Add Bria fibo-2

[muhammadali-afridi]

What does this PR do?

Adds fibo-2, Bria's text-to-image and image-editing model.

• BriaFibo2Transformer2DModel: an 8.1B flow-matching transformer with a single token stream, in the style of Z-Image. Its text encoder is Qwen3-VL-4B.

  • A Perceiver condenses Qwen3-VL's hidden states into gist tokens, which join the image tokens.
  • Five blocks also read Qwen3-VL through gated cross-attention.
  • Images to edit come in as context_latents. Each is patchified by the shared embedder and gets its own RoPE plane.

• BriaFibo2Pipeline:

  • text-to-image from structured JSON captions;
  • editing of up to 5 images, with an optional mask.

  Two checkpoints share the pipeline:

  • a distilled turbo checkpoint: 4 steps, no guidance;
  • a full checkpoint: 30 steps, guidance 5.

• scripts/convert_bria_fibo2_to_diffusers.py converts the original checkpoints.

• Docs and tests.

This is a draft until the weights are public. The repo ids in the docs are placeholders until then.

Verification

I compared the pipeline against Bria's reference inference code, using the original checkpoints, the same seeds and bf16:

• A100, SDPA:
  • Text-to-image is pixel-identical, for both the turbo and the full checkpoint.
  • Editing is pixel-identical when the text encoder gets the same images as the reference.
• H100: the reference code runs FlashAttention-3 there. With pipe.transformer.set_attention_backend("_flash_3_varlen_hub"):
  • text-to-image is pixel-identical for both checkpoints;
  • editing is pixel-identical too (4 turbo cases and 2 full-checkpoint cases).
• Text-encoder images: by default the pipeline thumbnails each image straight to 768 px for Qwen3-VL, as in training. The reference code upsamples 4× first, which changes that input slightly.
• Batching: in float32, a batch of prompts of different lengths matches the separate single-prompt calls. This includes the varlen attention backends; see the self-review notes.
• Checks:
  • make style, make quality and make fix-copies pass.
  • check_copies, check_dummies, check_support_list, check_forward_call_docstrings and check_return_annotations pass.
• New tests: they pass on a Mac (CPU/MPS), except the group offloading and layerwise casting memory tests. Those fail the same way for Z-Image and Flux 2 on that machine, because MPS has no CUDA streams and no float8.

Open before ready for review

• Hosted diffusers-format weights and the final repo ids.
• The license header of the new files.
• mask:
  • The pipeline greys out the region to regenerate, which is the convention of Bria's earlier edit models.
  • We're confirming that fibo-2 follows it.
  • The docs sentence on masks will be updated once that's confirmed.
• Caption score fields: training dropped aesthetic_score and preference_score from captions.
  • The pipeline strips them from edit prompts.
  • Text-to-image prompts are passed as given.
  • To decide: strip them for text-to-image too?
• Standard vs modular: this is a standard pipeline. Happy to add a modular one if that's preferred.

Self-review notes

I ran the self-review skill on the diff.

Fixed:

• The transformer forward docstring had no Returns: section, which check_forward_call_docstrings flags.
• Varlen backends kept the wrong keys.
  • With prompts of different lengths, the padded gist tokens sat between the real gist tokens and the image tokens.
  • The varlen backends (flash_varlen, _flash_3_varlen_hub, sage_varlen) keep the first mask.sum() keys, so they kept the wrong ones.
  • Padded gist tokens now go last. test_padded_batch_matches_single_prompts covers this.
• Removed dead code: a PyTorch 2.0 check in the attention processor, and casts that changed nothing.
• __call__ read the first image's size before check_inputs had validated it.

Not fixed, on purpose:

• Full-graph compile tests are skipped. RopeEmbedder, copied from Z-Image, builds its frequency table under a device context on first use. Z-Image skips these tests for the same reason.
• Gist counts break the graph under torch.compile(transformer).
  • With a text mask, the gist count per prompt comes from the text lengths through .tolist().
  • Regional compilation (compile_repeated_blocks) is not affected.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• 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?
• Did you write any new necessary tests?
• Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

I'll tag reviewers once the weights are public and the open items above are settled.

🤖 Generated with Claude Code