diff --git a/comfy/ldm/krea2/model.py b/comfy/ldm/krea2/model.py index 5f3c801c3..78e7e1952 100644 --- a/comfy/ldm/krea2/model.py +++ b/comfy/ldm/krea2/model.py @@ -158,11 +158,54 @@ class SingleStreamBlock(nn.Module): self.attn = Attention(features, heads, kvheads=kvheads, bias=bias, device=device, dtype=dtype, operations=operations) self.mlp = SwiGLU(features, multiplier, bias, device=device, dtype=dtype, operations=operations) - def forward(self, x, vec, freqs, mask=None, transformer_options={}): - prescale, preshift, pregate, postscale, postshift, postgate = self.mod(vec) - x = x + pregate * self.attn((1 + prescale) * self.prenorm(x) + preshift, freqs, mask, transformer_options=transformer_options) - x = x + postgate * self.mlp((1 + postscale) * self.postnorm(x) + postshift) - return x + def forward(self, x, vec, freqs, mask=None, transformer_options={}, vec_ref=None, split=None): + if vec_ref is not None and split is not None: + m = self.mod(vec) + r = self.mod(vec_ref) + + # prenorm and attention + h = self.prenorm(x) + h_mod = torch.cat( + ( + (1 + m[0]) * h[:, :split] + m[1], + (1 + r[0]) * h[:, split:] + r[1] + ), + dim=1 + ) + attn_out = self.attn(h_mod, freqs, mask, transformer_options=transformer_options) + attn_gate = torch.cat( + ( + m[2] * attn_out[:, :split], + r[2] * attn_out[:, split:] + ), + dim=1 + ) + x = x + attn_gate + + # postnorm and mlp + h = self.postnorm(x) + h_mod = torch.cat( + ( + (1 + m[3]) * h[:, :split] + m[4], + (1 + r[3]) * h[:, split:] + r[4] + ), + dim=1 + ) + mlp_out = self.mlp(h_mod) + mlp_gate = torch.cat( + ( + m[5] * mlp_out[:, :split], + r[5] * mlp_out[:, split:] + ), + dim=1 + ) + x = x + mlp_gate + return x + else: + prescale, preshift, pregate, postscale, postshift, postgate = self.mod(vec) + x = x + pregate * self.attn((1 + prescale) * self.prenorm(x) + preshift, freqs, mask, transformer_options=transformer_options) + x = x + postgate * self.mlp((1 + postscale) * self.postnorm(x) + postshift) + return x class LastLayer(nn.Module): @@ -259,7 +302,7 @@ class SingleStreamDiT(nn.Module): device = img.device ref_tokens_list, ref_pos_ids_list, ref_num_tokens = self._process_ref_latents( - ref_latents, ref_latents_method, device, bs, h_, w_ + ref_latents, ref_latents_method, device, bs, h_, w_, img.dtype ) if len(ref_num_tokens) > 0: @@ -280,8 +323,16 @@ class SingleStreamDiT(nn.Module): freqs = self.pe_embedder(pos) - for block in self.blocks: - combined = block(combined, tvec, freqs, None, transformer_options=transformer_options) + if len(ref_num_tokens) > 0: + # Compute tvec0 for timestep=0 (reference) + t0 = self.tmlp(timestep_embedding(torch.zeros_like(timesteps), self.tdim).unsqueeze(1).to(img.dtype)) + tvec0 = self.tproj(t0) + split = txtlen + imglen + for block in self.blocks: + combined = block(combined, tvec, freqs, None, transformer_options=transformer_options, vec_ref=tvec0, split=split) + else: + for block in self.blocks: + combined = block(combined, tvec, freqs, None, transformer_options=transformer_options) final = self.last(combined, t) out = final[:, txtlen:txtlen + imglen, :] @@ -303,43 +354,22 @@ class SingleStreamDiT(nn.Module): ) return context.reshape(b, seq, self.txtlayers, self.txtdim) - def _process_ref_latents(self, ref_latents, ref_latents_method, device, bs, h_main, w_main): + def _process_ref_latents(self, ref_latents, ref_latents_method, device, bs, h_main, w_main, dtype): ref_tokens_list = [] ref_pos_ids_list = [] ref_num_tokens = [] patch = self.patch if ref_latents is not None: - h = h_main - w = w_main - index = 0 - index_ref_method = (ref_latents_method == "index") or (ref_latents_method == "index_timestep_zero") - negative_ref_method = ref_latents_method == "negative_index" - - for ref in ref_latents: + for i, ref in enumerate(ref_latents): if ref.ndim == 5: ref_b5, ref_c5, ref_t5, ref_h5, ref_w5 = ref.shape ref = ref.reshape(ref_b5 * ref_t5, ref_c5, ref_h5, ref_w5) ref_pad = comfy.ldm.common_dit.pad_to_patch_size(ref, (patch, patch)) - ref_b, ref_c, ref_h, ref_w = ref_pad.shape - ref_gh = ref_h // patch - ref_gw = ref_w // patch - - if index_ref_method: - index += 1 - elif negative_ref_method: - index -= 1 - else: # offset/default - index = 0 # Stay in t_idx = 0 plane to remain 100% in-distribution - - gh_offset = 0 - gw_offset = 0 - if ref_gh + h > ref_gw + w: - gw_offset = w - else: - gh_offset = h - h = max(h, ref_gh + gh_offset) - w = max(w, ref_gw + gw_offset) + ref_pad = comfy.utils.repeat_to_batch_size(ref_pad, bs) + ref_pad = ref_pad.to(device=device, dtype=dtype) + ref_gh = ref_pad.shape[-2] // patch + ref_gw = ref_pad.shape[-1] // patch ref_tokens = rearrange(ref_pad, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch, pw=patch) ref_tokens = self.first(ref_tokens) @@ -347,9 +377,9 @@ class SingleStreamDiT(nn.Module): ref_num_tokens.append(ref_tokens.shape[1]) ref_pos_ids = torch.zeros(ref_gh, ref_gw, 3, device=device, dtype=torch.float32) - ref_pos_ids[..., 0] = index - ref_pos_ids[..., 1] = torch.arange(ref_gh, device=device, dtype=torch.float32)[:, None] + gh_offset - ref_pos_ids[..., 2] = torch.arange(ref_gw, device=device, dtype=torch.float32)[None, :] + gw_offset + ref_pos_ids[..., 0] = i + 1.0 + ref_pos_ids[..., 1] = torch.arange(ref_gh, device=device, dtype=torch.float32)[:, None] + ref_pos_ids[..., 2] = torch.arange(ref_gw, device=device, dtype=torch.float32)[None, :] ref_pos_ids = ref_pos_ids.reshape(1, ref_gh * ref_gw, 3).repeat(bs, 1, 1) ref_pos_ids_list.append(ref_pos_ids)