PureCC: Achieving "Pure" Personalization Without Damaging the Base Model

TL;DR

Concept customization in Text-to-Image (T2I) models usually comes at a cost: the better the model learns your dog "[V]", the worse it becomes at following general prompts or maintaining artistic backgrounds. PureCC (Pure Concept Customization) changes this trade-off. By decoupling the learning objective into "original prediction" and "concept guidance," and using a clever adaptive scale, it inserts new subjects into the model while keeping the model's foundational "DNA" intact.

The "Concept Pollution" Problem

Most weight-tuning methods like DreamBooth or LoRA treat the 3-5 reference images you provide as the new ground truth. The model tries so hard to minimize the loss on these few images that it "forgets" how to render anything else. If your reference photo has a specific lighting or background, the model erroneously bakes those features into the identifier [V].

As shown in the paper's motivation, this leads to:

1. Behavior Disruption: Changing the background/lighting when you only asked to change the subject.
2. Capability Degradation: A sharp drop in prompt adherence and aesthetic quality (distribution shift).

Methodology: The Dual-Branch Architecture

The core philosophy of PureCC is decoupling. Instead of merging everything into one weight update, PureCC splits the task.

1. The Representation Extractor (Stage 1)

First, the authors train a "specialist" model (using LoRA and layer-wise embeddings) to deeply understand the target concept. This model is then frozen to serve as a pure source of "what the target looks like."

2. Pure Learning Pipeline (Stage 2)

In this stage, a second, trainable model is used. The training objective is redefined: $v_t^{PureCC}=v_t^{original}+\lambda \cdot v_t^{target}$ Where:

• $v_t^{original}$ is what the base model would have predicted (preserving behavior).
• $v_t^{target}$ is the "bias" or delta provided by the frozen extractor (the concept injection).

3. The Adaptive Scale ${\lambda }^{\ast }$

Choosing a fixed $\lambda$ is risky—too high and you destroy the base model; too low and the subject doesn't look like the target. PureCC introduces an Adaptive Guidance Scale that uses a closed-form mathematical solution to dynamically find the optimal strength of guidance at every training step.

Experimental Excellence

The results are visually and quantitatively striking. In multi-concept scenarios, PureCC avoids "semantic entanglement" (e.g., color bleeding between two different customized objects) that plagues methods like Mix-of-Show.

Quantitatively, PureCC achieves a Seg-Cons (Segmentation Consistency) score of 69.37, whereas previous SOTA methods often hovered below 30.0. This means PureCC is effectively "the same model" in areas where the custom concept isn't present.

Conclusion and Deep Insight

The brilliance of PureCC lies in its transition from "re-training" to "guided prediction" within the training loop. By treating the identity of a concept as an additive vector in the velocity field of a flow-matching model, the authors have found a way to "patch" models without corrupting the underlying software.

Limitations: The primary drawback is the requirement of two training stages and slightly higher VRAM usage during the fine-tuning phase. However, since the inference cost remains identical to a standard model, it is a small price to pay for such high-fidelity results.

Reference: Zhichao Liao et al., "PureCC: Pure Learning for Text-to-Image Concept Customization", 2024.