Add working Qwen 2512 ControlNet (Fun ControlNet) support (#12359)

2026-02-14 08:53:52 +05:30
parent 712efb466b
commit dc9822b7df
2 changed files with 263 additions and 0 deletions
@@ -2,6 +2,196 @@ import torch
 import math

 from .model import QwenImageTransformer2DModel
+from .model import QwenImageTransformerBlock
+
+
+class QwenImageFunControlBlock(QwenImageTransformerBlock):
+    def __init__(self, dim, num_attention_heads, attention_head_dim, has_before_proj=False, dtype=None, device=None, operations=None):
+        super().__init__(
+            dim=dim,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=attention_head_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.has_before_proj = has_before_proj
+        if has_before_proj:
+            self.before_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.after_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+
+
+class QwenImageFunControlNetModel(torch.nn.Module):
+    def __init__(
+        self,
+        control_in_features=132,
+        inner_dim=3072,
+        num_attention_heads=24,
+        attention_head_dim=128,
+        num_control_blocks=5,
+        main_model_double=60,
+        injection_layers=(0, 12, 24, 36, 48),
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.main_model_double = main_model_double
+        self.injection_layers = tuple(injection_layers)
+        # Keep base hint scaling at 1.0 so user-facing strength behaves similarly
+        # to the reference Gen2/VideoX implementation around strength=1.
+        self.hint_scale = 1.0
+        self.control_img_in = operations.Linear(control_in_features, inner_dim, device=device, dtype=dtype)
+
+        self.control_blocks = torch.nn.ModuleList([])
+        for i in range(num_control_blocks):
+            self.control_blocks.append(
+                QwenImageFunControlBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    has_before_proj=(i == 0),
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+            )
+
+    def _process_hint_tokens(self, hint):
+        if hint is None:
+            return None
+        if hint.ndim == 4:
+            hint = hint.unsqueeze(2)
+
+        # Fun checkpoints are trained with 33 latent channels before 2x2 packing:
+        # [control_latent(16), mask(1), inpaint_latent(16)] -> 132 features.
+        # Default behavior (no inpaint input in stock Apply ControlNet) should use
+        # zeros for mask/inpaint branches, matching VideoX fallback semantics.
+        expected_c = self.control_img_in.weight.shape[1] // 4
+        if hint.shape[1] == 16 and expected_c == 33:
+            zeros_mask = torch.zeros_like(hint[:, :1])
+            zeros_inpaint = torch.zeros_like(hint)
+            hint = torch.cat([hint, zeros_mask, zeros_inpaint], dim=1)
+
+        bs, c, t, h, w = hint.shape
+        hidden_states = torch.nn.functional.pad(hint, (0, w % 2, 0, h % 2))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(
+            orig_shape[0],
+            orig_shape[1],
+            orig_shape[-3],
+            orig_shape[-2] // 2,
+            2,
+            orig_shape[-1] // 2,
+            2,
+        )
+        hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6)
+        hidden_states = hidden_states.reshape(
+            bs,
+            t * ((h + 1) // 2) * ((w + 1) // 2),
+            c * 4,
+        )
+
+        expected_in = self.control_img_in.weight.shape[1]
+        cur_in = hidden_states.shape[-1]
+        if cur_in < expected_in:
+            pad = torch.zeros(
+                (hidden_states.shape[0], hidden_states.shape[1], expected_in - cur_in),
+                device=hidden_states.device,
+                dtype=hidden_states.dtype,
+            )
+            hidden_states = torch.cat([hidden_states, pad], dim=-1)
+        elif cur_in > expected_in:
+            hidden_states = hidden_states[:, :, :expected_in]
+
+        return hidden_states
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        context,
+        attention_mask=None,
+        guidance: torch.Tensor = None,
+        hint=None,
+        transformer_options={},
+        base_model=None,
+        **kwargs,
+    ):
+        if base_model is None:
+            raise RuntimeError("Qwen Fun ControlNet requires a QwenImage base model at runtime.")
+
+        encoder_hidden_states_mask = attention_mask
+        # Keep attention mask disabled inside Fun control blocks to mirror
+        # VideoX behavior (they rely on seq lengths for RoPE, not masked attention).
+        encoder_hidden_states_mask = None
+
+        hidden_states, img_ids, _ = base_model.process_img(x)
+        hint_tokens = self._process_hint_tokens(hint)
+        if hint_tokens is None:
+            raise RuntimeError("Qwen Fun ControlNet requires a control hint image.")
+
+        if hint_tokens.shape[1] != hidden_states.shape[1]:
+            max_tokens = min(hint_tokens.shape[1], hidden_states.shape[1])
+            hint_tokens = hint_tokens[:, :max_tokens]
+            hidden_states = hidden_states[:, :max_tokens]
+            img_ids = img_ids[:, :max_tokens]
+
+        txt_start = round(
+            max(
+                ((x.shape[-1] + (base_model.patch_size // 2)) // base_model.patch_size) // 2,
+                ((x.shape[-2] + (base_model.patch_size // 2)) // base_model.patch_size) // 2,
+            )
+        )
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = base_model.pe_embedder(ids).to(x.dtype).contiguous()
+
+        hidden_states = base_model.img_in(hidden_states)
+        encoder_hidden_states = base_model.txt_norm(context)
+        encoder_hidden_states = base_model.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            base_model.time_text_embed(timesteps, hidden_states)
+            if guidance is None
+            else base_model.time_text_embed(timesteps, guidance, hidden_states)
+        )
+
+        c = self.control_img_in(hint_tokens)
+
+        for i, block in enumerate(self.control_blocks):
+            if i == 0:
+                c_in = block.before_proj(c) + hidden_states
+                all_c = []
+            else:
+                all_c = list(torch.unbind(c, dim=0))
+                c_in = all_c.pop(-1)
+
+            encoder_hidden_states, c_out = block(
+                hidden_states=c_in,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_hidden_states_mask=encoder_hidden_states_mask,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+                transformer_options=transformer_options,
+            )
+
+            c_skip = block.after_proj(c_out) * self.hint_scale
+            all_c += [c_skip, c_out]
+            c = torch.stack(all_c, dim=0)
+
+        hints = torch.unbind(c, dim=0)[:-1]
+
+        controlnet_block_samples = [None] * self.main_model_double
+        for local_idx, base_idx in enumerate(self.injection_layers):
+            if local_idx < len(hints) and base_idx < len(controlnet_block_samples):
+                controlnet_block_samples[base_idx] = hints[local_idx]
+
+        return {"input": controlnet_block_samples}


 class QwenImageControlNetModel(QwenImageTransformer2DModel):