Bria fibo #12545

galbria · 2025-10-26T17:11:02Z

What does this PR do?

Fixes # (issue)

Before submitting

This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
Did you read the contributor guideline?
Did you read our philosophy doc (important for complex PRs)?
Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
Did you make sure to update the documentation with your changes? Here are the
documentation guidelines, and
here are tips on formatting docstrings.
Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

sayakpaul

Thanks a lot for the PR. Excited for FIBO to make strides!

I have left a bunch of comments, most of which should be easily resolvable. If not, please let me know.

Additionally, I think:

It'd be nice to include a code snippet for folks to test it out (@linoytsaban @asomoza).
Remove the custom block implementations from the PR, host them on the Hub (just like this one), and guide the users about how to use them alongside the pipeline.

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

sayakpaul · 2025-10-27T06:35:47Z

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

+```bash
+hf auth login
+```
+


Feel free to talk a little more about how "control" is interfaced in the pipeline i.e., what users can do with the pipeline to take "control".

sayakpaul · 2025-10-27T06:36:07Z

src/diffusers/models/transformers/transformer_bria_fibo.py

@@ -0,0 +1,446 @@
+from typing import Any, Dict, List, Optional, Tuple, Union


Feel free to add the licensing header.

sayakpaul · 2025-10-27T06:36:54Z

src/diffusers/models/transformers/transformer_bria_fibo.py

+from ...models.modeling_utils import ModelMixin
+from ...models.normalization import AdaLayerNormContinuous, AdaLayerNormZeroSingle
+from ...models.transformers.transformer_bria import BriaAttnProcessor
+from ...models.transformers.transformer_flux import FluxTransformerBlock


@DN6, is this a pattern we want to avoid? 👀

Yup. It would be ideal if we can use #Copy from and define a BriaTransformerBlock inside this file.

src/diffusers/models/transformers/transformer_bria_fibo.py

tests/models/transformers/test_models_transformer_bria_fibo.py

tests/pipelines/bria_fibo/test_pipeline_bria_fibo.py

sayakpaul · 2025-10-27T06:55:45Z

tests/pipelines/bria_fibo/test_pipeline_bria_fibo.py

+        max_diff = np.abs(output_same_prompt - output_different_prompts).max()
+        assert max_diff > 1e-6
+
+    def test_image_output_shape(self):


Seems like we're only testing for a single image pair. Is that expected for this test?

tests/pipelines/bria_fibo/test_pipeline_bria_fibo.py

DN6 · 2025-10-27T07:50:52Z

src/diffusers/models/transformers/transformer_bria_fibo.py

+from ...models.modeling_utils import ModelMixin
+from ...models.normalization import AdaLayerNormContinuous, AdaLayerNormZeroSingle
+from ...models.transformers.transformer_bria import BriaAttnProcessor
+from ...models.transformers.transformer_flux import FluxTransformerBlock


Yup. It would be ideal if we can use #Copy from and define a BriaTransformerBlock inside this file.

DN6 · 2025-10-27T07:58:09Z

src/diffusers/models/transformers/transformer_bria_fibo.py

+
+        processor = BriaAttnProcessor()
+
+        self.attn = Attention(


Can be done in a follow up, but we're moving towards defining all components of a model within a single file (with some exceptions for timestep embeddings and norms). This means defining a dedicated Attention class per model. e.g BriaAttention

If it's the same as the FluxAttention, we can use #Copied from

Reference:

diffusers/src/diffusers/models/transformers/transformer_flux.py

Line 275 in 500b9cf

class FluxAttention(torch.nn.Module, AttentionModuleMixin):

DN6 · 2025-10-27T09:26:01Z

src/diffusers/pipelines/bria_fibo/pipeline_bria_fibo.py

+            latents_scaled = [latent / latents_std + latents_mean for latent in latents]
+            latents_scaled = torch.cat(latents_scaled, dim=0)
+            image = []
+            for scaled_latent in latents_scaled:


Think we can just use self.vae.decode on the latent here directly. Instance level decoding can be done by setting pipe.vae.enable_slicing()

DN6 · 2025-10-27T11:29:55Z

src/diffusers/pipelines/bria_fibo/pipeline_bria_fibo.py

+        return noise_scheduler, timesteps, num_inference_steps, mu
+
+    @staticmethod
+    def create_attention_matrix(attention_mask):


Could we name this _prepare_attention_mask for consistency with other pipelines?

diffusers/src/diffusers/pipelines/chroma/pipeline_chroma.py

Line 599 in dc6bd15

def _prepare_attention_mask(

DN6 · 2025-10-27T11:37:36Z

src/diffusers/pipelines/bria_fibo/pipeline_bria_fibo.py

+        return latents, latent_image_ids
+
+    @staticmethod
+    def init_inference_scheduler(height, width, device, image_seq_len, num_inference_steps=1000, noise_scheduler=None):


Possible to just have these steps in the __call__ method? Similar to

diffusers/src/diffusers/pipelines/flux/pipeline_flux.py

Lines 868 to 887 in dc6bd15

sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas

if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas:

sigmas = None

image_seq_len = latents.shape[1]

mu = calculate_shift(

image_seq_len,

self.scheduler.config.get("base_image_seq_len", 256),

self.scheduler.config.get("max_image_seq_len", 4096),

self.scheduler.config.get("base_shift", 0.5),

self.scheduler.config.get("max_shift", 1.15),

)

timesteps, num_inference_steps = retrieve_timesteps(

self.scheduler,

num_inference_steps,

device,

sigmas=sigmas,

mu=mu,

)

num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)

self._num_timesteps = len(timesteps)

DN6 · 2025-10-27T11:37:58Z

src/diffusers/pipelines/bria_fibo/pipeline_bria_fibo.py

+    def init_inference_scheduler(height, width, device, image_seq_len, num_inference_steps=1000, noise_scheduler=None):
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+
+        assert height % 16 == 0 and width % 16 == 0


Checks should be placed under check_inputs

galbria added 2 commits October 26, 2025 16:41

Bria FIBO pipeline

9e253a7

style fixs

371e5f5

sayakpaul reviewed Oct 27, 2025

View reviewed changes

sayakpaul requested a review from DN6 October 27, 2025 08:30

DN6 reviewed Oct 27, 2025

View reviewed changes

		@@ -0,0 +1,446 @@
		from typing import Any, Dict, List, Optional, Tuple, Union

	sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
	if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas:
	sigmas = None
	image_seq_len = latents.shape[1]
	mu = calculate_shift(
	image_seq_len,
	self.scheduler.config.get("base_image_seq_len", 256),
	self.scheduler.config.get("max_image_seq_len", 4096),
	self.scheduler.config.get("base_shift", 0.5),
	self.scheduler.config.get("max_shift", 1.15),
	)
	timesteps, num_inference_steps = retrieve_timesteps(
	self.scheduler,
	num_inference_steps,
	device,
	sigmas=sigmas,
	mu=mu,
	)
	num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
	self._num_timesteps = len(timesteps)

Uh oh!

Bria fibo #12545

Are you sure you want to change the base?

Bria fibo #12545

Conversation

galbria commented Oct 26, 2025

What does this PR do?

Before submitting

Who can review?

Uh oh!

sayakpaul left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants