diff --git a/comfy/ldm/krea2/model.py b/comfy/ldm/krea2/model.py index ecb16254f..d4081eb7c 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): @@ -253,21 +296,47 @@ class SingleStreamDiT(nn.Module): context = self.txtmlp(context) txtlen, imglen = context.shape[1], img.shape[1] - combined = torch.cat((context, img), dim=1) + + ref_latents = kwargs.get("ref_latents", None) + ref_latents_method = kwargs.get("ref_latents_method", "index_timestep_zero") + if ref_latents_method is None: + ref_latents_method = "index_timestep_zero" + if ref_latents_method != "index_timestep_zero": + raise ValueError(f"Unsupported Krea2 reference latent method: {ref_latents_method}") + device = img.device + + ref_tokens_list, ref_pos_ids_list, ref_num_tokens = self._process_ref_latents( + ref_latents, device, bs, h_, w_, img.dtype + ) + + if len(ref_num_tokens) > 0: + transformer_options = transformer_options.copy() + if "reference_image_num_tokens" not in transformer_options: + transformer_options["reference_image_num_tokens"] = [] + transformer_options["reference_image_num_tokens"].extend(ref_num_tokens) + + combined = torch.cat([context, img] + ref_tokens_list, dim=1) # Position ids: text at 0, image at (0, h_idx, w_idx). - device = combined.device txtpos = torch.zeros(bs, txtlen, 3, device=device, dtype=torch.float32) imgids = torch.zeros(h_, w_, 3, device=device, dtype=torch.float32) imgids[..., 1] = torch.arange(h_, device=device, dtype=torch.float32)[:, None] imgids[..., 2] = torch.arange(w_, device=device, dtype=torch.float32)[None, :] imgpos = imgids.reshape(1, h_ * w_, 3).repeat(bs, 1, 1) - pos = torch.cat((txtpos, imgpos), dim=1) + pos = torch.cat([txtpos, imgpos] + ref_pos_ids_list, dim=1) 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, :] @@ -288,3 +357,34 @@ class SingleStreamDiT(nn.Module): f"Load the text encoder with CLIPLoader type 'krea2'." ) return context.reshape(b, seq, self.txtlayers, self.txtdim) + + def _process_ref_latents(self, ref_latents, 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: + 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.movedim(2, 1).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_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) + ref_tokens_list.append(ref_tokens) + 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] = 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) + + return ref_tokens_list, ref_pos_ids_list, ref_num_tokens diff --git a/comfy/model_base.py b/comfy/model_base.py index 98f5ba48b..f1d71f77c 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -2317,12 +2317,31 @@ class Ideogram4(BaseModel): class Krea2(BaseModel): def __init__(self, model_config, model_type=ModelType.FLUX, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.krea2.model.SingleStreamDiT) + self.memory_usage_factor_conds = ("ref_latents",) def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) cross_attn = kwargs.get("cross_attn", None) if cross_attn is not None: out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + ref_latents = kwargs.get("reference_latents", None) + if ref_latents is not None: + latents = [] + for lat in ref_latents: + latents.append(self.process_latent_in(lat)) + out['ref_latents'] = comfy.conds.CONDList(latents) + + ref_latents_method = kwargs.get("reference_latents_method", None) + if ref_latents_method is not None: + out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method) + return out + + def extra_conds_shapes(self, **kwargs): + out = {} + ref_latents = kwargs.get("reference_latents", None) + if ref_latents is not None: + out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()[2:]), ref_latents))]) return out class HunyuanImage21(BaseModel):