Merge ba07c89407 into c4a14df9a3

For LTX Audio VAE, remove normalization of audio during MEL spectrogram creation.
This aligs inference with training and prevents loud audio from being attenuated.

README.md

@@ -108,7 +108,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/)
 - Latent previews with [TAESD](#how-to-show-high-quality-previews)
 - Works fully offline: core will never download anything unless you want to.
-- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview).
+- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview) disable with: `--disable-api-nodes`
 - [Config file](extra_model_paths.yaml.example) to set the search paths for models.
 
 Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/)
@@ -212,7 +212,7 @@ Python 3.14 works but you may encounter issues with the torch compile node. The
 
 Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
 
-torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old.
+torch 2.4 and above is supported but some features and optimizations might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old.
 
 ### Instructions:
 
@@ -229,7 +229,7 @@ AMD users can install rocm and pytorch with pip if you don't have it already ins
 
 ```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4```
 
-This is the command to install the nightly with ROCm 7.0 which might have some performance improvements:
+This is the command to install the nightly with ROCm 7.1 which might have some performance improvements:
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1```
 

comfy/ldm/lightricks/vae/audio_vae.py

@@ -103,20 +103,10 @@ class AudioPreprocessor:
             return waveform
         return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate)
 
-    @staticmethod
-    def normalize_amplitude(
-        waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5
-    ) -> torch.Tensor:
-        waveform = waveform - waveform.mean(dim=2, keepdim=True)
-        peak = torch.max(torch.abs(waveform)) + eps
-        scale = peak.clamp(max=max_amplitude) / peak
-        return waveform * scale
-
     def waveform_to_mel(
         self, waveform: torch.Tensor, waveform_sample_rate: int, device
     ) -> torch.Tensor:
         waveform = self.resample(waveform, waveform_sample_rate)
-        waveform = self.normalize_amplitude(waveform)
 
         mel_transform = torchaudio.transforms.MelSpectrogram(
             sample_rate=self.target_sample_rate,
@@ -189,9 +179,12 @@ class AudioVAE(torch.nn.Module):
         waveform = self.device_manager.move_to_load_device(waveform)
         expected_channels = self.autoencoder.encoder.in_channels
         if waveform.shape[1] != expected_channels:
-            raise ValueError(
-                f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}"
-            )
+            if waveform.shape[1] == 1:
+                waveform = waveform.expand(-1, expected_channels, *waveform.shape[2:])
+            else:
+                raise ValueError(
+                    f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}"
+                )
 
         mel_spec = self.preprocessor.waveform_to_mel(
             waveform, waveform_sample_rate, device=self.device_manager.load_device

comfy/ldm/lumina/model.py

@@ -13,10 +13,53 @@ from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
 from comfy.ldm.flux.math import apply_rope
 import comfy.patcher_extension
+import comfy.utils
 
 
-def modulate(x, scale):
-    return x * (1 + scale.unsqueeze(1))
+def invert_slices(slices, length):
+    sorted_slices = sorted(slices)
+    result = []
+    current = 0
+
+    for start, end in sorted_slices:
+        if current < start:
+            result.append((current, start))
+        current = max(current, end)
+
+    if current < length:
+        result.append((current, length))
+
+    return result
+
+
+def modulate(x, scale, timestep_zero_index=None):
+    if timestep_zero_index is None:
+        return x * (1 + scale.unsqueeze(1))
+    else:
+        scale = (1 + scale.unsqueeze(1))
+        actual_batch = scale.size(0) // 2
+        slices = timestep_zero_index
+        invert = invert_slices(timestep_zero_index, x.shape[1])
+        for s in slices:
+            x[:, s[0]:s[1]] *= scale[actual_batch:]
+        for s in invert:
+            x[:, s[0]:s[1]] *= scale[:actual_batch]
+        return x
+
+
+def apply_gate(gate, x, timestep_zero_index=None):
+    if timestep_zero_index is None:
+        return gate * x
+    else:
+        actual_batch = gate.size(0) // 2
+
+        slices = timestep_zero_index
+        invert = invert_slices(timestep_zero_index, x.shape[1])
+        for s in slices:
+            x[:, s[0]:s[1]] *= gate[actual_batch:]
+        for s in invert:
+            x[:, s[0]:s[1]] *= gate[:actual_batch]
+        return x
 
 #############################################################################
 #                               Core NextDiT Model                              #
@@ -258,6 +301,7 @@ class JointTransformerBlock(nn.Module):
         x_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
         adaln_input: Optional[torch.Tensor]=None,
+        timestep_zero_index=None,
         transformer_options={},
     ):
         """
@@ -276,18 +320,18 @@ class JointTransformerBlock(nn.Module):
             assert adaln_input is not None
             scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1)
 
-            x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2(
+            x = x + apply_gate(gate_msa.unsqueeze(1).tanh(), self.attention_norm2(
                 clamp_fp16(self.attention(
-                    modulate(self.attention_norm1(x), scale_msa),
+                    modulate(self.attention_norm1(x), scale_msa, timestep_zero_index=timestep_zero_index),
                     x_mask,
                     freqs_cis,
                     transformer_options=transformer_options,
-                ))
+                ))), timestep_zero_index=timestep_zero_index
             )
-            x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
+            x = x + apply_gate(gate_mlp.unsqueeze(1).tanh(), self.ffn_norm2(
                 clamp_fp16(self.feed_forward(
-                    modulate(self.ffn_norm1(x), scale_mlp),
-                ))
+                    modulate(self.ffn_norm1(x), scale_mlp, timestep_zero_index=timestep_zero_index),
+                ))), timestep_zero_index=timestep_zero_index
             )
         else:
             assert adaln_input is None
@@ -345,13 +389,37 @@ class FinalLayer(nn.Module):
             ),
         )
 
-    def forward(self, x, c):
+    def forward(self, x, c, timestep_zero_index=None):
         scale = self.adaLN_modulation(c)
-        x = modulate(self.norm_final(x), scale)
+        x = modulate(self.norm_final(x), scale, timestep_zero_index=timestep_zero_index)
         x = self.linear(x)
         return x
 
 
+def pad_zimage(feats, pad_token, pad_tokens_multiple):
+    pad_extra = (-feats.shape[1]) % pad_tokens_multiple
+    return torch.cat((feats, pad_token.to(device=feats.device, dtype=feats.dtype, copy=True).unsqueeze(0).repeat(feats.shape[0], pad_extra, 1)), dim=1), pad_extra
+
+
+def pos_ids_x(start_t, H_tokens, W_tokens, batch_size, device, transformer_options={}):
+    rope_options = transformer_options.get("rope_options", None)
+    h_scale = 1.0
+    w_scale = 1.0
+    h_start = 0
+    w_start = 0
+    if rope_options is not None:
+        h_scale = rope_options.get("scale_y", 1.0)
+        w_scale = rope_options.get("scale_x", 1.0)
+
+        h_start = rope_options.get("shift_y", 0.0)
+        w_start = rope_options.get("shift_x", 0.0)
+    x_pos_ids = torch.zeros((batch_size, H_tokens * W_tokens, 3), dtype=torch.float32, device=device)
+    x_pos_ids[:, :, 0] = start_t
+    x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
+    x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
+    return x_pos_ids
+
+
 class NextDiT(nn.Module):
     """
     Diffusion model with a Transformer backbone.
@@ -378,6 +446,7 @@ class NextDiT(nn.Module):
         time_scale=1.0,
         pad_tokens_multiple=None,
         clip_text_dim=None,
+        siglip_feat_dim=None,
         image_model=None,
         device=None,
         dtype=None,
@@ -491,6 +560,41 @@ class NextDiT(nn.Module):
                 for layer_id in range(n_layers)
             ]
         )
+
+        if siglip_feat_dim is not None:
+            self.siglip_embedder = nn.Sequential(
+                operation_settings.get("operations").RMSNorm(siglip_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+                operation_settings.get("operations").Linear(
+                    siglip_feat_dim,
+                    dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+            self.siglip_refiner = nn.ModuleList(
+                [
+                    JointTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        modulation=False,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+            self.siglip_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype))
+        else:
+            self.siglip_embedder = None
+            self.siglip_refiner = None
+            self.siglip_pad_token = None
+
         # This norm final is in the lumina 2.0 code but isn't actually used for anything.
         # self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
         self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings)
@@ -531,70 +635,168 @@ class NextDiT(nn.Module):
             imgs = torch.stack(imgs, dim=0)
         return imgs
 
-    def patchify_and_embed(
-        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, transformer_options={}
-    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
-        bsz = len(x)
-        pH = pW = self.patch_size
-        device = x[0].device
-        orig_x = x
-
-        if self.pad_tokens_multiple is not None:
-            pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
-            cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype, copy=True).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1)
+    def embed_cap(self, cap_feats=None, offset=0, bsz=1, device=None, dtype=None):
+        if cap_feats is not None:
+            cap_feats = self.cap_embedder(cap_feats)
+            cap_feats_len = cap_feats.shape[1]
+            if self.pad_tokens_multiple is not None:
+                cap_feats, _ = pad_zimage(cap_feats, self.cap_pad_token, self.pad_tokens_multiple)
+        else:
+            cap_feats_len = 0
+            cap_feats = self.cap_pad_token.to(device=device, dtype=dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1)
 
         cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device)
-        cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0
+        cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 + offset
+        embeds = (cap_feats,)
+        freqs_cis = (self.rope_embedder(cap_pos_ids).movedim(1, 2),)
+        return embeds, freqs_cis, cap_feats_len
+
+    def embed_all(self, x, cap_feats=None, siglip_feats=None, offset=0, omni=False, transformer_options={}):
+        bsz = 1
+        pH = pW = self.patch_size
+        device = x.device
+        embeds, freqs_cis, cap_feats_len = self.embed_cap(cap_feats, offset=offset, bsz=bsz, device=device, dtype=x.dtype)
+
+        if (not omni) or self.siglip_embedder is None:
+            cap_feats_len = embeds[0].shape[1] + offset
+            embeds += (None,)
+            freqs_cis += (None,)
+        else:
+            cap_feats_len += offset
+            if siglip_feats is not None:
+                b, h, w, c = siglip_feats.shape
+                siglip_feats = siglip_feats.permute(0, 3, 1, 2).reshape(b, h * w, c)
+                siglip_feats = self.siglip_embedder(siglip_feats)
+                siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device)
+                siglip_pos_ids[:, :, 0] = cap_feats_len + 2
+                siglip_pos_ids[:, :, 1] = (torch.linspace(0, h * 8 - 1, steps=h, dtype=torch.float32, device=device).floor()).view(-1, 1).repeat(1, w).flatten()
+                siglip_pos_ids[:, :, 2] = (torch.linspace(0, w * 8 - 1, steps=w, dtype=torch.float32, device=device).floor()).view(1, -1).repeat(h, 1).flatten()
+                if self.siglip_pad_token is not None:
+                    siglip_feats, pad_extra = pad_zimage(siglip_feats, self.siglip_pad_token, self.pad_tokens_multiple)  # TODO: double check
+                    siglip_pos_ids = torch.nn.functional.pad(siglip_pos_ids, (0, 0, 0, pad_extra))
+            else:
+                if self.siglip_pad_token is not None:
+                    siglip_feats = self.siglip_pad_token.to(device=device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1)
+                    siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device)
+
+            if siglip_feats is None:
+                embeds += (None,)
+                freqs_cis += (None,)
+            else:
+                embeds += (siglip_feats,)
+                freqs_cis += (self.rope_embedder(siglip_pos_ids).movedim(1, 2),)
 
         B, C, H, W = x.shape
         x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
-
-        rope_options = transformer_options.get("rope_options", None)
-        h_scale = 1.0
-        w_scale = 1.0
-        h_start = 0
-        w_start = 0
-        if rope_options is not None:
-            h_scale = rope_options.get("scale_y", 1.0)
-            w_scale = rope_options.get("scale_x", 1.0)
-
-            h_start = rope_options.get("shift_y", 0.0)
-            w_start = rope_options.get("shift_x", 0.0)
-
-        H_tokens, W_tokens = H // pH, W // pW
-        x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device)
-        x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1
-        x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
-        x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
-
+        x_pos_ids = pos_ids_x(cap_feats_len + 1, H // pH, W // pW, bsz, device, transformer_options=transformer_options)
         if self.pad_tokens_multiple is not None:
-            pad_extra = (-x.shape[1]) % self.pad_tokens_multiple
-            x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1)
+            x, pad_extra = pad_zimage(x, self.x_pad_token, self.pad_tokens_multiple)
             x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra))
 
-        freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
+        embeds += (x,)
+        freqs_cis += (self.rope_embedder(x_pos_ids).movedim(1, 2),)
+        return embeds, freqs_cis, cap_feats_len + len(freqs_cis) - 1
+
+
+    def patchify_and_embed(
+        self, x: torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={}
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
+        bsz = x.shape[0]
+        cap_mask = None  # TODO?
+        main_siglip = None
+        orig_x = x
+
+        embeds = ([], [], [])
+        freqs_cis = ([], [], [])
+        leftover_cap = []
+
+        start_t = 0
+        omni = len(ref_latents) > 0
+        if omni:
+            for i, ref in enumerate(ref_latents):
+                if i < len(ref_contexts):
+                    ref_con = ref_contexts[i]
+                else:
+                    ref_con = None
+                if i < len(siglip_feats):
+                    sig_feat = siglip_feats[i]
+                else:
+                    sig_feat = None
+
+                out = self.embed_all(ref, ref_con, sig_feat, offset=start_t, omni=omni, transformer_options=transformer_options)
+                for i, e in enumerate(out[0]):
+                    if e is not None:
+                        embeds[i].append(comfy.utils.repeat_to_batch_size(e, bsz))
+                        freqs_cis[i].append(out[1][i])
+                start_t = out[2]
+            leftover_cap = ref_contexts[len(ref_latents):]
+
+        H, W = x.shape[-2], x.shape[-1]
+        img_sizes = [(H, W)] * bsz
+        out = self.embed_all(x, cap_feats, main_siglip, offset=start_t, omni=omni, transformer_options=transformer_options)
+        img_len = out[0][-1].shape[1]
+        cap_len = out[0][0].shape[1]
+        for i, e in enumerate(out[0]):
+            if e is not None:
+                e = comfy.utils.repeat_to_batch_size(e, bsz)
+                embeds[i].append(e)
+                freqs_cis[i].append(out[1][i])
+        start_t = out[2]
+
+        for cap in leftover_cap:
+            out = self.embed_cap(cap, offset=start_t, bsz=bsz, device=x.device, dtype=x.dtype)
+            cap_len += out[0][0].shape[1]
+            embeds[0].append(comfy.utils.repeat_to_batch_size(out[0][0], bsz))
+            freqs_cis[0].append(out[1][0])
+            start_t += out[2]
 
         patches = transformer_options.get("patches", {})
 
         # refine context
+        cap_feats = torch.cat(embeds[0], dim=1)
+        cap_freqs_cis = torch.cat(freqs_cis[0], dim=1)
         for layer in self.context_refiner:
-            cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options)
+            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options)
+
+        feats = (cap_feats,)
+        fc = (cap_freqs_cis,)
+
+        if omni and len(embeds[1]) > 0:
+            siglip_mask = None
+            siglip_feats_combined = torch.cat(embeds[1], dim=1)
+            siglip_feats_freqs_cis = torch.cat(freqs_cis[1], dim=1)
+            if self.siglip_refiner is not None:
+                for layer in self.siglip_refiner:
+                    siglip_feats_combined = layer(siglip_feats_combined, siglip_mask, siglip_feats_freqs_cis, transformer_options=transformer_options)
+            feats += (siglip_feats_combined,)
+            fc += (siglip_feats_freqs_cis,)
 
         padded_img_mask = None
+        x = torch.cat(embeds[-1], dim=1)
+        fc_x = torch.cat(freqs_cis[-1], dim=1)
+        if omni:
+            timestep_zero_index = [(x.shape[1] - img_len, x.shape[1])]
+        else:
+            timestep_zero_index = None
+
         x_input = x
         for i, layer in enumerate(self.noise_refiner):
-            x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options)
+            x = layer(x, padded_img_mask, fc_x, t, timestep_zero_index=timestep_zero_index, transformer_options=transformer_options)
             if "noise_refiner" in patches:
                 for p in patches["noise_refiner"]:
-                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"})
+                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": fc_x, "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"})
                     if "img" in out:
                         x = out["img"]
 
-        padded_full_embed = torch.cat((cap_feats, x), dim=1)
+        padded_full_embed = torch.cat(feats + (x,), dim=1)
+        if timestep_zero_index is not None:
+            ind = padded_full_embed.shape[1] - x.shape[1]
+            timestep_zero_index = [(ind + x.shape[1] - img_len, ind + x.shape[1])]
+            timestep_zero_index.append((feats[0].shape[1] - cap_len, feats[0].shape[1]))
+
         mask = None
-        img_sizes = [(H, W)] * bsz
-        l_effective_cap_len = [cap_feats.shape[1]] * bsz
-        return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
+        l_effective_cap_len = [padded_full_embed.shape[1] - img_len] * bsz
+        return padded_full_embed, mask, img_sizes, l_effective_cap_len, torch.cat(fc + (fc_x,), dim=1), timestep_zero_index
 
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -604,7 +806,11 @@ class NextDiT(nn.Module):
         ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
 
     # def forward(self, x, t, cap_feats, cap_mask):
-    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, transformer_options={}, **kwargs):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={}, **kwargs):
+        omni = len(ref_latents) > 0
+        if omni:
+            timesteps = torch.cat([timesteps * 0, timesteps], dim=0)
+
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask
@@ -619,8 +825,6 @@ class NextDiT(nn.Module):
         t = self.t_embedder(t * self.time_scale, dtype=x.dtype)  # (N, D)
         adaln_input = t
 
-        cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute
-
         if self.clip_text_pooled_proj is not None:
             pooled = kwargs.get("clip_text_pooled", None)
             if pooled is not None:
@@ -632,7 +836,7 @@ class NextDiT(nn.Module):
 
         patches = transformer_options.get("patches", {})
         x_is_tensor = isinstance(x, torch.Tensor)
-        img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options)
+        img, mask, img_size, cap_size, freqs_cis, timestep_zero_index = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, ref_latents=ref_latents, ref_contexts=ref_contexts, siglip_feats=siglip_feats, transformer_options=transformer_options)
         freqs_cis = freqs_cis.to(img.device)
 
         transformer_options["total_blocks"] = len(self.layers)
@@ -640,7 +844,7 @@ class NextDiT(nn.Module):
         img_input = img
         for i, layer in enumerate(self.layers):
             transformer_options["block_index"] = i
-            img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
+            img = layer(img, mask, freqs_cis, adaln_input, timestep_zero_index=timestep_zero_index, transformer_options=transformer_options)
             if "double_block" in patches:
                 for p in patches["double_block"]:
                     out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
@@ -649,8 +853,7 @@ class NextDiT(nn.Module):
                     if "txt" in out:
                         img[:, :cap_size[0]] = out["txt"]
 
-        img = self.final_layer(img, adaln_input)
+        img = self.final_layer(img, adaln_input, timestep_zero_index=timestep_zero_index)
         img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w]
-
         return -img
 

comfy/model_base.py

@@ -1150,6 +1150,7 @@ class CosmosPredict2(BaseModel):
 class Lumina2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1169,6 +1170,35 @@ class Lumina2(BaseModel):
         if clip_text_pooled is not None:
             out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled)
 
+        clip_vision_outputs = kwargs.get("clip_vision_outputs", list(map(lambda a: a.get("clip_vision_output"), kwargs.get("unclip_conditioning", [{}]))))  # Z Image omni
+        if clip_vision_outputs is not None and len(clip_vision_outputs) > 0:
+            sigfeats = []
+            for clip_vision_output in clip_vision_outputs:
+                if clip_vision_output is not None:
+                    image_size = clip_vision_output.image_sizes[0]
+                    shape = clip_vision_output.last_hidden_state.shape
+                    sigfeats.append(clip_vision_output.last_hidden_state.reshape(shape[0], image_size[1] // 16, image_size[2] // 16, shape[-1]))
+            if len(sigfeats) > 0:
+                out['siglip_feats'] = comfy.conds.CONDList(sigfeats)
+
+        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_contexts = kwargs.get("reference_latents_text_embeds", None)
+        if ref_contexts is not None:
+            out['ref_contexts'] = comfy.conds.CONDList(ref_contexts)
+
+        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 WAN21(BaseModel):

comfy/model_detection.py

@@ -253,7 +253,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["image_model"] = "chroma_radiance"
                 dit_config["in_channels"] = 3
                 dit_config["out_channels"] = 3
-                dit_config["patch_size"] = 16
+                dit_config["patch_size"] = state_dict.get('{}img_in_patch.weight'.format(key_prefix)).size(dim=-1)
                 dit_config["nerf_hidden_size"] = 64
                 dit_config["nerf_mlp_ratio"] = 4
                 dit_config["nerf_depth"] = 4
@@ -446,6 +446,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["time_scale"] = 1000.0
             if '{}cap_pad_token'.format(key_prefix) in state_dict_keys:
                 dit_config["pad_tokens_multiple"] = 32
+            sig_weight = state_dict.get('{}siglip_embedder.0.weight'.format(key_prefix), None)
+            if sig_weight is not None:
+                dit_config["siglip_feat_dim"] = sig_weight.shape[0]
 
         return dit_config
 

comfy/text_encoders/ovis.py

@@ -61,6 +61,7 @@ def te(dtype_llama=None, llama_quantization_metadata=None):
             if dtype_llama is not None:
                 dtype = dtype_llama
             if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
                 model_options["quantization_metadata"] = llama_quantization_metadata
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return OvisTEModel_

comfy/text_encoders/z_image.py

@@ -40,6 +40,7 @@ def te(dtype_llama=None, llama_quantization_metadata=None):
             if dtype_llama is not None:
                 dtype = dtype_llama
             if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
                 model_options["quantization_metadata"] = llama_quantization_metadata
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return ZImageTEModel_

comfy/utils.py

@@ -639,6 +639,8 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
                         "proj_out.bias": "linear2.bias",
                         "attn.norm_q.weight": "norm.query_norm.scale",
                         "attn.norm_k.weight": "norm.key_norm.scale",
+                        "attn.to_qkv_mlp_proj.weight": "linear1.weight", # Flux 2
+                        "attn.to_out.weight": "linear2.weight", # Flux 2
                     }
 
         for k in block_map:

comfy_api/latest/_io.py

@@ -1000,20 +1000,38 @@ class Autogrow(ComfyTypeI):
             names = [f"{prefix}{i}" for i in range(max)]
         # need to create a new input based on the contents of input
         template_input = None
-        for _, dict_input in input.items():
-            # for now, get just the first value from dict_input
+        template_required = True
+        for _input_type, dict_input in input.items():
+            # for now, get just the first value from dict_input; if not required, min can be ignored
+            if len(dict_input) == 0:
+                continue
             template_input = list(dict_input.values())[0]
+            template_required = _input_type == "required"
+            break
+        if template_input is None:
+            raise Exception("template_input could not be determined from required or optional; this should never happen.")
         new_dict = {}
+        new_dict_added_to = False
+        # first, add possible inputs into out_dict
         for i, name in enumerate(names):
             expected_id = finalize_prefix(curr_prefix, name)
+            # required
+            if i < min and template_required:
+                out_dict["required"][expected_id] = template_input
+                type_dict = new_dict.setdefault("required", {})
+            # optional
+            else:
+                out_dict["optional"][expected_id] = template_input
+                type_dict = new_dict.setdefault("optional", {})
             if expected_id in live_inputs:
-                # required
-                if i < min:
-                    type_dict = new_dict.setdefault("required", {})
-                # optional
-                else:
-                    type_dict = new_dict.setdefault("optional", {})
+                # NOTE: prefix gets added in parse_class_inputs
                 type_dict[name] = template_input
+                new_dict_added_to = True
+        # account for the edge case that all inputs are optional and no values are received
+        if not new_dict_added_to:
+            finalized_prefix = finalize_prefix(curr_prefix)
+            out_dict["dynamic_paths"][finalized_prefix] = finalized_prefix
+            out_dict["dynamic_paths_default_value"][finalized_prefix] = DynamicPathsDefaultValue.EMPTY_DICT
         parse_class_inputs(out_dict, live_inputs, new_dict, curr_prefix)
 
 @comfytype(io_type="COMFY_DYNAMICCOMBO_V3")
@@ -1151,6 +1169,8 @@ class V3Data(TypedDict):
     'Dictionary where the keys are the hidden input ids and the values are the values of the hidden inputs.'
     dynamic_paths: dict[str, Any]
     'Dictionary where the keys are the input ids and the values dictate how to turn the inputs into a nested dictionary.'
+    dynamic_paths_default_value: dict[str, Any]
+    'Dictionary where the keys are the input ids and the values are a string from DynamicPathsDefaultValue for the inputs if value is None.'
     create_dynamic_tuple: bool
     'When True, the value of the dynamic input will be in the format (value, path_key).'
 
@@ -1504,6 +1524,7 @@ def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], i
         "required": {},
         "optional": {},
         "dynamic_paths": {},
+        "dynamic_paths_default_value": {},
     }
     d = d.copy()
     # ignore hidden for parsing
@@ -1513,8 +1534,12 @@ def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], i
         out_dict["hidden"] = hidden
     v3_data = {}
     dynamic_paths = out_dict.pop("dynamic_paths", None)
-    if dynamic_paths is not None:
+    if dynamic_paths is not None and len(dynamic_paths) > 0:
         v3_data["dynamic_paths"] = dynamic_paths
+    # this list is used for autogrow, in the case all inputs are optional and no values are passed
+    dynamic_paths_default_value = out_dict.pop("dynamic_paths_default_value", None)
+    if dynamic_paths_default_value is not None and len(dynamic_paths_default_value) > 0:
+        v3_data["dynamic_paths_default_value"] = dynamic_paths_default_value
     return out_dict, hidden, v3_data
 
 def parse_class_inputs(out_dict: dict[str, Any], live_inputs: dict[str, Any], curr_dict: dict[str, Any], curr_prefix: list[str] | None=None) -> None:
@@ -1551,11 +1576,16 @@ def add_to_dict_v1(i: Input, d: dict):
 def add_to_dict_v3(io: Input | Output, d: dict):
     d[io.id] = (io.get_io_type(), io.as_dict())
 
+class DynamicPathsDefaultValue:
+    EMPTY_DICT = "empty_dict"
+
 def build_nested_inputs(values: dict[str, Any], v3_data: V3Data):
     paths = v3_data.get("dynamic_paths", None)
+    default_value_dict = v3_data.get("dynamic_paths_default_value", {})
     if paths is None:
         return values
     values = values.copy()
+
     result = {}
 
     create_tuple = v3_data.get("create_dynamic_tuple", False)
@@ -1569,6 +1599,11 @@ def build_nested_inputs(values: dict[str, Any], v3_data: V3Data):
 
             if is_last:
                 value = values.pop(key, None)
+                if value is None:
+                    # see if a default value was provided for this key
+                    default_option = default_value_dict.get(key, None)
+                    if default_option == DynamicPathsDefaultValue.EMPTY_DICT:
+                        value = {}
                 if create_tuple:
                     value = (value, key)
                 current[p] = value

comfy_api_nodes/apis/bria.py

@@ -0,0 +1,61 @@
+from typing import TypedDict
+
+from pydantic import BaseModel, Field
+
+
+class InputModerationSettings(TypedDict):
+    prompt_content_moderation: bool
+    visual_input_moderation: bool
+    visual_output_moderation: bool
+
+
+class BriaEditImageRequest(BaseModel):
+    instruction: str | None = Field(...)
+    structured_instruction: str | None = Field(
+        ...,
+        description="Use this instead of instruction for precise, programmatic control.",
+    )
+    images: list[str] = Field(
+        ...,
+        description="Required. Publicly available URL or Base64-encoded. Must contain exactly one item.",
+    )
+    mask: str | None = Field(
+        None,
+        description="Mask image (black and white). Black areas will be preserved, white areas will be edited. "
+        "If omitted, the edit applies to the entire image. "
+        "The input image and the the input mask must be of the same size.",
+    )
+    negative_prompt: str | None = Field(None)
+    guidance_scale: float = Field(...)
+    model_version: str = Field(...)
+    steps_num: int = Field(...)
+    seed: int = Field(...)
+    ip_signal: bool = Field(
+        False,
+        description="If true, returns a warning for potential IP content in the instruction.",
+    )
+    prompt_content_moderation: bool = Field(
+        False, description="If true, returns 422 on instruction moderation failure."
+    )
+    visual_input_content_moderation: bool = Field(
+        False, description="If true, returns 422 on images or mask moderation failure."
+    )
+    visual_output_content_moderation: bool = Field(
+        False, description="If true, returns 422 on visual output moderation failure."
+    )
+
+
+class BriaStatusResponse(BaseModel):
+    request_id: str = Field(...)
+    status_url: str = Field(...)
+    warning: str | None = Field(None)
+
+
+class BriaResult(BaseModel):
+    structured_prompt: str = Field(...)
+    image_url: str = Field(...)
+
+
+class BriaResponse(BaseModel):
+    status: str = Field(...)
+    result: BriaResult | None = Field(None)

comfy_api_nodes/apis/wavespeed.py

@@ -0,0 +1,35 @@
+from pydantic import BaseModel, Field
+
+
+class SeedVR2ImageRequest(BaseModel):
+    image: str = Field(...)
+    target_resolution: str = Field(...)
+    output_format: str = Field("png")
+    enable_sync_mode: bool = Field(False)
+
+
+class FlashVSRRequest(BaseModel):
+    target_resolution: str = Field(...)
+    video: str = Field(...)
+    duration: float = Field(...)
+
+
+class TaskCreatedDataResponse(BaseModel):
+    id: str = Field(...)
+
+
+class TaskCreatedResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskCreatedDataResponse | None = Field(None)
+
+
+class TaskResultDataResponse(BaseModel):
+    status: str = Field(...)
+    outputs: list[str] = Field([])
+
+
+class TaskResultResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskResultDataResponse | None = Field(None)

comfy_api_nodes/nodes_bria.py

@@ -0,0 +1,198 @@
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.bria import (
+    BriaEditImageRequest,
+    BriaResponse,
+    BriaStatusResponse,
+    InputModerationSettings,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    convert_mask_to_image,
+    download_url_to_image_tensor,
+    get_number_of_images,
+    poll_op,
+    sync_op,
+    upload_images_to_comfyapi,
+)
+
+
+class BriaImageEditNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="BriaImageEditNode",
+            display_name="Bria Image Edit",
+            category="api node/image/Bria",
+            description="Edit images using Bria latest model",
+            inputs=[
+                IO.Combo.Input("model", options=["FIBO"]),
+                IO.Image.Input("image"),
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Instruction to edit image",
+                ),
+                IO.String.Input("negative_prompt", multiline=True, default=""),
+                IO.String.Input(
+                    "structured_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="A string containing the structured edit prompt in JSON format. "
+                    "Use this instead of usual prompt for precise, programmatic control.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=1,
+                    min=1,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                ),
+                IO.Float.Input(
+                    "guidance_scale",
+                    default=3,
+                    min=3,
+                    max=5,
+                    step=0.01,
+                    display_mode=IO.NumberDisplay.number,
+                    tooltip="Higher value makes the image follow the prompt more closely.",
+                ),
+                IO.Int.Input(
+                    "steps",
+                    default=50,
+                    min=20,
+                    max=50,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                ),
+                IO.DynamicCombo.Input(
+                    "moderation",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "true",
+                            [
+                                IO.Boolean.Input(
+                                    "prompt_content_moderation", default=False
+                                ),
+                                IO.Boolean.Input(
+                                    "visual_input_moderation", default=False
+                                ),
+                                IO.Boolean.Input(
+                                    "visual_output_moderation", default=True
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option("false", []),
+                    ],
+                    tooltip="Moderation settings",
+                ),
+                IO.Mask.Input(
+                    "mask",
+                    tooltip="If omitted, the edit applies to the entire image.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                IO.Image.Output(),
+                IO.String.Output(display_name="structured_prompt"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.04}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: Input.Image,
+        prompt: str,
+        negative_prompt: str,
+        structured_prompt: str,
+        seed: int,
+        guidance_scale: float,
+        steps: int,
+        moderation: InputModerationSettings,
+        mask: Input.Image | None = None,
+    ) -> IO.NodeOutput:
+        if not prompt and not structured_prompt:
+            raise ValueError(
+                "One of prompt or structured_prompt is required to be non-empty."
+            )
+        if get_number_of_images(image) != 1:
+            raise ValueError("Exactly one input image is required.")
+        mask_url = None
+        if mask is not None:
+            mask_url = (
+                await upload_images_to_comfyapi(
+                    cls,
+                    convert_mask_to_image(mask),
+                    max_images=1,
+                    mime_type="image/png",
+                    wait_label="Uploading mask",
+                )
+            )[0]
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="proxy/bria/v2/image/edit", method="POST"),
+            data=BriaEditImageRequest(
+                instruction=prompt if prompt else None,
+                structured_instruction=structured_prompt if structured_prompt else None,
+                images=await upload_images_to_comfyapi(
+                    cls,
+                    image,
+                    max_images=1,
+                    mime_type="image/png",
+                    wait_label="Uploading image",
+                ),
+                mask=mask_url,
+                negative_prompt=negative_prompt if negative_prompt else None,
+                guidance_scale=guidance_scale,
+                seed=seed,
+                model_version=model,
+                steps_num=steps,
+                prompt_content_moderation=moderation.get(
+                    "prompt_content_moderation", False
+                ),
+                visual_input_content_moderation=moderation.get(
+                    "visual_input_moderation", False
+                ),
+                visual_output_content_moderation=moderation.get(
+                    "visual_output_moderation", False
+                ),
+            ),
+            response_model=BriaStatusResponse,
+        )
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"),
+            status_extractor=lambda r: r.status,
+            response_model=BriaResponse,
+        )
+        return IO.NodeOutput(
+            await download_url_to_image_tensor(response.result.image_url),
+            response.result.structured_prompt,
+        )
+
+
+class BriaExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            BriaImageEditNode,
+        ]
+
+
+async def comfy_entrypoint() -> BriaExtension:
+    return BriaExtension()

comfy_api_nodes/nodes_wavespeed.py

@@ -0,0 +1,178 @@
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.wavespeed import (
+    FlashVSRRequest,
+    TaskCreatedResponse,
+    TaskResultResponse,
+    SeedVR2ImageRequest,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    download_url_to_video_output,
+    poll_op,
+    sync_op,
+    upload_video_to_comfyapi,
+    validate_container_format_is_mp4,
+    validate_video_duration,
+    upload_images_to_comfyapi,
+    get_number_of_images,
+    download_url_to_image_tensor,
+)
+
+
+class WavespeedFlashVSRNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="WavespeedFlashVSRNode",
+            display_name="FlashVSR Video Upscale",
+            category="api node/video/WaveSpeed",
+            description="Fast, high-quality video upscaler that "
+            "boosts resolution and restores clarity for low-resolution or blurry footage.",
+            inputs=[
+                IO.Video.Input("video"),
+                IO.Combo.Input("target_resolution", options=["720p", "1080p", "2K", "4K"]),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["target_resolution"]),
+                expr="""
+                (
+                  $price_for_1sec := {"720p": 0.012, "1080p": 0.018, "2k": 0.024, "4k": 0.032};
+                  {
+                    "type":"usd",
+                    "usd": $lookup($price_for_1sec, widgets.target_resolution),
+                    "format":{"suffix": "/second", "approximate": true}
+                  }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        target_resolution: str,
+    ) -> IO.NodeOutput:
+        validate_container_format_is_mp4(video)
+        validate_video_duration(video, min_duration=5, max_duration=60 * 10)
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/wavespeed/api/v3/wavespeed-ai/flashvsr", method="POST"),
+            response_model=TaskCreatedResponse,
+            data=FlashVSRRequest(
+                target_resolution=target_resolution.lower(),
+                video=await upload_video_to_comfyapi(cls, video),
+                duration=video.get_duration(),
+            ),
+        )
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation fails with code={initial_res.code} and message={initial_res.message}")
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wavespeed/api/v3/predictions/{initial_res.data.id}/result"),
+            response_model=TaskResultResponse,
+            status_extractor=lambda x: "failed" if x.data is None else x.data.status,
+            poll_interval=10.0,
+            max_poll_attempts=480,
+        )
+        if final_response.code != 200:
+            raise ValueError(
+                f"Task processing failed with code={final_response.code} and message={final_response.message}"
+            )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.data.outputs[0]))
+
+
+class WavespeedImageUpscaleNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="WavespeedImageUpscaleNode",
+            display_name="WaveSpeed Image Upscale",
+            category="api node/image/WaveSpeed",
+            description="Boost image resolution and quality, upscaling photos to 4K or 8K for sharp, detailed results.",
+            inputs=[
+                IO.Combo.Input("model", options=["SeedVR2", "Ultimate"]),
+                IO.Image.Input("image"),
+                IO.Combo.Input("target_resolution", options=["2K", "4K", "8K"]),
+            ],
+            outputs=[
+                IO.Image.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model"]),
+                expr="""
+                (
+                  $prices := {"seedvr2": 0.01, "ultimate": 0.06};
+                  {"type":"usd", "usd": $lookup($prices, widgets.model)}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: Input.Image,
+        target_resolution: str,
+    ) -> IO.NodeOutput:
+        if get_number_of_images(image) != 1:
+            raise ValueError("Exactly one input image is required.")
+        if model == "SeedVR2":
+            model_path = "seedvr2/image"
+        else:
+            model_path = "ultimate-image-upscaler"
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wavespeed/api/v3/wavespeed-ai/{model_path}", method="POST"),
+            response_model=TaskCreatedResponse,
+            data=SeedVR2ImageRequest(
+                target_resolution=target_resolution.lower(),
+                image=(await upload_images_to_comfyapi(cls, image, max_images=1))[0],
+            ),
+        )
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation fails with code={initial_res.code} and message={initial_res.message}")
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/wavespeed/api/v3/predictions/{initial_res.data.id}/result"),
+            response_model=TaskResultResponse,
+            status_extractor=lambda x: "failed" if x.data is None else x.data.status,
+            poll_interval=10.0,
+            max_poll_attempts=480,
+        )
+        if final_response.code != 200:
+            raise ValueError(
+                f"Task processing failed with code={final_response.code} and message={final_response.message}"
+            )
+        return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.outputs[0]))
+
+
+class WavespeedExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            WavespeedFlashVSRNode,
+            WavespeedImageUpscaleNode,
+        ]
+
+
+async def comfy_entrypoint() -> WavespeedExtension:
+    return WavespeedExtension()

comfy_api_nodes/util/__init__.py

@@ -11,6 +11,7 @@ from .conversions import (
     audio_input_to_mp3,
     audio_to_base64_string,
     bytesio_to_image_tensor,
+    convert_mask_to_image,
     downscale_image_tensor,
     image_tensor_pair_to_batch,
     pil_to_bytesio,
@@ -72,6 +73,7 @@ __all__ = [
     "audio_input_to_mp3",
     "audio_to_base64_string",
     "bytesio_to_image_tensor",
+    "convert_mask_to_image",
     "downscale_image_tensor",
     "image_tensor_pair_to_batch",
     "pil_to_bytesio",

comfy_api_nodes/util/conversions.py

@@ -451,6 +451,12 @@ def resize_mask_to_image(
     return mask
 
 
+def convert_mask_to_image(mask: Input.Image) -> torch.Tensor:
+    """Make mask have the expected amount of dims (4) and channels (3) to be recognized as an image."""
+    mask = mask.unsqueeze(-1)
+    return torch.cat([mask] * 3, dim=-1)
+
+
 def text_filepath_to_base64_string(filepath: str) -> str:
     """Converts a text file to a base64 string."""
     with open(filepath, "rb") as f:

comfy_extras/nodes_zimage.py

@@ -0,0 +1,88 @@
+import node_helpers
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io
+import math
+import comfy.utils
+
+
+class TextEncodeZImageOmni(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncodeZImageOmni",
+            category="advanced/conditioning",
+            is_experimental=True,
+            inputs=[
+                io.Clip.Input("clip"),
+                io.ClipVision.Input("image_encoder", optional=True),
+                io.String.Input("prompt", multiline=True, dynamic_prompts=True),
+                io.Boolean.Input("auto_resize_images", default=True),
+                io.Vae.Input("vae", optional=True),
+                io.Image.Input("image1", optional=True),
+                io.Image.Input("image2", optional=True),
+                io.Image.Input("image3", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, clip, prompt, image_encoder=None, auto_resize_images=True, vae=None, image1=None, image2=None, image3=None) -> io.NodeOutput:
+        ref_latents = []
+        images = list(filter(lambda a: a is not None, [image1, image2, image3]))
+
+        prompt_list = []
+        template = None
+        if len(images) > 0:
+            prompt_list = ["<|im_start|>user\n<|vision_start|>"]
+            prompt_list += ["<|vision_end|><|vision_start|>"] * (len(images) - 1)
+            prompt_list += ["<|vision_end|><|im_end|>"]
+            template = "<|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n<|vision_start|>"
+
+        encoded_images = []
+
+        for i, image in enumerate(images):
+            if image_encoder is not None:
+                encoded_images.append(image_encoder.encode_image(image))
+
+            if vae is not None:
+                if auto_resize_images:
+                    samples = image.movedim(-1, 1)
+                    total = int(1024 * 1024)
+                    scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+                    width = round(samples.shape[3] * scale_by / 8.0) * 8
+                    height = round(samples.shape[2] * scale_by / 8.0) * 8
+
+                    image = comfy.utils.common_upscale(samples, width, height, "area", "disabled").movedim(1, -1)
+                ref_latents.append(vae.encode(image))
+
+        tokens = clip.tokenize(prompt, llama_template=template)
+        conditioning = clip.encode_from_tokens_scheduled(tokens)
+
+        extra_text_embeds = []
+        for p in prompt_list:
+            tokens = clip.tokenize(p, llama_template="{}")
+            text_embeds = clip.encode_from_tokens_scheduled(tokens)
+            extra_text_embeds.append(text_embeds[0][0])
+
+        if len(ref_latents) > 0:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_latents": ref_latents}, append=True)
+        if len(encoded_images) > 0:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"clip_vision_outputs": encoded_images}, append=True)
+        if len(extra_text_embeds) > 0:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_latents_text_embeds": extra_text_embeds}, append=True)
+
+        return io.NodeOutput(conditioning)
+
+
+class ZImageExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            TextEncodeZImageOmni,
+        ]
+
+
+async def comfy_entrypoint() -> ZImageExtension:
+    return ZImageExtension()

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.9.2"
+__version__ = "0.10.0"

folder_paths.py

@@ -105,6 +105,14 @@ if not os.path.exists(input_directory):
     except:
         logging.error("Failed to create input directory")
 
+custom_nodes_paths, _ = folder_names_and_paths["custom_nodes"]
+for directory in custom_nodes_paths:
+    if not os.path.exists(directory):
+        try:
+            os.makedirs(directory)
+        except:
+            logging.error(f"Failed to create custom_nodes directory: {directory}")
+
 def set_output_directory(output_dir: str) -> None:
     global output_directory
     output_directory = output_dir

nodes.py

@@ -2373,6 +2373,7 @@ async def init_builtin_extra_nodes():
         "nodes_kandinsky5.py",
         "nodes_wanmove.py",
         "nodes_image_compare.py",
+        "nodes_zimage.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.9.2"
+version = "0.10.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.36.14
-comfyui-workflow-templates==0.8.11
+comfyui-frontend-package==1.37.11
+comfyui-workflow-templates==0.8.15
 comfyui-embedded-docs==0.4.0
 torch
 torchsde