---
title: 'PureCC: Text-to-Image Concept Customization'
url: https://www.emergentmind.com/topics/purecc
type: topic
---

# PureCC: Text-to-Image Concept Customization

Searching arXiv for the PureCC paper and closely related concept-customization work.
arxiv_search.query({"search_query":"ti:\"PureCC\" OR abs:\"PureCC\"","start":0,"max_results":10})
Searching for DreamBooth and related baselines mentioned alongside PureCC.
arxiv_search.query({"search_query":"all:DreamBooth text-to-image personalization", "start":0, "max_results":5})
PureCC is a text-to-image concept customization method designed to “purely” learn a new concept while preserving the original model’s behavior and capabilities. In the formulation reported for flow-matching models, specifically Rectified Flow and SD3.5-M, the method addresses the standard personalization setting in which a model is adapted from a few reference images of a user-specific instance or style and is prompted through an identifier token such as \([V]\). Its central claim is that concept insertion should affect concept-related attributes without unnecessarily changing background, lighting, style, composition, prompt adherence, or the broader generative prior of the pretrained model. To that end, PureCC combines a decoupled learning objective in velocity space, a dual-branch training pipeline with a frozen extractor and a trainable flow model, and an adaptive guidance scale \(\lambda^\star\) that balances customization fidelity and model preservation [2603.07561].

## 1. Problem setting and preservation objective

In PureCC, “text-to-image concept customization” denotes the adaptation of a pretrained T2I model from a small custom set, typically 3–5 images, so that prompts containing a personalized identifier token generate the intended subject or style in arbitrary scenes. The reported setup distinguishes three textual conditions: a base text describing the scene without personalization, a target text containing the personalized insertion such as \([V]\) dog, and a complete text formed by combining the two. The complete text is encoded to a text embedding \(y_{complete}\) by a pretrained encoder \(E(\cdot)\), and the model is fine-tuned so that prompts containing \([V]\) produce images with the personalized concept [2603.07561].

The paper identifies two failures of standard tuning-based customization. The first is disruption of the original model’s behavior. Existing methods learn from the entire custom image-text pair, and with very few examples the model cannot disentangle “what is the concept” from “scene-specific details” such as background, lighting, and pose. The second is degradation of the original model’s capabilities. After fine-tuning for a concept, metrics such as CLIP-T and HPSv2.1 drop, and the model’s “original distribution” drifts toward the narrow distribution of the custom set. In the paper’s terminology, PureCC therefore treats customization fidelity and model preservation as joint optimization targets rather than assuming that improved identity fidelity alone is sufficient [2603.07561].

For a flow-based model with conditional flow matching, the standard concept customization loss is written as
\[
\mathcal{L}_{CC} = \mathbb{E}_{t, x_t} \left\| \bm{v}_t\big(x_t\big) - \bm{v}_t^\theta\big(x_t \mid y_{complete}\big) \right\|_2^2,
\]
with \(x_t = (1-t)x_0 + t x_1\) and \(\bm{v}_t(x_t) = x_1 - x_0\). PureCC argues that this objective is insufficient because it treats the complete text as a monolithic condition and does not explicitly preserve the base model’s conditional behavior [2603.07561].

## 2. Decoupled objective in velocity space

The central methodological move in PureCC is to define the desired customization target as a sum of two terms: an “original conditional prediction” and a target-concept-specific guidance term. The paper writes the core decomposition as
\[
\bm{v}_t^{PureCC} = \bm{v}_t^{original} + \lambda \cdot \bm{v}_t^{target}.
\]
Here, \(\bm{v}_t^{original}\) is the velocity predicted under the base text alone, and \(\bm{v}_t^{target}\) is an implicit guidance vector for the target concept computed from a frozen extractor branch. This explicitly separates what the original model should do for the base prompt from what should be added when the personalized concept is inserted [2603.07561].

The paper makes the analogy to classifier-free guidance explicit. In PureCC, the “original” term is
\[
\bm{v}_t^{original} = \bm{v}_t^{\theta_2}(x_t \mid y_{base}),
\]
while the target-concept term is defined as
\[
\bm{v}_t^{target} = \mathbf{R}(y_{tar}) = \bm{v}_t^{\theta_1}(x_t \mid y_{tar}) - \bm{v}_t^{\theta_1}(x_t \mid \emptyset),
\]
where \(y_{tar} = \{\mathbf{Y}_{tar}^l\}_{l=1}^L\) denotes layer-wise embeddings for the personalized concept. The refined training target is then
\[
\bm{v}_t^{PureCC} =
\underbrace{\bm{v}_t^{\theta_2}(x_t \mid y_{base})}_{\text{original}}
+
\lambda^\star
\underbrace{\big( \bm{v}_t^{\theta_1}(x_t \mid y_{tar}) - \bm{v}_t^{\theta_1}(x_t \mid \emptyset) \big)}_{\text{target}}.
\]

The corresponding loss is
\[
\mathcal{L}_{PureCC} = \mathbb{E}_{t, x_t} \left\| \bm{v}_t^{PureCC} - \bm{v}_t^{\theta_2}(x_t \mid y_{complete}) \right\|_2^2,
\]
and the total optimization objective is
\[
\mathcal{L}_{PCC} = \mathcal{L}_{CC} + \eta \cdot \mathcal{L}_{PureCC}.
\]

This construction is the paper’s primary novelty claim. It does not constrain parameter drift directly; instead, it constrains predictions in velocity space so that complete-text behavior is learned as base-text behavior plus a controlled concept-specific offset. A plausible implication is that the method reframes personalization from “fit the custom set” to “inject only the concept-specific increment,” which is the sense in which the paper uses the term “pure learning” [2603.07561].

## 3. Dual-branch training pipeline

PureCC implements the decoupled objective through a two-stage, dual-branch training pipeline. Both branches use the same base architecture, the SD3.5-M rectified flow transformer, but they play different roles. The first branch is a frozen representation extractor \(\bm{v}_t^{\theta_1}\); the second is a trainable customization model \(\bm{v}_t^{\theta_2}\) [2603.07561].

In Stage 1, the extractor is trained on the custom set using LoRA modules inserted in transformer layers and layer-wise tunable concept embeddings \(\{\mathbf{Y}_{tar}^l\}_{l=1}^L\). The layer-wise textual embedding is written as
\[
\mathbf{Y}^{l}_{complete} = [y_{base}; \mathbf{Y}_{tar}^{l}],
\]
and the extractor is optimized with the standard customization loss
\[
\mathcal{L}^{Rep}_{CC} = \mathbb{E}_{t,x_0,x_1} \left\| (x_1 - x_0) - \bm{v}_t^{\theta_1}(x_t \mid y_{complete}) \right\|_2^2.
\]
Only the LoRA weights and \(\{\mathbf{Y}_{tar}^l\}\) are updated, while the base SD3.5-M backbone remains fixed. After this stage, the extractor and its concept embeddings are frozen [2603.07561].

In Stage 2, a second flow model is initialized from the same pretrained SD3.5-M and trained with \(\mathcal{L}_{PCC}\). The frozen branch receives \(x_t\) with target-only and null conditions and produces
\[
\bm{v}_t^{\theta_1}(x_t \mid y_{tar}), \qquad \bm{v}_t^{\theta_1}(x_t \mid \emptyset),
\]
whose difference defines the target guidance. The trainable branch receives \(x_t\) with both base and complete conditions and produces
\[
\bm{v}_t^{\theta_2}(x_t \mid y_{base}), \qquad \bm{v}_t^{\theta_2}(x_t \mid y_{complete}).
\]
The former is used as the original conditional prediction; the latter is trained to match the decoupled target [2603.07561].

The paper attributes two distinct functions to this architecture. The frozen extractor is intended to provide a “purified” target concept representation that is isolated from incidental scene details, while the trainable branch preserves the original model’s base-prompt behavior and learns to integrate the concept guidance additively. At inference time, only the trainable model \(\bm{v}_t^{\theta_2}\) is used, so inference cost is the same as standard LoRA or flow fine-tuning [2603.07561].

## 4. Adaptive guidance scale and optimization dynamics

PureCC replaces a fixed guidance coefficient with an adaptive scale \(\lambda^\star\). The paper first defines the trainable model’s learned concept representation as
\[
\mathbf{R}(y_{complete}, y_{base}) = \bm{v}_t^{\theta_2}(x_t \mid y_{complete}) - \bm{v}_t^{\theta_2}(x_t \mid y_{base}),
\]
and then chooses \(\lambda^\star\) as the minimizer of
\[
\lambda^\star = \arg\min_\lambda \left\| \mathbf{R}(y_{complete}, y_{base}) - \lambda \cdot \mathbf{R}(y_{tar}) \right\|_2^2.
\]
The resulting closed form is
\[
\lambda^\star =
\frac{\langle \mathbf{R}(y_{complete}, y_{base}), \mathbf{R}(y_{tar}) \rangle}
{\left\| \mathbf{R}(y_{tar}) \right\|^2}.
\]
This is a projection coefficient: it measures how well the trainable model’s concept direction aligns with the frozen extractor’s target direction and adjusts the guidance strength accordingly [2603.07561].

The reported ablation shows why this term matters. For instance/style settings, \(\lambda=1.0\) yields \(\Delta\)CLIP-T (base) of \(-0.18 / -0.67\) and CLIP-I(target)/CSD of \(0.43 / 0.26\); \(\lambda=3.0\) yields \(-0.51 / -0.93\) and \(0.58 / 0.61\); \(\lambda=5.0\) yields \(-2.67 / -4.21\) and \(0.73 / 0.42\); and \(\lambda=\lambda^\star\) yields **\(-0.31 / -0.26\)** and **\(0.81 / 0.63\)**. In the paper’s reading, fixed small \(\lambda\) preserves the base model but weakens concept fidelity, fixed large \(\lambda\) improves concept expression but harms prompt adherence, and adaptive \(\lambda^\star\) is the mechanism that balances preservation and responsiveness [2603.07561].

Optimization details are correspondingly specific. The reported backbone is SD 3.5-M Rectified Flow, with LoRA rank 4, learning rate \(1.0\times10^{-4}\) for both flow model and layer-wise concept embeddings, NVIDIA A100 hardware in BF16, batch size 2, 400 steps for Stage 1, 400 steps for Stage 2, and 28 timesteps for inference with the default SD3.5-M sampler [2603.07561].

## 5. Evaluation protocol and empirical performance

The evaluation uses both qualitative and quantitative benchmarks. The custom concepts consist of 14 personalized concepts from the DreamBooth dataset and 16 additional concepts collected by the authors, including 11 instance concepts and 5 style concepts, for a total of 30 concepts in qualitative evaluation. The quantitative benchmark, DreamBenchPCC, extends DreamBench with 12 additional style concepts and is described as balanced between instance and style concepts. Reference images per concept number 3–5, and new images are captioned with Claude 3.5 Sonnet for textual labels [2603.07561].

PureCC evaluates both concept fidelity and preservation. For instance concepts, fidelity is measured by CLIP-I (target) and DINO. For style concepts, fidelity is measured by CSD. Preservation is tracked through differential metrics
\[
\Delta M = M_{\mathrm{custom}(I(y_{\mathrm{complete}}))} - M_{\mathrm{original}(I(y_{\mathrm{base}}))},
\]
where \(M\) is CLIP-T, HPSv2.1, or PickScore. Smaller magnitude, values closer to 0, or positive values mean better preservation. Behavior-level preservation is measured by Seg-Cons, which uses SAM segmentations and compares original-model outputs conditioned on base text with customized-model outputs conditioned on complete text [2603.07561].

On DreamBenchPCC, the reported instance-preservation results are: DreamBooth with \(\Delta\)CLIP-T \(-4.81\), \(\Delta\)HPSv2.1 \(-2.17\), and Seg-Cons \(18.38\); CIFC with \(\Delta\)CLIP-T \(-1.93\), \(\Delta\)HPSv2.1 \(-1.62\), and Seg-Cons \(13.23\); and PureCC with **\(\Delta\)CLIP-T \(-0.31\)**, **\(\Delta\)HPSv2.1 \(+0.10\)**, \(\Delta\)PickScore \(-0.67\), and **Seg-Cons \(69.37\)**. For instance concept responsiveness, DreamBooth reports CLIP-I(target) \(0.63\) and DINO \(0.62\), CIFC reports \(0.78\) and \(0.65\), and PureCC reports **\(0.81\)** and **\(0.73\)**. For style preservation, DreamBooth reports \(\Delta\)CLIP-T \(-6.23\), CIFC reports \(-4.70\), and PureCC reports **\(-0.26\)** with \(\Delta\)HPSv2.1 \(-0.92\). For style responsiveness, DreamBooth reports CSD \(0.57\), CIFC \(0.64\), and PureCC \(0.63\), which the paper describes as competitive [2603.07561].

The ablation on pure learning further isolates the effect of the method. Using only \(\mathcal{L}_{CC}\) gives \(\Delta\)CLIP-T \(-4.52\), \(\Delta\)HPSv2.1 \(-2.01\), Seg-Cons \(23.74\), and CLIP-I(target) \(0.65\). A merged training stage without a pretrained extractor gives \(\Delta\)CLIP-T \(-1.17\), \(\Delta\)HPSv2.1 \(-0.34\), Seg-Cons \(54.37\), and CLIP-I(target) \(0.50\). Full PureCC gives **\(\Delta\)CLIP-T \(-0.31\)**, **\(\Delta\)HPSv2.1 \(+0.10\)**, **Seg-Cons \(69.37\)**, and **CLIP-I(target) \(0.81\)**. The paper interprets this as evidence that \(\mathcal{L}_{PureCC}\) is crucial for preservation and that two-stage training is necessary for strong concept fidelity [2603.07561].

## 6. Position in the literature, practical considerations, and limitations

PureCC is positioned against several families of personalization methods. Relative to direct fine-tuning methods such as DreamBooth, it uses limited parameter updates through LoRA together with a preservation-aware objective. Relative to embedding-based methods such as Textual Inversion, it still modifies model parameters, but it supplements concept embeddings with a frozen extractor and velocity-space supervision. Relative to EWC-style and continual-learning regularizers, it does not operate primarily at the parameter level; instead, it decouples the velocity field into an original term and a concept-guidance term. The paper presents the decoupled velocity-space objective, the dual-branch architecture, and the adaptive \(\lambda^\star\) as its three claimed novelties [2603.07561].

The reported qualitative findings are consistent with that positioning. In single-concept generation, PureCC is described as changing only the target concept while keeping background, composition, and lighting similar to what the base model would produce for the base text. In multi-concept composition, the paper reports that baselines show cross-concept interference, whereas PureCC maintains each concept’s identity and integrates them coherently. In instance-plus-style settings, it is reported to better balance style transfer and content preservation. The user study is summarized as showing that participants overwhelmingly prefer PureCC for original behavior consistency, approximately \(95\)–\(99\%\) versus baselines, as well as for aesthetic quality and target fidelity [2603.07561].

Practical overhead is modest but nonzero. The supplementary material reports Stage 1 memory of 28 GB and time of 0.13 A100 hours per concept, Stage 2 memory of 30 GB and time of 0.20 A100 hours, for a total of approximately 0.33 A100 hours per concept. Comparative costs are reported as DreamBooth \(0.25\)h / \(4.46\)s, LoRA \(0.13\)h / \(4.72\)s, Mix-of-Show \(0.23\)h / \(5.80\)s, CIFC \(0.28\)h / \(5.48\)s, and PureCC \(0.33\)h / \(4.72\)s. This suggests that the method adds training complexity through two stages and dual-branch optimization, while keeping inference cost comparable to ordinary LoRA-based personalization [2603.07561].

The paper also states clear constraints. Hyperparameters such as \(\eta\) and, when fixed, \(\lambda\), affect the trade-off between preservation and fidelity; the supplementary ablation reports that \(\eta=1.0\) is the best trade-off, whereas \(\eta=0.5\) worsens preservation and lowers concept fidelity and \(\eta>1.5\) increases artifacts or over-injection of concept. The method is conceptually model-agnostic but is evaluated on the SD3.5-M rectified flow model rather than across all diffusion and flow backbones. Extremely complex or ambiguous concepts remain difficult to isolate in the representation extractor, and the additional training stage is more complex than simple LoRA or DreamBooth fine-tuning [2603.07561].

Source: https://www.emergentmind.com/topics/purecc