Nudity Detection & Safe Model Routing · Open Skill 公开技能

Nudity Detection & Safe Model Routing

Nudity Detection & Safe Model Routing 是 Digen 上的公开 Open Skill。打开后即可了解该工作流能制作什么。

@Ragazul

Custom

86 3 12

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/1856918/2026-09-03/gyj6ptuh.png

# Digen Skill – Nudity Detection & Safe Model Routing

Safety-first model routing: before any generation, editing, animation, or processing of an uploaded image, first run a visual content check, then pick an officially permitted channel. The purpose of this skill is model routing and compatibility checking — NOT censorship, redesign, or creative reinterpretation of the uploaded person.

## When to use
- The user uploads an image or a reference image and asks to generate, edit, animate, or process it (image-to-image, image-to-video, reference-to-video, etc.).
- A task where the source content might contain nudity or adult material and the model channel must be chosen accordingly.
- Any workflow where the user expects a safe, compliant model to be selected automatically before generation.

## Steps

### 1. Content classification (mandatory, before model selection)
1. First call `recognize_image` on the uploaded image with a classification prompt, asking the recognizer to return exactly one of these five categories plus a short basis:
   1. **No nudity** – fully clothed, non-explicit content.
   2. **Non-sexual adult nudity** – artistic nude photography, bathing, changing clothes, non-sexual full or partial nudity, body-reference imagery, fine-art nudity. The person is clearly an adult; the content is not sexual.
   3. **Artistic / implied adult nudity** – implied, suggestive, artistic partial nudity where the person is clearly an adult.
   4. **Explicit sexual content** – explicit sexual acts or explicit sexual poses/genital detail.
   5. **Age unclear or potentially under 18** – age cannot be confidently determined as adult, or the person appears to be / could be a minor.
2. Do not begin generation before this classification is completed.
3. If the result is borderline between two categories, route as the MORE restrictive category (safer choice).

### 2. Age safety gate (highest priority)
- If the category is **5 (age unclear / potentially under 18)**: stop immediately. Do NOT route the image to any nudity-capable or adult-content model. Never generate, enhance, or sexualize nude or sexualized content involving minors or age-ambiguous persons. Ask the user for a different, clearly adult and permitted source image, or abandon the task.
- Any hint that the person is a minor leads to the same stop.

### 3. Model routing by category
Platform channel catalog (documented, official capability):
- **Images**: bare `t2i` (text-to-image) and `i2i` (image-to-image) support adult / higher-intensity content. All `.hd` / `.hd.lite` image channels (`t2i.hd`, `i2i.hd`, `i2i.hd.lite`) do **NOT** support NSFW — never route adult content to them.
- **Video**: `i2v.nsfw` is the preference channel for adult / higher-intensity image-to-video. Bare `i2v` is for regular content. `ref2v` is for reference-to-video and `t2v` for pure text-to-video — use only when their officially supported content capability matches the detected category.

Routing table:
| Detected category | Image channels | Video channels |
|---|---|---|
| 1. No nudity | Default: `t2i` / `i2i` (any channel suited to the task) | Default: `i2v` / `ref2v` / `t2v` |
| 2. Non-sexual adult nudity | Only officially permitted channels: bare `t2i` / `i2i`. Never `.hd` variants. | `i2v.nsfw` (or another channel confirmed via `list_models` to support adult non-sexual nudity) |
| 3. Artistic / implied adult nudity | Bare `t2i` / `i2i` — a channel explicitly supporting the category. | `i2v.nsfw` (or another channel confirmed to support it) |
| 4. Explicit sexual content | Only channels that officially allow explicit content. If none exists, **stop the task** and clearly report that no compatible model is available. | Only channels that officially allow explicit content. If none exists, **stop and report**. |

- For category 2, do **not** automatically censor, cover, redesign, or alter the uploaded person unless the user requests it — preserve the uploaded image faithfully.
- For category 3, mainta