init

utils/attn_utils.py

+# *************************************************************************
+# This file may have been modified by Bytedance Inc. (“Bytedance Inc.'s Mo-
+# difications”). All Bytedance Inc.'s Modifications are Copyright (2023) B-
+# ytedance Inc..  
+# *************************************************************************
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from einops import rearrange, repeat
+
+
+class AttentionBase:
+
+    def __init__(self):
+        self.cur_step = 0
+        self.num_att_layers = -1
+        self.cur_att_layer = 0
+
+    def after_step(self):
+        pass
+
+    def __call__(self, q, k, v, is_cross, place_in_unet, num_heads, **kwargs):
+        out = self.forward(q, k, v, is_cross, place_in_unet, num_heads, **kwargs)
+        self.cur_att_layer += 1
+        if self.cur_att_layer == self.num_att_layers:
+            self.cur_att_layer = 0
+            self.cur_step += 1
+            # after step
+            self.after_step()
+        return out
+
+    def forward(self, q, k, v, is_cross, place_in_unet, num_heads, **kwargs):
+        out = F.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return out
+
+    def reset(self):
+        self.cur_step = 0
+        self.cur_att_layer = 0
+
+
+class MutualSelfAttentionControl(AttentionBase):
+
+    def __init__(self, start_step=4, start_layer=10, layer_idx=None, step_idx=None, total_steps=50, guidance_scale=7.5):
+        """
+        Mutual self-attention control for Stable-Diffusion model
+        Args:
+            start_step: the step to start mutual self-attention control
+            start_layer: the layer to start mutual self-attention control
+            layer_idx: list of the layers to apply mutual self-attention control
+            step_idx: list the steps to apply mutual self-attention control
+            total_steps: the total number of steps
+        """
+        super().__init__()
+        self.total_steps = total_steps
+        self.start_step = start_step
+        self.start_layer = start_layer
+        self.layer_idx = layer_idx if layer_idx is not None else list(range(start_layer, 16))
+        self.step_idx = step_idx if step_idx is not None else list(range(start_step, total_steps))
+        # store the guidance scale to decide whether there are unconditional branch
+        self.guidance_scale = guidance_scale
+        print("step_idx: ", self.step_idx)
+        print("layer_idx: ", self.layer_idx)
+
+    def forward(self, q, k, v, is_cross, place_in_unet, num_heads, **kwargs):
+        """
+        Attention forward function
+        """
+        if is_cross or self.cur_step not in self.step_idx or self.cur_att_layer // 2 not in self.layer_idx:
+            return super().forward(q, k, v, is_cross, place_in_unet, num_heads, **kwargs)
+
+        if self.guidance_scale > 1.0:
+            qu, qc = q[0:2], q[2:4]
+            ku, kc = k[0:2], k[2:4]
+            vu, vc = v[0:2], v[2:4]
+
+            # merge queries of source and target branch into one so we can use torch API
+            qu = torch.cat([qu[0:1], qu[1:2]], dim=2)
+            qc = torch.cat([qc[0:1], qc[1:2]], dim=2)
+
+            out_u = F.scaled_dot_product_attention(qu, ku[0:1], vu[0:1], attn_mask=None, dropout_p=0.0, is_causal=False)
+            out_u = torch.cat(out_u.chunk(2, dim=2), dim=0) # split the queries into source and target batch
+            out_u = rearrange(out_u, 'b h n d -> b n (h d)')
+
+            out_c = F.scaled_dot_product_attention(qc, kc[0:1], vc[0:1], attn_mask=None, dropout_p=0.0, is_causal=False)
+            out_c = torch.cat(out_c.chunk(2, dim=2), dim=0) # split the queries into source and target batch
+            out_c = rearrange(out_c, 'b h n d -> b n (h d)')
+
+            out = torch.cat([out_u, out_c], dim=0)
+        else:
+            q = torch.cat([q[0:1], q[1:2]], dim=2)
+            out = F.scaled_dot_product_attention(q, k[0:1], v[0:1], attn_mask=None, dropout_p=0.0, is_causal=False)
+            out = torch.cat(out.chunk(2, dim=2), dim=0) # split the queries into source and target batch
+            out = rearrange(out, 'b h n d -> b n (h d)')
+        return out
+
+# forward function for default attention processor
+# modified from __call__ function of AttnProcessor in diffusers
+def override_attn_proc_forward(attn, editor, place_in_unet):
+    def forward(x, encoder_hidden_states=None, attention_mask=None, context=None, mask=None):
+        """
+        The attention is similar to the original implementation of LDM CrossAttention class
+        except adding some modifications on the attention
+        """
+        if encoder_hidden_states is not None:
+            context = encoder_hidden_states
+        if attention_mask is not None:
+            mask = attention_mask
+
+        to_out = attn.to_out
+        if isinstance(to_out, nn.modules.container.ModuleList):
+            to_out = attn.to_out[0]
+        else:
+            to_out = attn.to_out
+
+        h = attn.heads
+        q = attn.to_q(x)
+        is_cross = context is not None
+        context = context if is_cross else x
+        k = attn.to_k(context)
+        v = attn.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
+
+        # the only difference
+        out = editor(
+            q, k, v, is_cross, place_in_unet,
+            attn.heads, scale=attn.scale)
+
+        return to_out(out)
+
+    return forward
+
+# forward function for lora attention processor
+# modified from __call__ function of LoRAAttnProcessor2_0 in diffusers v0.17.1
+def override_lora_attn_proc_forward(attn, editor, place_in_unet):
+    def forward(hidden_states, encoder_hidden_states=None, attention_mask=None):
+        residual = hidden_states
+        input_ndim = hidden_states.ndim
+        is_cross = encoder_hidden_states is not None
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        # query = attn.to_q(hidden_states) + lora_scale * attn.to_q.lora_layer(hidden_states)
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        # key = attn.to_k(encoder_hidden_states) + lora_scale * attn.to_k.lora_layer(encoder_hidden_states)
+        # value = attn.to_v(encoder_hidden_states) + lora_scale * attn.to_v.lora_layer(encoder_hidden_states)
+        key, value = attn.to_k(encoder_hidden_states), attn.to_v(encoder_hidden_states)
+
+        query, key, value = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=attn.heads), (query, key, value))
+
+        # the only difference
+        hidden_states = editor(
+            query, key, value, is_cross, place_in_unet,
+            attn.heads, scale=attn.scale)
+
+        # linear proj
+        # hidden_states = attn.to_out[0](hidden_states) + lora_scale * attn.to_out[0].lora_layer(hidden_states)
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+    return forward
+
+def register_attention_editor_diffusers(model, editor: AttentionBase, attn_processor='attn_proc'):
+    """
+    Register a attention editor to Diffuser Pipeline, refer from [Prompt-to-Prompt]
+    """
+    def register_editor(net, count, place_in_unet):
+        for name, subnet in net.named_children():
+            if net.__class__.__name__ == 'Attention':  # spatial Transformer layer
+                if attn_processor == 'attn_proc':
+                    net.forward = override_attn_proc_forward(net, editor, place_in_unet)
+                elif attn_processor == 'lora_attn_proc':
+                    net.forward = override_lora_attn_proc_forward(net, editor, place_in_unet)
+                else:
+                    raise NotImplementedError("not implemented")
+                return count + 1
+            elif hasattr(net, 'children'):
+                count = register_editor(subnet, count, place_in_unet)
+        return count
+
+    cross_att_count = 0
+    for net_name, net in model.unet.named_children():
+        if "down" in net_name:
+            cross_att_count += register_editor(net, 0, "down")
+        elif "mid" in net_name:
+            cross_att_count += register_editor(net, 0, "mid")
+        elif "up" in net_name:
+            cross_att_count += register_editor(net, 0, "up")
+    editor.num_att_layers = cross_att_count

utils/drag_utils.py

+# *************************************************************************
+# Copyright (2023) Bytedance Inc.
+#
+# Copyright (2023) DragDiffusion Authors 
+#
+# Licensed under the Apache License, Version 2.0 (the "License"); 
+# you may not use this file except in compliance with the License. 
+# You may obtain a copy of the License at 
+#
+#     http://www.apache.org/licenses/LICENSE-2.0 
+#
+# Unless required by applicable law or agreed to in writing, software 
+# distributed under the License is distributed on an "AS IS" BASIS, 
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 
+# See the License for the specific language governing permissions and 
+# limitations under the License. 
+# *************************************************************************
+
+import copy
+import torch
+import torch.nn.functional as F
+
+
+def point_tracking(F0,
+                   F1,
+                   handle_points,
+                   handle_points_init,
+                   args):
+    with torch.no_grad():
+        _, _, max_r, max_c = F0.shape
+        for i in range(len(handle_points)):
+            pi0, pi = handle_points_init[i], handle_points[i]
+            f0 = F0[:, :, int(pi0[0]), int(pi0[1])]
+
+            r1, r2 = max(0,int(pi[0])-args.r_p), min(max_r,int(pi[0])+args.r_p+1)
+            c1, c2 = max(0,int(pi[1])-args.r_p), min(max_c,int(pi[1])+args.r_p+1)
+            F1_neighbor = F1[:, :, r1:r2, c1:c2]
+            all_dist = (f0.unsqueeze(dim=-1).unsqueeze(dim=-1) - F1_neighbor).abs().sum(dim=1)
+            all_dist = all_dist.squeeze(dim=0)
+            row, col = divmod(all_dist.argmin().item(), all_dist.shape[-1])
+            # handle_points[i][0] = pi[0] - args.r_p + row
+            # handle_points[i][1] = pi[1] - args.r_p + col
+            handle_points[i][0] = r1 + row
+            handle_points[i][1] = c1 + col
+        return handle_points
+
+def check_handle_reach_target(handle_points,
+                              target_points):
+    # dist = (torch.cat(handle_points,dim=0) - torch.cat(target_points,dim=0)).norm(dim=-1)
+    all_dist = list(map(lambda p,q: (p-q).norm(), handle_points, target_points))
+    return (torch.tensor(all_dist) < 2.0).all()
+
+# obtain the bilinear interpolated feature patch centered around (x, y) with radius r
+def interpolate_feature_patch(feat,
+                              y1,
+                              y2,
+                              x1,
+                              x2):
+    x1_floor = torch.floor(x1).long()
+    x1_cell = x1_floor + 1
+    dx = torch.floor(x2).long() - torch.floor(x1).long()
+
+    y1_floor = torch.floor(y1).long()
+    y1_cell = y1_floor + 1
+    dy = torch.floor(y2).long() - torch.floor(y1).long()
+
+    wa = (x1_cell.float() - x1) * (y1_cell.float() - y1)
+    wb = (x1_cell.float() - x1) * (y1 - y1_floor.float())
+    wc = (x1 - x1_floor.float()) * (y1_cell.float() - y1)
+    wd = (x1 - x1_floor.float()) * (y1 - y1_floor.float())
+
+    Ia = feat[:, :, y1_floor : y1_floor+dy, x1_floor : x1_floor+dx]
+    Ib = feat[:, :, y1_cell : y1_cell+dy, x1_floor : x1_floor+dx]
+    Ic = feat[:, :, y1_floor : y1_floor+dy, x1_cell : x1_cell+dx]
+    Id = feat[:, :, y1_cell : y1_cell+dy, x1_cell : x1_cell+dx]
+
+    return Ia * wa + Ib * wb + Ic * wc + Id * wd
+
+def drag_diffusion_update(model,
+                          init_code,
+                          text_embeddings,
+                          t,
+                          handle_points,
+                          target_points,
+                          mask,
+                          args):
+
+    assert len(handle_points) == len(target_points), \
+        "number of handle point must equals target points"
+    if text_embeddings is None:
+        text_embeddings = model.get_text_embeddings(args.prompt)
+
+    # the init output feature of unet
+    with torch.no_grad():
+        unet_output, F0 = model.forward_unet_features(init_code, t,
+            encoder_hidden_states=text_embeddings,
+            layer_idx=args.unet_feature_idx, interp_res_h=args.sup_res_h, interp_res_w=args.sup_res_w)
+        x_prev_0,_ = model.step(unet_output, t, init_code)
+        # init_code_orig = copy.deepcopy(init_code)
+
+    # prepare optimizable init_code and optimizer
+    init_code.requires_grad_(True)
+    optimizer = torch.optim.Adam([init_code], lr=args.lr)
+
+    # prepare for point tracking and background regularization
+    handle_points_init = copy.deepcopy(handle_points)
+    interp_mask = F.interpolate(mask, (init_code.shape[2],init_code.shape[3]), mode='nearest')
+    using_mask = interp_mask.sum() != 0.0
+
+    # prepare amp scaler for mixed-precision training
+    scaler = torch.cuda.amp.GradScaler()
+    for step_idx in range(args.n_pix_step):
+        with torch.autocast(device_type='cuda', dtype=torch.float16):
+            unet_output, F1 = model.forward_unet_features(init_code, t,
+                encoder_hidden_states=text_embeddings,
+                layer_idx=args.unet_feature_idx, interp_res_h=args.sup_res_h, interp_res_w=args.sup_res_w)
+            x_prev_updated,_ = model.step(unet_output, t, init_code)
+
+            # do point tracking to update handle points before computing motion supervision loss
+            if step_idx != 0:
+                handle_points = point_tracking(F0, F1, handle_points, handle_points_init, args)
+                print('new handle points', handle_points)
+
+            # break if all handle points have reached the targets
+            if check_handle_reach_target(handle_points, target_points):
+                break
+
+            loss = 0.0
+            _, _, max_r, max_c = F0.shape
+            for i in range(len(handle_points)):
+                pi, ti = handle_points[i], target_points[i]
+                # skip if the distance between target and source is less than 1
+                if (ti - pi).norm() < 2.:
+                    continue
+
+                di = (ti - pi) / (ti - pi).norm()
+
+                # motion supervision
+                # with boundary protection
+                r1, r2 = max(0,int(pi[0])-args.r_m), min(max_r,int(pi[0])+args.r_m+1)
+                c1, c2 = max(0,int(pi[1])-args.r_m), min(max_c,int(pi[1])+args.r_m+1)
+                f0_patch = F1[:,:,r1:r2, c1:c2].detach()
+                f1_patch = interpolate_feature_patch(F1,r1+di[0],r2+di[0],c1+di[1],c2+di[1])
+
+                # original code, without boundary protection
+                # f0_patch = F1[:,:,int(pi[0])-args.r_m:int(pi[0])+args.r_m+1, int(pi[1])-args.r_m:int(pi[1])+args.r_m+1].detach()
+                # f1_patch = interpolate_feature_patch(F1, pi[0] + di[0], pi[1] + di[1], args.r_m)
+                loss += ((2*args.r_m+1)**2)*F.l1_loss(f0_patch, f1_patch)
+
+            # masked region must stay unchanged
+            if using_mask:
+                loss += args.lam * ((x_prev_updated-x_prev_0)*(1.0-interp_mask)).abs().sum()
+            # loss += args.lam * ((init_code_orig-init_code)*(1.0-interp_mask)).abs().sum()
+            print('loss total=%f'%(loss.item()))
+
+        scaler.scale(loss).backward()
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+    return init_code
+
+def drag_diffusion_update_gen(model,
+                              init_code,
+                              text_embeddings,
+                              t,
+                              handle_points,
+                              target_points,
+                              mask,
+                              args):
+
+    assert len(handle_points) == len(target_points), \
+        "number of handle point must equals target points"
+    if text_embeddings is None:
+        text_embeddings = model.get_text_embeddings(args.prompt)
+
+    # positive prompt embedding
+    if args.guidance_scale > 1.0:
+        unconditional_input = model.tokenizer(
+            [args.neg_prompt],
+            padding="max_length",
+            max_length=77,
+            return_tensors="pt"
+        )
+        unconditional_emb = model.text_encoder(unconditional_input.input_ids.to(text_embeddings.device))[0].detach()
+        text_embeddings = torch.cat([unconditional_emb, text_embeddings], dim=0)
+
+    # the init output feature of unet
+    with torch.no_grad():
+        if args.guidance_scale > 1.:
+            model_inputs_0 = copy.deepcopy(torch.cat([init_code] * 2))
+        else:
+            model_inputs_0 = copy.deepcopy(init_code)
+        unet_output, F0 = model.forward_unet_features(model_inputs_0, t, encoder_hidden_states=text_embeddings,
+            layer_idx=args.unet_feature_idx, interp_res_h=args.sup_res_h, interp_res_w=args.sup_res_w)
+        if args.guidance_scale > 1.:
+            # strategy 1: discard the unconditional branch feature maps
+            # F0 = F0[1].unsqueeze(dim=0)
+            # strategy 2: concat pos and neg branch feature maps for motion-sup and point tracking
+            # F0 = torch.cat([F0[0], F0[1]], dim=0).unsqueeze(dim=0)
+            # strategy 3: concat pos and neg branch feature maps with guidance_scale consideration
+            coef = args.guidance_scale / (2*args.guidance_scale - 1.0)
+            F0 = torch.cat([(1-coef)*F0[0], coef*F0[1]], dim=0).unsqueeze(dim=0)
+
+            unet_output_uncon, unet_output_con = unet_output.chunk(2, dim=0)
+            unet_output = unet_output_uncon + args.guidance_scale * (unet_output_con - unet_output_uncon)
+        x_prev_0,_ = model.step(unet_output, t, init_code)
+        # init_code_orig = copy.deepcopy(init_code)
+
+    # prepare optimizable init_code and optimizer
+    init_code.requires_grad_(True)
+    optimizer = torch.optim.Adam([init_code], lr=args.lr)
+
+    # prepare for point tracking and background regularization
+    handle_points_init = copy.deepcopy(handle_points)
+    interp_mask = F.interpolate(mask, (init_code.shape[2],init_code.shape[3]), mode='nearest')
+    using_mask = interp_mask.sum() != 0.0
+
+    # prepare amp scaler for mixed-precision training
+    scaler = torch.cuda.amp.GradScaler()
+    for step_idx in range(args.n_pix_step):
+        with torch.autocast(device_type='cuda', dtype=torch.float16):
+            if args.guidance_scale > 1.:
+                model_inputs = init_code.repeat(2,1,1,1)
+            else:
+                model_inputs = init_code
+            unet_output, F1 = model.forward_unet_features(model_inputs, t, encoder_hidden_states=text_embeddings,
+                layer_idx=args.unet_feature_idx, interp_res_h=args.sup_res_h, interp_res_w=args.sup_res_w)
+            if args.guidance_scale > 1.:
+                # strategy 1: discard the unconditional branch feature maps
+                # F1 = F1[1].unsqueeze(dim=0)
+                # strategy 2: concat positive and negative branch feature maps for motion-sup and point tracking
+                # F1 = torch.cat([F1[0], F1[1]], dim=0).unsqueeze(dim=0)
+                # strategy 3: concat pos and neg branch feature maps with guidance_scale consideration
+                coef = args.guidance_scale / (2*args.guidance_scale - 1.0)
+                F1 = torch.cat([(1-coef)*F1[0], coef*F1[1]], dim=0).unsqueeze(dim=0)
+
+                unet_output_uncon, unet_output_con = unet_output.chunk(2, dim=0)
+                unet_output = unet_output_uncon + args.guidance_scale * (unet_output_con - unet_output_uncon)
+            x_prev_updated,_ = model.step(unet_output, t, init_code)
+
+            # do point tracking to update handle points before computing motion supervision loss
+            if step_idx != 0:
+                handle_points = point_tracking(F0, F1, handle_points, handle_points_init, args)
+                print('new handle points', handle_points)
+
+            # break if all handle points have reached the targets
+            if check_handle_reach_target(handle_points, target_points):
+                break
+
+            loss = 0.0
+            _, _, max_r, max_c = F0.shape
+            for i in range(len(handle_points)):
+                pi, ti = handle_points[i], target_points[i]
+                # skip if the distance between target and source is less than 1
+                if (ti - pi).norm() < 2.:
+                    continue
+
+                di = (ti - pi) / (ti - pi).norm()
+
+                # motion supervision
+                # with boundary protection
+                r1, r2 = max(0,int(pi[0])-args.r_m), min(max_r,int(pi[0])+args.r_m+1)
+                c1, c2 = max(0,int(pi[1])-args.r_m), min(max_c,int(pi[1])+args.r_m+1)
+                f0_patch = F1[:,:,r1:r2, c1:c2].detach()
+                f1_patch = interpolate_feature_patch(F1,r1+di[0],r2+di[0],c1+di[1],c2+di[1])
+
+                # original code, without boundary protection
+                # f0_patch = F1[:,:,int(pi[0])-args.r_m:int(pi[0])+args.r_m+1, int(pi[1])-args.r_m:int(pi[1])+args.r_m+1].detach()
+                # f1_patch = interpolate_feature_patch(F1, pi[0] + di[0], pi[1] + di[1], args.r_m)
+
+                loss += ((2*args.r_m+1)**2)*F.l1_loss(f0_patch, f1_patch)
+
+            # masked region must stay unchanged
+            if using_mask:
+                loss += args.lam * ((x_prev_updated-x_prev_0)*(1.0-interp_mask)).abs().sum()
+            # loss += args.lam * ((init_code_orig - init_code)*(1.0-interp_mask)).abs().sum()
+            print('loss total=%f'%(loss.item()))
+
+        scaler.scale(loss).backward()
+        scaler.step(optimizer)
+        scaler.update()
+        optimizer.zero_grad()
+
+    return init_code
+