FOCUS: A Principled Route to Multi‑Subject Fidelity

Eric Tillmann Bill, Enis Simsar, Thomas Hofmann

Abstract. Text-to-image (T2I) models excel on single-entity prompts but struggle with multi-subject descriptions, often showing attribute leakage, identity entanglement, and subject omissions. We introduce the first theoretical framework with a principled, optimizable objective for steering sampling dynamics toward multi-subject fidelity. Viewing flow matching (FM) through stochastic optimal control (SOC), we formulate subject disentanglement as control over a trained FM sampler. This yields two architecture-agnostic algorithms: (i) a training-free test‑time controller that perturbs the base velocity with a single-pass update, and (ii) Adjoint Matching, a lightweight fine-tuning rule that regresses a control network to a backward adjoint signal while preserving base-model capabilities. The same formulation unifies prior attention heuristics, extends to diffusion models via a flow–diffusion correspondence, and provides the first fine-tuning route explicitly designed for multi-subject fidelity. Empirically, on Stable Diffusion 3.5, FLUX, and Stable Diffusion XL, both algorithms consistently improve multi-subject alignment while maintaining base-model style. Test-time control runs efficiently on commodity GPUs, and fine-tuned controllers trained on limited prompts generalize to unseen ones. We further highlight FOCUS (Flow Optimal Control for Unentangled Subjects), which achieves state-of-the-art multi-subject fidelity across models.

                Base Models                                                 Base Models + FOCUS (Ours)
               

Highlights

• Two ways to use FOCUS
  • Test‑time controller: training‑free, perturbs the base sampler to encourage disentanglement.
  • Adjoint Matching: efficient fine‑tuning that learns a control network without degrading base capabilities.
• Model‑agnostic: demonstrated with SD3.5, FLUX, and SDXL.
• Plug‑and‑play: minimal changes to official pipelines; FOCUS controller is passed into the sampler.
• Efficient: test‑time control runs on commodity GPUs (≈12 GB VRAM works with memory‑saving modes).

---

Installation

# Clone the repository
git clone https://github.com/ericbill21/FOCUS.git
cd FOCUS

# Install dependencies
pip install -r requirements.txt

Note: The example notebooks rely on the official SD3 / FLUX pipelines with a small wrapper to accept the Controller object.

---

Test‑Time Control

We provide two simple notebooks to run FOCUS without changing model weights:

• sample_sd3_otf.ipynb (Stable Diffusion 3.5)
• sample_flux_otf.ipynb (FLUX‑1 [dev])

Each pipeline is a lightly modified copy of the official implementation to accept a Controller that specifies the heuristic, λ (lambda), and subject indices.

from focus.controller import Controller

controller = Controller(
    t5_ids=[[1], [5]],      # indices of prompt subjects in the T5 encoder
    clip_ids=[[2], [5]],    # indices of prompt subjects in the CLIP encoder
    lambda_scale=4.0,
    heuristic="focus",        # one of: "focus", "conform", "attend_and_excite", "divide_and_bind"
    model="SD3"            # "SD3" or "FLUX"
)

# Pass `controller` to the sampling pipeline in the notebook or your script.

Tips

• Set lambda_scale=0 to disable the controller (baseline).
• Start with moderate λ (e.g., 1–5) and tune per prompt.

---

Fine‑Tuning (Adjoint Matching)

We provide two training scripts (one per base model). Both follow the same logic.

Inference on FOCUS+Finetuned Models

Effective Batch Size

Let num_traj be the number of trajectories per iteration and k the length (time steps) of each trajectory. The effective batch size is num_traj × k. We further sub‑sample time steps via sub-start, sub-end, and sub-extra to reduce memory and avoid overfitting.

Common Arguments

Group	Flag	Default	Description
Setup	--res-dir	–	Directory to save results/checkpoints
	--dataset	–	Path to a pre‑encoded dataset. We provided examples in datasets/data_finetuning/
	--ckpt-every	100	Save a checkpoint every N iterations
	--verbose	1	0: warning, 1: info, 2: debug
Training	--num-iterations	400	Number of training iterations
	--learning-rate	1e-5	Optimizer learning rate
	--dtype	bfloat16	Compute dtype (e.g., bfloat16, float32)
Scheduler / Solver	--num-traj	5	Trajectories sampled per iteration
	--k	28	Time steps per trajectory
	--lambda-value	1.0	λ in the reward/objective
Sub‑sampling	--sub-start	0	Start index for sub‑sampling time steps
	--sub-end	0	End index for sub‑sampling time steps
	--sub-extra	5	Additional random time steps
LoRA	--lora-rank	4	LoRA rank
	--lora-alpha	16	LoRA α
	--lora-dropout	0.05	LoRA dropout
Misc	--seed	42	Random seed
	--image-size	256	Training image size

Example

python finetune_sd3.py \
  --dataset datasets/data_finetuning/example.yaml \
  --res-dir runs/sd3_focus \
  --num-iterations 200 --learning-rate 5e-5 \
  --num-traj 5 --k 28 --lambda-value 1.0 \
  --sub-start 0 --sub-end 0 --sub-extra 16 \
  --lora-rank 4 --lora-alpha 16 --lora-dropout 0.05 \
  --dtype bfloat16 --image-size 256 --seed 42

---

Sampling Script (sample.py)

sample.py can generate images using test‑time control, fine‑tuned weights, or both. Disable the test-time controller by setting --lambda-scale 0.

Key Arguments

Group	Flag	Default	Description
Paths	--exp-dir	images	Output directory
	--dataset	(required)	Path to a dataset YAML of prompts. We provide examples in datasets/
Weights	--path	–	Path to fine‑tuned weights (optional)
Sampler	--image-size	512	Output resolution
	--num-steps	model default	ODE/diffusion steps
	--guidance-scale	model default	Classifier‑free guidance
Seeds	--seed	0	Single seed
	--seed-range	–	Range [a b] to sweep multiple seeds
Controller	--lambda-scale	1.0	λ for controller; set 0 to disable
	--heuristic	focus	One of {focus, conform, attend_and_excite, divide_and_bind}
	--model	SD3	Base model: {SD3, FLUX}
Memory	--save-memory	0	0: none, 1: offload, 2: sequential offload, 3: + grad checkpointing

Example

python sample.py \
  --dataset datasets/natural_prompts.yaml \
  --exp-dir outputs/sd3_focus \
  --model SD3 --heuristic focus --lambda-scale 4 \
  --image-size 512 --seed-range 0 10

---

Metrics & Evaluation

Use eval_metrics.py to compute all metrics reported in the paper.

1. Generate a set of images with sample.py.
2. Run eval_metrics.py on the produced images.
3. The script writes a CSV next to the image folder with per‑image metrics.

python eval_metrics.py \
  --data-dir outputs/sd3_focus \
  --batch-size 16

---

Datasets

We provide all used datasets in the folder datasets/, which are simple YAML files, containing for each prompt, the subjects, and their corresponding T5 and CLIP indices. For the finetuning, we further provide already encoded versions of datasets/two_objects.yaml and datasets/horse_bear.yaml, in the folder datasets/data_finetuning/.

---

Model Checkpoints

We released our best perfoming checkpoints for FLUX and SD3 on Huggingface.

---

Troubleshooting

• OOM / CUDA out of memory: reduce --image-size, use --save-memory 2 or 3, or reduce --k / --num-traj.
• No improvement from controller: tune --lambda-scale, verify subject indices (t5_ids, clip_ids), and try alternative heuristics (conform, attend_and_excite, divide_and_bind).
• Slow sampling: decrease steps, enable memory‑saving modes, or disable gradient‑heavy options in notebooks.

---

📄 Citation

If you find our work useful, please consider citing:

@misc{bill2025focus,
  title         = {Optimal Control Meets Flow Matching: A Principled Route to Multi-Subject Fidelity},
  author        = {Eric Tillmann Bill and Enis Simsar and Thomas Hofmann},
  year          = {2025},
  eprint        = {2505.19166},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

---

Questions or issues? Please open a GitHub issue or reach out via the project website.