Fix SDXL area composition sometimes not using the right pooled output.

Author: comfyanonymous

comfy/model_base.py

@@ -126,9 +126,15 @@ def blank_inpaint_image_like(latent_image):
                         cond_concat.append(blank_inpaint_image_like(noise))
             data = torch.cat(cond_concat, dim=1)
             out['c_concat'] = comfy.conds.CONDNoiseShape(data)
+
         adm = self.encode_adm(**kwargs)
         if adm is not None:
             out['y'] = comfy.conds.CONDRegular(adm)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+
         return out
 
     def load_model_weights(self, sd, unet_prefix=""):
@@ -322,6 +328,10 @@ def extra_conds(self, **kwargs):
 
         out['c_concat'] = comfy.conds.CONDNoiseShape(latent_image)
 
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+
         if "time_conditioning" in kwargs:
             out["time_context"] = comfy.conds.CONDCrossAttn(kwargs["time_conditioning"])
 

comfy/samplers.py

@@ -599,6 +599,10 @@ def sample(model, noise, positive, negative, cfg, device, sampler, sigmas, model
     calculate_start_end_timesteps(model, negative)
     calculate_start_end_timesteps(model, positive)
 
+    if hasattr(model, 'extra_conds'):
+        positive = encode_model_conds(model.extra_conds, positive, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask)
+        negative = encode_model_conds(model.extra_conds, negative, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask)
+
     #make sure each cond area has an opposite one with the same area
     for c in positive:
         create_cond_with_same_area_if_none(negative, c)
@@ -613,9 +617,6 @@ def sample(model, noise, positive, negative, cfg, device, sampler, sigmas, model
     if latent_image is not None:
         latent_image = model.process_latent_in(latent_image)
 
-    if hasattr(model, 'extra_conds'):
-        positive = encode_model_conds(model.extra_conds, positive, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask)
-        negative = encode_model_conds(model.extra_conds, negative, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask)
 
     extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": model_options, "seed":seed}