor[i], pos)# B×1
+signs = sigmoid(rel_dist, 0.01)
+rank_i = torch.sum(signs)# Get rank:1×1
+acc_q += sigmoid(q - rank_i, 1)
+batch_acc = acc_q / len(anchor)
+loss = -1 * batch_acc
+return loss
+# Approximate step function with Sigmoid
+def sigmoid(tensor, tau):
+tensor = tensor / tau
+y = 1.0 / (1.0 + torch.exp(-tensor))
+return y
+Acc.@qSTL-10 with ResNet-18 using batch size b = 32. Horizontal axis is relative to the wall-clock training time in seconds.
+
+
+
+in
+
+
+
+DocEdit-v2 framework performs multimodal grounding and edit command generation via Doc2Command, utilizes LLM-based Command Reformulation prompting to refine the command into LMM instruction format (<Action><Component>, <Initial State>, <Final State>), and employs LMMs to edit the HTML structure using multimodal (edit instruction and grounded RoI) prompt.
+
+
+
+
+
+
+
+
+