Fibo Edit: multi-reference conditioning and batching

docs/source/en/api/pipelines/bria_fibo_edit.md

@@ -16,18 +16,20 @@ Fibo Edit is an 8B parameter image-to-image model that introduces a new paradigm
 Featuring native masking for granular precision, it moves beyond simple prompt-based diffusion to offer explicit, interpretable control optimized for production environments.
 Its lightweight architecture is designed for deep customization, empowering researchers to build specialized "Edit" models for domain-specific tasks while delivering top-tier aesthetic quality
 
+Refer to the Bria Fibo Edit Hugging Face [page](https://huggingface.co/briaai/Fibo-Edit-1.5-base) to learn more. A distilled checkpoint is available at [Fibo-Edit-1.5-turbo](https://huggingface.co/briaai/Fibo-Edit-1.5-turbo).
+
 ## Usage
-_As the model is gated, before using it with diffusers you first need to go to the [Bria Fibo Hugging Face page](https://huggingface.co/briaai/Fibo-Edit), fill in the form and accept the gate. Once you are in, you need to login so that your system knows you’ve accepted the gate._
+
+_As the model is gated, before using it with diffusers you first need to go to the [Bria Fibo Edit Hugging Face page](https://huggingface.co/briaai/Fibo-Edit-1.5-base), fill in the form and accept the gate. Once you are in, you need to login so that your system knows you’ve accepted the gate._
 
 Use the command below to log in:
 
 ```bash
 hf auth login
 ```
 
-
 ## BriaFiboEditPipeline
 
 [[autodoc]] BriaFiboEditPipeline
 	- all
-	- __call__
+	- __call__

src/diffusers/models/transformers/transformer_bria_fibo.py

@@ -441,6 +441,7 @@ class BriaFiboTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, From
     """
 
     _supports_gradient_checkpointing = True
+    _repeated_blocks = ["BriaFiboTransformerBlock", "BriaFiboSingleTransformerBlock"]
 
     @register_to_config
     def __init__(

src/diffusers/pipelines/bria_fibo/pipeline_bria_fibo.py

@@ -432,16 +432,6 @@ def prepare_latents(
 
         return latents, latent_image_ids
 
-    @staticmethod
-    def _prepare_attention_mask(attention_mask):
-        attention_matrix = torch.einsum("bi,bj->bij", attention_mask, attention_mask)
-
-        # convert to 0 - keep, -inf ignore
-        attention_matrix = torch.where(
-            attention_matrix == 1, 0.0, -torch.inf
-        )  # Apply -inf to ignored tokens for nulling softmax score
-        return attention_matrix
-
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -615,13 +605,15 @@ def __call__(
         if guidance_scale > 1:
             latent_attention_mask = latent_attention_mask.repeat(2, 1)
 
-        attention_mask = torch.cat([prompt_attention_mask, latent_attention_mask], dim=1)
-        attention_mask = self._prepare_attention_mask(attention_mask)  # batch, seq => batch, seq, seq
-        attention_mask = attention_mask.unsqueeze(dim=1).to(dtype=self.transformer.dtype)  # for head broadcasting
+        attention_mask = torch.cat([prompt_attention_mask, latent_attention_mask], dim=1).bool()
 
         if self._joint_attention_kwargs is None:
             self._joint_attention_kwargs = {}
-        self._joint_attention_kwargs["attention_mask"] = attention_mask
+        if not attention_mask.all():
+            # Bool key-padding mask (batch, 1, 1, seq): every real query attends to the same keys as
+            # with a full (seq, seq) matrix, and varlen backends require bool. When nothing is padded
+            # the mask is a no-op, and omitting it keeps backends without mask support usable.
+            self._joint_attention_kwargs["attention_mask"] = attention_mask[:, None, None, :]
 
         # Adapt scheduler to dynamic shifting (resolution dependent)
 
@@ -630,7 +622,7 @@ def __call__(
         else:
             seq_len = (height // self.vae_scale_factor) * (width // self.vae_scale_factor)
 
-        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        sigmas = None if timesteps is not None else np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
 
         mu = calculate_shift(
             seq_len,
@@ -646,7 +638,7 @@ def __call__(
             self.scheduler,
             num_inference_steps=num_inference_steps,
             device=device,
-            timesteps=None,
+            timesteps=timesteps,
             sigmas=sigmas,
             mu=mu,
         )