add fibo-2 LoRA training and loading

docs/source/en/api/loaders/lora.md

@@ -39,6 +39,7 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`ErnieImageLoraLoaderMixin`] provides similar functions for [Ernie-Image](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ernie_image).
 - [`LTX2LoraLoaderMixin`] provides similar functions for [Flux2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx2).
 - [`MiniMaxH3LoraLoaderMixin`] provides similar functions for [MiniMax-H3](https://huggingface.co/docs/diffusers/main/en/api/pipelines/minimax_h3).
+- [`BriaFibo2LoraLoaderMixin`] provides similar functions for [fibo-2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/bria_fibo2).
 - [`LoraBaseMixin`] provides a base class with several utility methods to fuse, unfuse, unload, LoRAs and more.
 
 > [!TIP]
@@ -72,6 +73,10 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 
 [[autodoc]] loaders.lora_pipeline.Flux3LoraLoaderMixin
 
+## BriaFibo2LoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.BriaFibo2LoraLoaderMixin
+
 ## ErnieImageLoraLoaderMixin
 
 [[autodoc]] loaders.lora_pipeline.ErnieImageLoraLoaderMixin

examples/dreambooth/README_bria_fibo2.md

@@ -0,0 +1,204 @@
+# DreamBooth LoRA training for fibo-2 (Bria)
+
+[fibo-2](https://huggingface.co/briaai) is Bria's flow-matching text-to-image and image-editing model
+(`BriaFibo2Pipeline`): a `BriaFibo2Transformer2DModel` denoiser conditioned on a Qwen3-VL text encoder, with an
+`AutoencoderKLFlux2` VAE. This directory has two LoRA trainers:
+
+- `train_dreambooth_lora_bria_fibo2.py` — text-to-image.
+- `train_dreambooth_lora_bria_fibo2_img2img.py` — image editing (image-to-image).
+
+Both train a LoRA adapter on the transformer only; the text encoder and VAE are frozen.
+
+> [!NOTE]
+> fibo-2 is trained on **structured JSON captions** and renders them best, but a plain-text short description is
+> accepted too (and used as-is). Keep your training captions in the form you intend to prompt with. The trainers
+> clean captions the way the pipeline does, so a JSON caption and its cleaned, compact form train the same.
+
+## Captions
+
+fibo-2 was trained on **structured JSON captions**, so prompt and train in the same form. The trainers pass each
+caption through the pipeline's own helpers, cleaning and compacting a JSON caption exactly as inference does (empty
+values and the `aesthetic_score`/`preference_score` keys are dropped), so the model sees the same text both times.
+
+**Text-to-image.** A caption is a JSON object; include only the keys you need:
+
+```json
+{
+  "short_description": "A red bicycle leaning against a white brick wall on a sunny morning.",
+  "objects": [{"description": "a red city bicycle with a wicker basket", "location": "center"}],
+  "background_setting": "a white brick wall with a green door",
+  "lighting": {"conditions": "bright morning sun", "direction": "from the right"},
+  "style_medium": "photograph"
+}
+```
+
+Pass one caption per image through `--caption_column` (each value a JSON string, or a plain sentence used as the
+`short_description`), or a single `--instance_prompt` for the whole set. To place literal text in the image, fibo-2
+uses a `text_render` field; the trainers here do not train that path, but the base model still renders short quoted
+text in a prompt reasonably well.
+
+**Image editing.** The edit instruction (`--caption_column`, with `--instance_prompt` as the fallback) takes either
+form the pipeline accepts, and the two give different text embeddings — train in the one you will prompt with:
+
+- a **plain instruction** — `"Make the background a snowy forest"`; or
+- a **JSON object** with the instruction under `edit_instruction` and a structured caption of the *result*:
+  `{"short_description": "...", "edit_instruction": "Make the background a snowy forest"}` (the caption may instead
+  sit under a `structured_caption` key).
+
+## Running locally
+
+### Installing the dependencies
+
+Install `diffusers` from source and then the trainer requirements:
+
+```bash
+pip install -e .
+cd examples/dreambooth
+pip install -r requirements_bria_fibo2.txt
+```
+
+And initialize an [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+Or for a default accelerate configuration without answering questions about your environment:
+
+```bash
+accelerate config default
+```
+
+fibo-2's Qwen3-VL text encoder needs `transformers>=5.20.1`.
+
+## Text-to-image
+
+Train a LoRA on a folder of instance images (one trigger prompt for all of them):
+
+```bash
+export MODEL_NAME="briaai/fibo-2"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-fibo2-lora"
+
+accelerate launch train_dreambooth_lora_bria_fibo2.py \
+  --pretrained_model_name_or_path=$MODEL_NAME \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --mixed_precision="bf16" \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --learning_rate=1e-4 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=1000 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed=0 \
+  --push_to_hub
+```
+
+To train on a dataset with per-image captions, pass `--dataset_name` and `--caption_column` (the captions can be
+structured JSON captions or plain short descriptions) instead of `--instance_data_dir`:
+
+```bash
+accelerate launch train_dreambooth_lora_bria_fibo2.py \
+  --pretrained_model_name_or_path="briaai/fibo-2" \
+  --dataset_name="your-dataset" \
+  --caption_column="text" \
+  --instance_prompt="in the style of TOK" \
+  --output_dir="trained-fibo2-lora" \
+  --mixed_precision="bf16" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --learning_rate=1e-4 \
+  --max_train_steps=1000
+```
+
+### Inference
+
+```python
+from diffusers import BriaFibo2Pipeline
+import torch
+
+pipeline = BriaFibo2Pipeline.from_pretrained("briaai/fibo-2", torch_dtype=torch.bfloat16).to("cuda")
+pipeline.load_lora_weights("trained-fibo2-lora", weight_name="pytorch_lora_weights.safetensors")
+image = pipeline("A photo of sks dog in a bucket").images[0]
+image.save("fibo2.png")
+```
+
+## Image editing (image-to-image)
+
+The editing trainer learns from **(conditioning image, target image, edit instruction)** triplets, so it needs a
+dataset (`--dataset_name`) with a source-image column, a target-image column and a caption column holding the
+instruction. The conditioning and target images are given the same geometry so each pair stays aligned; the
+conditioning image is both shown to the Qwen3-VL text encoder (as the image to edit) and VAE-encoded into the
+transformer's clean context tokens, exactly as `BriaFibo2Pipeline` does at inference.
+
+```bash
+accelerate launch train_dreambooth_lora_bria_fibo2_img2img.py \
+  --pretrained_model_name_or_path="briaai/fibo-2" \
+  --dataset_name="your-edit-dataset" \
+  --image_column="target_image" \
+  --cond_image_column="source_image" \
+  --caption_column="instruction" \
+  --instance_prompt="make it a watercolor painting" \
+  --output_dir="trained-fibo2-edit-lora" \
+  --mixed_precision="bf16" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --learning_rate=1e-4 \
+  --max_train_steps=2000 \
+  --validation_prompt="make it a watercolor painting" \
+  --validation_image="path/to/source.png" \
+  --validation_epochs=25
+```
+
+`--instance_prompt` is the fallback instruction used when a pair has no caption. `--caption_column` holds the edit
+instruction as plain text.
+
+> [!NOTE]
+> For editing, the conditioning image is shown to Qwen3-VL, and its vision tokens count towards
+> `--max_sequence_length` (a 768px image is up to ~576 tokens). The editing trainer defaults it to 1024 for that
+> reason; raise it further only if you pair large images with long structured captions.
+
+### Inference
+
+```python
+from diffusers import BriaFibo2Pipeline
+from diffusers.utils import load_image
+import torch
+
+pipeline = BriaFibo2Pipeline.from_pretrained("briaai/fibo-2", torch_dtype=torch.bfloat16).to("cuda")
+pipeline.load_lora_weights("trained-fibo2-edit-lora", weight_name="pytorch_lora_weights.safetensors")
+image = load_image("path/to/source.png")
+edited = pipeline("make it a watercolor painting", image=image).images[0]
+edited.save("fibo2_edit.png")
+```
+
+## Notes
+
+- **LoRA layers.** By default the adapter targets **every linear layer** of the transformer (PEFT's `"all-linear"`),
+  matching Bria's reference trainer. Pass `--lora_layers` with a comma-separated list to target a narrower set, e.g.
+  `--lora_layers="to_q,to_k,to_v,to_out.0,w1,w2,w3"` for attention and feed-forward only, or
+  `--lora_layers="to_q,to_k,to_v,to_out.0"` for attention only.
+- **Aspect-ratio bucketing.** Pass `--use_aspect_ratio_buckets` to bucket images by aspect ratio (optionally with an
+  explicit `--aspect_ratio_buckets` list), instead of resizing everything to a square `--resolution`.
+- **Caption dropout.** `--caption_dropout 0.1` trains a fraction of samples unconditionally, which improves
+  classifier-free guidance. For editing, the conditioning image is always kept; only the instruction is dropped.
+- **Memory.** Use `--cache_latents` to pre-encode images, `--gradient_checkpointing`, `--use_8bit_adam`, and
+  `--offload` to move the VAE and text encoder to CPU when idle. `--bnb_quantization_config_path` enables
+  bitsandbytes-quantized training of the transformer, and `--do_fp8_training` enables FP8 (TorchAO) training.
+- **Prodigy.** Pass `--optimizer="prodigy"` (with `--learning_rate=1.0`) to use the Prodigy optimizer.
+
+> [!NOTE]
+> These scripts are provided as a starting point; tune hyperparameters (rank, learning rate, steps) for your data.
+
+## More resources
+
+Bria publishes a single-GPU reference LoRA trainer, the full caption schema, and their recommended settings in the
+model's [`fine_tuning` folder](https://huggingface.co/briaai/Fibo-2.0-dev/tree/main/fine_tuning). Their suggested
+starting point is rank 64 (alpha = rank), learning rate 1e-4 constant, 1024px, gradient accumulation 8, and LoRA on
+every linear layer. The trainers here also default to every linear layer; consider raising `--rank`/`--lora_alpha`
+(the defaults are small) and lowering the target set with `--lora_layers` if you want a lighter adapter.

examples/dreambooth/requirements_bria_fibo2.txt

@@ -0,0 +1,11 @@
+accelerate>=0.31.0
+torchvision
+transformers>=5.20.1
+ftfy
+tensorboard
+Jinja2
+peft>=0.11.1
+sentencepiece
+bitsandbytes
+prodigyopt
+datasets>=3.1.0