EditIDv2: Editable Identity Customization

• EditIDv2 is a character and identity customization technique designed for long narrative prompts that balances maintaining recognizable features with flexible semantic editing.
• It employs a Flux-based DiT with flow-matching and a decomposed PerceiverAttention module to dynamically control identity influence during the denoising process.
• The method leverages minimal 'data lubrication'—around 3K multi-attribute samples—to train its identity integration module, achieving a trade-off between fidelity and editability.

EditIDv2 is a character and identity customization method for text-to-image generation whose central objective is not merely to keep a face recognizable, but to keep that identity editable under long, story-like prompts and high-complexity narrative scenes. It is presented as a tuning-free solution built on a Flux-based DiT plus flow-matching framework, and it attributes failures of prior character editing methods under long prompts to degraded editing capabilities, semantic understanding bias, contextual and temporal inconsistency, and identity consistency breakdowns. Its core claim is that editable ID customization can be achieved with only a small amount of additional supervised data, termed “data lubrication,” by redesigning the ID feature integration module rather than simply strengthening identity extraction (Li et al., 6 Sep 2025).

1. Problem setting and conceptual objective

EditIDv2 targets the regime in which prompts contain multiple semantic layers, temporal logic, environmental cues, pose changes, emotional description, and narrative context. In that setting, the method argues that earlier identity-preserving systems often exhibit a characteristic tension: identity injection either overpowers the prompt, producing a rigid “paste-copy phenomenon,” or becomes too weak, causing drift away from the target person. The method is therefore framed around “editable ID customization,” meaning preservation of identity under substantial semantic change rather than static likeness preservation alone (Li et al., 6 Sep 2025).

The paper defines “data lubrication” as introducing a small amount of labeled identity data—multiple images of the same person under different attributes such as pose, lighting, expression, accessories, and scene variation—to lubricate the transfer of identity information from the character feature branch into the image generation branch. In practice, “minimal data lubrication” means about 3K finely labeled samples total, gathered from public data such as MyStyle plus crawled images, with a small number of identities but many attribute-diverse images per identity. The paper explicitly states that the richness comes from many different attribute images per ID rather than many IDs (Li et al., 6 Sep 2025).

This framing places EditIDv2 in a narrower problem class than general-purpose image editors. It is not primarily a real-image editing platform in the style of broader systems that optimize instruction following, image consistency, and image quality across heterogeneous edit types; rather, it is designed for identity-conditioned generation under long prompts and complex scenes. A plausible implication is that its principal axis of evaluation is the editability–consistency trade-off within customized text-to-image generation, not generalized editing coverage across arbitrary real-image transformations (Li et al., 6 Sep 2025, Wang et al., 5 Jun 2025).

2. System decomposition and backbone

Architecturally, EditIDv2 retains the broad decomposition introduced by EditID: an image generation main branch and a character feature branch. The character feature branch is further decomposed into ID feature extraction, feature fusion, and feature integration. The image generation branch is a Flux-based DiT/flow-matching model, and the paper presents the diffusion preliminaries as

$$x_t = a(t) \cdot x_0 + b(t) \cdot \epsilon, \quad \epsilon \sim \mathcal{N}(0, I)$$

with reverse dynamics

$\frac{dx}{dt} = f_\theta(x_t, t, c)$

and flow-matching objective

$$\mathcal{L}(\theta) = \mathbb{E}_{t, x_0, x_1} \left[ \| f_\theta((1-t)x_0 + t x_1, t, c) - (x_1 - x_0) \|^2 \right].$$

For DiT conditioning, the paper writes

$$f_\theta = \text{Projection}(\text{MultiHeadAttention}(x_t W_Q, c W_K, c W_V))$$

and for Flux it uses

$$a(t) = \cos^2(\pi t / 2), \quad b(t) = \sin^2(\pi t / 2)$$

to improve smoothness and stability (Li et al., 6 Sep 2025).

Within this backbone, the character feature branch uses Antelopev2 as the face recognition model and EVA-CLIP as the image encoder. The extraction branch therefore provides identity-related embeddings derived from reference images. The fusion stage combines these identity descriptors into mapping features used for injection. Unlike EditID, the experiments state that EditIDv2 does not directly use EditID’s shift and mapping feature scheme; instead, it retains the original PuLID feature combination. This difference is substantial: EditID concentrated on semantic compression of local features across network depths and inference-time control of ID integration in a training-free regime, whereas EditIDv2 shifts the center of gravity to the feature integration module itself (Li et al., 6 Sep 2025, Li et al., 16 Mar 2025).

The conceptual thesis is correspondingly different. EditID argued that editability depended on decomposing the character feature branch into feature extraction, fusion, and integration, then using mapping features and shift features to form an editable feature space (Li et al., 16 Mar 2025). EditIDv2 accepts the same broad branch decomposition, but positions the ID feature integration module—the bridge between character features and the generator—as the locus where editability is largely determined (Li et al., 6 Sep 2025).

3. ID feature integration and diffusion-time control

The key technical contribution of EditIDv2 is a decomposition of PerceiverAttention into separately controllable query, key, and value paths. The paper gives

$$Q = W_Q \cdot x_{id}, \quad K = W_K \cdot x_{gen}, \quad V = W_V \cdot x_{gen}$$

where $x_{id}$ is the identity feature vector from the character feature branch and $x_{gen}$ is the current state of the image generation branch. The point of this decomposition is that identity conditioning is no longer treated as a monolithic attention block; it can be modulated path by path, especially on the query path (Li et al., 6 Sep 2025).

The principal control is an adjustable mixing between the identity-derived query and a query compensation term derived from the noise image:

$$Q' = \alpha \cdot Q + (1 - \alpha) \cdot Q_{noise}.$$

The paper interprets $Q_{noise}$ as injecting “additional degrees of freedom in the latent space,” thereby preserving editability by allowing prompt semantics and generation noise to influence the result rather than forcing strict identity reconstruction. This is the mechanism by which long-prompt scene descriptions, pose changes, and expression changes can still manifest even when identity guidance is present (Li et al., 6 Sep 2025).

This control is explicitly dynamic over denoising time:

$\frac{dx}{dt} = f_\theta(x_t, t, c)$0

The intended interpretation is that identity should be stronger in the early stages, when global structure and subject identity need to be established, and then progressively relaxed so that later denoising steps can realize fine semantic edits from the prompt. The paper states that this avoids the darkening and instability problems of the blunt control strategy in EditID and better handles multi-level semantic edits such as actions, expressions, and environmental changes under long inputs (Li et al., 6 Sep 2025).

The method further introduces an identity-preservation loss as an auxiliary supervision signal:

$\frac{dx}{dt} = f_\theta(x_t, t, c)$1

As written in the paper, this formula appears to be missing a closing brace around the numerator; faithfully interpreted, it is intended to be cosine similarity subtracted from 1. The total objective is

$\frac{dx}{dt} = f_\theta(x_t, t, c)$2

with $\frac{dx}{dt} = f_\theta(x_t, t, c)$3 in the experiments. The paper also gives an identity-guided denoising ODE,

$\frac{dx}{dt} = f_\theta(x_t, t, c)$4

where $\frac{dx}{dt} = f_\theta(x_t, t, c)$5 is an identity guidance function based on cosine-similarity gradients and $\frac{dx}{dt} = f_\theta(x_t, t, c)$6 uses cosine decay scheduling. The paper does not provide an explicit closed-form expression for $\frac{dx}{dt} = f_\theta(x_t, t, c)$7, only that it uses cosine decay scheduling (Li et al., 6 Sep 2025).

4. Data lubrication, joint training, and offline fusion

The training setup is described as “tuning-free,” but the paper itself requires a narrower reading of that phrase. EditIDv2 is not fully training-free in the literal sense; it preserves a tuning-free deployment style relative to heavy full-model adaptation and fine-tunes only a small part of the system. Specifically, the Flux model is initialized from pretrained weights; only the cross-attention module parameters are trainable, while the VAE, text encoder, and visual encoder are frozen. Training uses AdamW and a cosine learning-rate scheduler (Li et al., 6 Sep 2025).

The data preparation strategy is central to this lightweight adaptation. The authors collect images from MyStyle and internet sources, focusing on multiple images of the same identity with different attributes. They divide the data into three dimensions: rich character-scene data with varied poses, lighting, accessories, and related factors; enhanced data for Asian people in Chinese scenes; and balanced samples under each identity. GPT-4o is used for labeling, and after filtering they obtain about 3K finely labeled training samples. The paper argues that these varied attribute images provide “rich unlocking degrees of freedom” for learning editability: the same identity can legitimately appear under many prompt-compatible changes, which teaches the integration module not to collapse into reconstruction (Li et al., 6 Sep 2025).

The joint dynamic training scheme couples diffusion training to the ID objective rather than treating identity preservation as a terminal regularizer only. The description says that the system randomly samples time step $\frac{dx}{dt} = f_\theta(x_t, t, c)$8, generates noisy images, predicts noise through the model, and jointly optimizes diffusion loss and ID loss. This suggests that the learned integration module is aligned with the diffusion trajectory itself, not just with final image similarity. A plausible implication is that EditIDv2’s editability gains come from making identity negotiable across denoising time rather than fixing it as an always-on constraint (Li et al., 6 Sep 2025).

A further practical mechanism is the offline fusion strategy. Multiple independent attention variants are trained with different emphases—for example, one more strongly weighted toward identity consistency by increasing $\frac{dx}{dt} = f_\theta(x_t, t, c)$9, and another more oriented toward editability by reducing the initial identity-strength parameter $$\mathcal{L}(\theta) = \mathbb{E}_{t, x_0, x_1} \left[ \| f_\theta((1-t)x_0 + t x_1, t, c) - (x_1 - x_0) \|^2 \right].$$0. These are fused offline by weighted averaging:

$$\mathcal{L}(\theta) = \mathbb{E}_{t, x_0, x_1} \left[ \| f_\theta((1-t)x_0 + t x_1, t, c) - (x_1 - x_0) \|^2 \right].$$1

where $$\mathcal{L}(\theta) = \mathbb{E}_{t, x_0, x_1} \left[ \| f_\theta((1-t)x_0 + t x_1, t, c) - (x_1 - x_0) \|^2 \right].$$2 is the weight set for the $$\mathcal{L}(\theta) = \mathbb{E}_{t, x_0, x_1} \left[ \| f_\theta((1-t)x_0 + t x_1, t, c) - (x_1 - x_0) \|^2 \right].$$3-th variant and $$\mathcal{L}(\theta) = \mathbb{E}_{t, x_0, x_1} \left[ \| f_\theta((1-t)x_0 + t x_1, t, c) - (x_1 - x_0) \|^2 \right].$$4 is the fusion coefficient. The coefficients are selected based on validation performance. This is an offline procedure rather than online optimization, and the paper presents it as a way to precompute fused attention weights tailored to the desired consistency/editability trade-off (Li et al., 6 Sep 2025).

5. Inference behavior and relation to EditID

At inference time, EditIDv2 behaves as a lightweight identity-conditioned text-to-image system. A reference identity image and a prompt—potentially long and narratively rich—are provided. Identity features are extracted by the frozen character feature branch and integrated into the generation process via the tuned or fused PerceiverAttention module. No optimization happens during inference; the identity-guided attention mechanism and any multi-weight fusion are already prepared offline. The paper does not define multiple named modes, but functionally it describes at least a consistency-biased setting and an editability-biased setting (Li et al., 6 Sep 2025).

The relation to the original EditID is a matter of architectural emphasis rather than simple iteration. EditID had already argued that many identity-preserving systems behave like implicit identity reconstruction systems and therefore suppress prompt-driven changes. It introduced a decomposition into an image generation branch and a character feature branch, and within the character feature branch it emphasized feature extraction, feature fusion, and feature integration, using mapping features and shift features to form an editable feature space in a training-free way (Li et al., 16 Mar 2025). EditIDv2 retains the same high-level branch structure, but redirects the main intervention to the PerceiverAttention-based integration point, adds an explicit ID loss, performs joint dynamic training with the diffusion model, and introduces offline fusion of multiple attention variants (Li et al., 6 Sep 2025).

The difference from PuLID Flux is equally central to the paper’s interpretation. PuLID Flux is repeatedly described as having too-strong consistency, producing the “paste-copy phenomenon,” where the output face is too close to the input face and prompt responsiveness is limited. EditIDv2 deliberately gives up a small amount of face similarity to gain much stronger editable behavior under long prompts. Compared with methods like InfiniteYou, the paper argues that EditIDv2 attains a better trade-off because InfiniteYou can have strong pose diversity but suffers from poor generation quality and weaker overall stability (Li et al., 6 Sep 2025).

This positioning also differentiates EditIDv2 from adjacent lines of work. SeedEdit 3.0 addresses a related practical tension—edit instruction following versus preservation of “ID/IP” content on real-image inputs—but does so as a broader general-purpose editing platform built around metadata-conditioned data mixing and reward models rather than a narrowly identity-customization formulation (Wang et al., 5 Jun 2025). EditedID, by contrast, is a training-free post-edit facial restoration framework that aligns latent trajectories, switches solvers, and gates attention to restore identity after a multimodal editor has already produced an edited portrait (Dong et al., 21 Feb 2026). EditIDv2 sits between these poles: it is neither a generalist real-image editor nor a pure post-hoc restorer, but an identity-conditioned text-to-image generator optimized for editable long-prompt customization (Li et al., 6 Sep 2025).

6. Evaluation, empirical trade-offs, and limitations

The experiments are centered on IBench, especially ChineseID with editable long prompts. The base model is the Flux version of PuLID, which is also the base for EditID. Comparative baselines are PuLID (SDXL), PuLID (Flux), InfiniteYou, and EditID. Generation settings include 20 sampling steps, guidance 3.5, cfg_scale 1, and Euler sampling. All experiments run on 4 NVIDIA A100 GPUs, with ComfyUI as the inference framework (Li et al., 6 Sep 2025).

On ChineseID with editable long prompts, EditIDv2 reports Aesthetic 0.691, Image Quality 0.437, Posediv Yaw 18.17, Pitch 9.641, Roll 11.39, Landmarkdiff 0.096, Exprdiv 0.611, Facesim 0.659, ClipI 0.804, and ClipT 0.253. FID is shown as “-” for EditIDv2 in the table. Relative to PuLID Flux, Facesim drops from 0.735 to 0.659, but the paper treats this as a desirable reduction in excessive identity rigidity. At the same time, editability metrics improve substantially: compared with PuLID Flux, EditIDv2 improves Posediv by about 8.872 in Yaw, 3.769 in Pitch, and 1.917 in Roll, totaling more than 14 points across the three Euler-angle measures; Landmarkdiff rises from 0.070 to 0.096; Exprdiv increases from 0.562 to 0.611; ClipI rises from 0.757 to 0.804; and ClipT becomes 0.253 versus 0.243 for PuLID Flux and 0.249 for EditID (Li et al., 6 Sep 2025).

Compared specifically with EditID, EditIDv2 improves Aesthetic from 0.683 to 0.691, ClipI from 0.769 to 0.804, and major editability metrics such as Yaw, Pitch, Roll, Landmarkdiff, and Exprdiv, while accepting a reduction in Facesim from 0.714 to 0.659. The paper interprets this not as a regression in the ordinary sense but as evidence that the method is optimized for long-prompt editable customization rather than maximal static identity fidelity (Li et al., 6 Sep 2025).

The qualitative results reinforce that interpretation. The paper uses long prompts covering modern urban suspense, mythological epic battlefield, rural pastoral memories, and steampunk adventure, and it argues that EditIDv2 better realizes action, scene composition, and multi-layer semantics than EditID, InfiniteYou, and PuLID. The ablations focus on the components the paper claims matter most. Removing the ID loss weakens identity consistency and harms semantic behavior under complex prompts. Using a single set of attention weights struggles to balance consistency and editability under long text input, whereas offline fusion better presents multi-level semantics. The paper repeatedly attributes performance gains to PerceiverAttention decomposition and offline fusion under minimal data lubrication (Li et al., 6 Sep 2025).

The paper also implies several limitations. First, despite the “tuning-free” label, the method requires fine-tuning the cross-attention or PerceiverAttention module on about 3K labeled samples, so its practicality depends on access to curated same-ID multi-attribute data. Second, it intentionally sacrifices some face similarity relative to rigid identity methods such as PuLID Flux. Third, offline fusion adds model-preparation complexity because multiple optimized weight variants must be trained and validated before deployment. Fourth, some implementation details remain underspecified, including the exact $$\mathcal{L}(\theta) = \mathbb{E}_{t, x_0, x_1} \left[ \| f_\theta((1-t)x_0 + t x_1, t, c) - (x_1 - x_0) \|^2 \right].$$5 schedule and the complete cross-attention implementation after decomposition. Overall, EditIDv2 is best understood as a pragmatic middle ground: not full customization retraining, not rigid train-free identity pasting, but editable ID customization with minimal data lubrication for narrative text-to-image generation (Li et al., 6 Sep 2025).