TIDE: Target-Instructed Diffusion Enhancing

• TIDE is a framework for balanced subject-driven image generation that uses target-supervised triplet alignment to explicitly optimize both subject preservation and instruction adherence.
• It replaces self-supervised reconstruction with target supervision and employs a Direct Subject Diffusion objective to learn the preservation-compliance trade-off effectively.
• The method trains a lightweight multimodal adapter on a frozen Stable Diffusion backbone, achieving competitive metrics and improved scalability without test-time fine-tuning.

Target-Instructed Diffusion Enhancing (TIDE) is a framework for balanced subject-driven image generation (SDIG) in which a diffusion model receives a reference subject image and a textual instruction, and is trained to produce outputs that preserve subject identity while complying with the requested edit. In the formulation introduced in “TIDE: Achieving Balanced Subject-Driven Image Generation via Target-Instructed Diffusion Enhancement,” TIDE replaces self-supervised reconstruction with target-supervised triplet learning and adapts preference optimization to diffusion models through the Direct Subject Diffusion (DSD) objective, using paired winning and losing targets to learn the preservation-compliance trade-off without test-time fine-tuning (Lin et al., 8 Sep 2025).

1. Problem formulation and conceptual scope

TIDE is defined around a specific failure mode in personalized or subject-driven generation: subject identity preservation and instruction compliance are jointly required, but existing methods often over-optimize one at the expense of the other. In the paper’s SDIG formulation, the inputs are a source or reference image $x_i$ representing the subject and a text instruction $x_p$; the desired output should depict the same subject while executing edits such as changes in pose, background, style, or local attributes (Lin et al., 8 Sep 2025).

The method is motivated by a distinction between two established paradigms. Fine-tuning-based methods adapt the diffusion model per subject at test time and can preserve identity strongly, but the paper associates them with poor scalability and catastrophic forgetting. Fine-tuning-free methods inject subject information through external encoders or adapters, but the paper argues that they are usually trained with self-supervised reconstruction, so the input image becomes a pseudo-target. Under that regime, “preserve the subject” is conflated with “reconstruct the input,” and the model is not explicitly taught how a subject should change under edited instructions.

A central conceptual claim is therefore that the SDIG problem is not adequately specified by reconstruction alone. TIDE reformulates it as a target-supervised and preference-aligned diffusion problem in which desirable edited outcomes are explicitly contrasted against undesirable ones. This suggests that TIDE treats balanced generation not as a latent by-product of conditioning, but as a directly supervised optimization target.

2. Target-supervised triplet alignment and the C4DD dataset

The core supervision unit in TIDE is the triplet

$(x_p, x_i, y_0^w, y_0^l),$

where $x_p$ is the instruction text, $x_i$ is the subject image, $y_0^w$ is a winning target image, and $y_0^l$ is a losing target image (Lin et al., 8 Sep 2025). The paper refers to this as target-supervised triplet alignment. Instead of reconstructing the input subject image, the model is trained against explicit edited targets that encode what a good preservation-compliance balance should look like.

To support this training regime, the paper constructs C4DD, expanded as Concept101 for DPO Dataset. C4DD is built from Concept101, which provides multiple instructions per subject. For each subject-instruction pair, candidate outputs are generated using existing subject-driven generation methods, including both fine-tuning and fine-tuning-free systems. These candidates are then ranked using two similarity signals: $$S_{text} = \mathrm{mean}\left(\mathrm{CLIP}(x_p) \cdot \mathrm{CLIP}(y_0)^{\top}\right),$$

$$S_{visual} = \cos(\mathrm{DINO}(x_i), \mathrm{DINO}(y_0)),$$

combined as

$$Q = \varphi \cdot S_{text} + (1-\varphi)\cdot S_{visual},$$

with

$x_p$0

The ranking protocol divides candidates into five levels by $x_p$1: Level 5 is best, Level 4 next best, and Level 1 worst. Winning samples are defined as Level 5 and Level 4; losing samples are defined as Level 3, 2, and 1. The resulting dataset contains 10,000 high-quality samples and 6,000 curated (winning, losing) pairs (Lin et al., 8 Sep 2025).

This ranking scheme is important because the preference labels are not described as human ratings alone; they are automatically constructed from CLIP- and DINO-based signals. A plausible implication is that TIDE’s supervision pipeline externalizes the preservation-compliance trade-off into a measurable ranking criterion before diffusion training begins.

3. Direct Subject Diffusion and implicit reward modeling

The optimization core of TIDE is Direct Subject Diffusion (DSD), a diffusion-form preference objective derived from DPO-style training. The paper first restates the standard diffusion denoising objective as

$x_p$2

It then reviews diffusion-preference training through DPO-Diffusion, and extends that construction to multimodal subject conditioning (Lin et al., 8 Sep 2025).

The DSD loss is given as

$x_p$3

The paper decomposes this loss into four functional roles: Preference Alignment, Reference Regularization, Dispreference Repulsion, and Reference Normalization. The trainable model is pulled toward the winning target, pushed away from the losing target, and evaluated relative to a frozen reference model rather than in isolation. This makes the objective contrastive in both target quality and model deviation.

TIDE also frames DSD as implicit reward modeling. Instead of learning an explicit reward model $x_p$4, it uses paired preferences directly in the diffusion loss. In that sense, the trainable adapter becomes the preference-aligned policy. This suggests that TIDE imports the logic of Bradley–Terry and DPO into SDIG while staying in noise-prediction space rather than probability-ratio space alone.

4. Conditioning architecture and trainable components

TIDE is built on Stable Diffusion v1.5 and keeps the base diffusion model frozen, while updating only a lightweight multimodal adapter (Lin et al., 8 Sep 2025). The trainable components are:

• IPM: Image Projection Module, an MLP mapping CLIP image features into the cross-attention space.
• ICAM: Image Cross-Attention Module, adapted from Stable Diffusion’s text cross-attention.

The paper states that only about 1.33% of base model parameters are trained. The conditioning pipeline is explicitly multimodal. The instruction is encoded with the frozen CLIP text encoder; the reference image is encoded with the CLIP image encoder; image features are projected through IPM; and text and image features are fused through ICAM into a joint conditioning representation.

The image feature extraction and conditioning fusion are written as

$x_p$5

$x_p$6

$x_p$7

This fusion formula makes the architectural intent explicit: text guidance specifies the edit semantics, while subject anchoring injects identity-specific visual evidence. The paper later uses the same $x_p$8 parameter for text-image interpolation and reports that $x_p$9 yields smooth transitions.

At inference time, no per-subject optimization is performed. The reference image and instruction are provided once, and generation proceeds with the frozen diffusion model plus the trained adapter. The paper therefore distinguishes training-time adaptation of the adapter from test-time fine-tuning of the backbone.

5. Training regime, benchmarks, quantitative results, and ablations

The reported training setup uses 2 NVIDIA V100 GPUs (32GB), 50 epochs, about 40,000 steps, batch size 8, learning rate 1e-5, and a warmup scheduler, with training performed on C4DD (Lin et al., 8 Sep 2025). Evaluation is conducted on Concept101 and DreamBench, both under a common SD v1.5 backbone for fair comparison. The baselines include fine-tuning-based methods—Textual Inversion, DreamBooth, and Custom Diffusion—and fine-tuning-free methods—ELITE, BLIP-Diffusion, IP-Adapter, SSR-Encoder, and Subject-Diffusion. The reported metrics are CLIP-T for text-image similarity or instruction compliance, CLIP-I for image-image similarity or subject fidelity, and DINO for visual similarity or subject alignment.

On Concept101, TIDE reports CLIP-T 0.304, CLIP-I 0.814, and DINO 0.690. On DreamBench, it reports CLIP-T 0.314, CLIP-I 0.826, and DINO 0.676. The paper characterizes these as best or tied-best on all three reported metrics for Concept101, and as best on CLIP-T and CLIP-I with second-best DINO on DreamBench.

The ablation study isolates the contributions of dataset construction, target supervision, DSD, and multimodal fusion. The full model on DreamBench reports CLIP-T 0.314, CLIP-I 0.826, and DINO 0.676. Removing training on C4DD yields 0.274 / 0.809 / 0.608; removing DSD yields 0.275 / 0.820 / 0.665; removing target supervision yields 0.265 / 0.815 / 0.668; and removing IPM-ICAM yields 0.299 / 0.719 / 0.637. These results are used in the paper to support three claims: target supervision provides more informative training signals than self-reconstruction, preference learning improves semantic balance, and feature fusion is necessary for strong subject fidelity.

The same section of the paper also treats the ranking coefficient $(x_p, x_i, y_0^w, y_0^l),$0 as an ablation variable and reports best performance at

$(x_p, x_i, y_0^w, y_0^l),$1

which prioritizes instruction compliance while still preserving subject identity.

6. Extensions, related directions, and acronym disambiguation

The paper presents TIDE as extensible beyond the base SDIG setting. It reports compatibility with ControlNet for structural-conditioned generation using depth maps, normal maps, pose skeletons, canny edges, and free-form scribbles, and states that this compatibility does not require additional fine-tuning (Lin et al., 8 Sep 2025). It also applies the method to image-to-image generation, including style transfer settings such as watercolor, painterly style, and synthetic art style, and to text-image interpolation through the conditioning parameter $(x_p, x_i, y_0^w, y_0^l),$2.

In the broader diffusion-alignment landscape, a conceptually related but differently named framework is AliDiff, which aligns pretrained target-aware ligand diffusion models toward preferred functional properties through preference optimization and Exact Energy Preference Optimization in structure-based drug design (Gu et al., 2024). A plausible comparison is that both methods shift a pretrained diffusion distribution toward a target-defined preference signal, but TIDE does so for subject-driven image generation, whereas AliDiff does so for protein-conditioned molecule generation.

The acronym TIDE is also reused across several unrelated research threads, which is a recurrent source of confusion in bibliographic and search contexts. Distinct expansions include “Task-Isolated Diffusion for Unified Video Editing and Generation” (Liu et al., 6 Jun 2026), “Token Identity Delivered Everywhere” (Jaiswal et al., 7 May 2026), “Turning the TIDE” for cross-architecture distillation in diffusion LLMs (Zhang et al., 29 Apr 2026), “Text-Informed Dynamic Extrapolation with Step-Aware Temperature Control for Diffusion Transformers” (Liu et al., 9 Mar 2026), and “Two-Stage Inverse Degradation Estimation with Guided Prior Disentanglement” for underwater image restoration (Venkatraman et al., 8 Dec 2025). In the present sense, however, TIDE denotes Target-Instructed Diffusion Enhancing, specifically the SDIG framework centered on target-supervised triplet alignment, DSD-based preference learning, and adapter-only conditioning over a frozen Stable Diffusion v1.5 backbone (Lin et al., 8 Sep 2025).