Feature request: Codex-managed private Style Profiles for GPT Image 2 (LoRA-like adapters)

[xiuyidesign]

Feature request: Codex-managed private Style Profiles for GPT Image 2 (LoRA-like adapters)

Executive summary

Please add a Codex-managed, private style-adaptation workflow for GPT Image 2. The user-facing product should be a versioned Style Profile trained from a small, curated set of authorized images. The implementation could use LoRA or an OpenAI-native adapter; users would not need access to the base model or raw adapter weights.

This would combine the strongest part of open image-model ecosystems - repeatable customization - with GPT Image's quality, instruction following, editing workflow, safety systems, and ease of use.

GPT Image 2 currently does not support fine-tuning. Reference-image workflows are useful and high fidelity, but the documentation also acknowledges that recurring characters and brand elements may still vary across generations. A managed Style Profile would address the gap between prompt-time reference and reusable learned adaptation.

Official context:

• GPT Image 2 model page
• OpenAI image generation guide

The user problem

Text-only skills encode instructions, not visual features. Even a carefully built multimodal skill can still drift between generations.

One real workflow uses:

• a detailed visual grammar describing line weight, directional hatching, tonal hierarchy, negative space, composition, and prohibited artifacts;
• five curated style references, with two to four references supplied per generation;
• explicit preservation rules for geometry, perspective, proportions, and identifying details;
• a review-and-retry rubric.

This is substantially better than a plain prompt, but it is still inference-time conditioning. It does not create a persistent representation of the style. Users repeatedly spend time selecting references, rewriting prompts, rejecting drifted outputs, and re-establishing the same visual language.

Open-weight image ecosystems solve this class of problem with LoRA and related adapters. Creators who need repeatable style identity therefore have a strong reason to leave an otherwise simpler GPT Image workflow.

Proposed product experience

1. Create a private Style Profile in Codex

The user selects an authorized set of example images and chooses Create Style Profile. Codex helps with:

• duplicate and near-duplicate detection;
• image-quality and diversity checks;
• automatic captions and editable style descriptions;
• separation of style references from subject-identity references;
• rights attestation and privacy settings;
• a cost and training-time estimate.

2. Train a managed adapter

OpenAI trains a private, model-compatible adapter. LoRA is one possible implementation, but the public contract should be outcome-based so OpenAI can use a safer or more effective internal method.

The profile should expose clear states such as validating, training, ready, failed, and deleted, plus immutable versions for reproducible work.

3. Reuse it across Codex image workflows

Users select a Style Profile in the Codex image-generation UI or reference it in an image tool call. An illustrative interface could be:

{
  "style_profile": {
    "id": "sp_example",
    "version": 1,
    "strength": 0.75
  }
}

This is a proposed interface, not a claim about the current API. The profile should work with generation, editing, multi-turn refinement, and batch exploration.

4. Evaluate, version, and delete

Codex generates a fixed evaluation set before activation. Users compare versions, promote or roll back a version, and permanently delete the profile and its training assets.

Safety, ownership, and privacy

A managed service can be safer than arbitrary third-party LoRA distribution:

• require explicit rights attestation for training images;
• keep profiles private to the user or workspace by default;
• do not use submitted images or adapters to improve the base model without separate opt-in consent;
• preserve existing image safety checks during training and inference;
• provide workspace access controls, audit events, retention settings, and deletion;
• detect attempts to imitate protected living artists or misuse real-person identity;
• do not expose adapter downloads in the MVP.

Suggested MVP

Launch an opt-in limited beta with:

• private single-style profiles;
• a guided dataset-preparation flow in Codex;
• managed training and version status;
• use in GPT Image 2 generation and editing;
• one adjustable style-strength control;
• evaluation previews, rollback, and deletion;
• no raw-weight export or adapter composition.

Later versions could add organization libraries, approved brand profiles, subject adapters, controlled adapter composition, API management, and shared evaluation suites.

Business and ecosystem impact

This feature would expand the audience for both Codex and GPT Image among:

• industrial, automotive, fashion, and graphic designers;
• game, film, animation, and publishing teams;
• brand and advertising studios;
• independent creators who currently maintain local GPU workflows;
• enterprises that need consistent visual identity with governed data handling.

It would also create a natural paid workflow: dataset preparation in Codex, adapter training, profile storage, evaluation, and repeated image generation. More importantly, it would turn Codex skills from prompt libraries into repeatable production systems without requiring users to operate an open-model training stack.

Success metrics

• fewer rejected generations and manual prompt revisions;
• higher style-adherence and cross-generation consistency ratings;
• lower time to produce an approved asset set;
• repeat use of the same profile across projects;
• retention of creators who would otherwise move to local LoRA workflows;
• safe deletion and low rates of ownership or impersonation disputes.

Evidence plan

An authorized comparison board will use the same source content, prompt, model, output settings, and evaluation rubric across:

1. text-only style instructions;
2. the current detailed skill plus multiple reference images;
3. the expected acceptance criteria for a trained Style Profile.

Only images owned by the submitter or explicitly licensed for public use will be posted. Until those images are confirmed, this proposal intentionally contains no public image attachments.

---

中文摘要

建议为 GPT Image 2 增加由 Codex 管理的私有“风格配置文件”。用户提供一组拥有合法使用权的参考图，Codex 负责样本检查、去重、描述、训练、版本管理、评估、调用和删除。底层可以采用 LoRA，也可以采用 OpenAI 自研适配器；产品不必开放基础模型或原始适配器权重。

目前，纯文字 skill 只能保存提示规则，多参考图也只是推理时条件控制，无法学习一个可跨任务复用的视觉表示。实际工作中，即使已经建立详细的风格语法、参考图库和质量检查流程，线条、材质、构图和品牌元素仍会在多次生成中漂移。

该功能可以把开源图像生态的可定制优势，与 GPT Image 的生成质量、编辑能力、安全机制和易用性结合起来，吸引设计师、品牌团队、影视游戏工作室和依赖本地 LoRA 工作流的创作者使用 Codex。