Image Bot · öffentlicher Open Skill

Image Bot

**Image Bot · Text to Picture to Spoken-Free Stills · t2i / i2i Native · Locked Face, Style & Anatomy** *The agent skill for makers who want a finished **photograph or illustration**—not a video, a slogan dump, a restyled identity, or a leftover model carried onto a new picture* **Image Bot** is a production-native Agent Skill for Digen. It orchestrates a complete **text-to-image (t2i) + image-to-image (i2i)** pipeline for one still: **Plain ask → matching designer → Who / Where / To what / Keep → native-size generate → silent self-check → one delivered picture** No video. No voiceover. No music. No copywriting. No small talk. No storyboard grids. No asking ratio before the first cut. Generate leaves pixel size to the model unless you name one. **2K / 4K** goes through **upscale**, not the `.hd` model channel. Speech is not a track — this skill makes **stills**. Identity, whole-figure art style, and anatomy stay locked unless you asked to change them. A background you named is followed. A new picture with no background gets one matched to the subject. An existing picture keeps its background until you ask to change it. ## What you get | For creators | For the production stack | | --- | --- | | **Beginner-first** — a sentence or an upload is enough; the agent auto-matches ratio and runs to the picture without stopping | **Opinionated pipeline** — t2i or i2i, never mixed; five ratios; one locked model; English prompt from a short edit plan | | **Still-first** — what is in the frame, not a mute mood board or a clip | **t2i / i2i-first** — no accidental video / TTS / mix_clips / Director chains | | **Identity of people & style** — face, kit, and whole-figure drawing lock for the whole edit | **Source binding** — first edit uses the upload; accepted tweaks use the last result; unhappy redo uses the **image that was being edited**, never the rejected output | | **Living edit grammar** — clothes, pose, background, lighting, camera, combine, crop, zoom, outpaint, upscale in one coherent still | **Dense prompts** — Who / Where / To what / Keep + Anatomy + Art style, phrased from the model file for this call | | **Model that stays — until the picture changes** — pick **General image**, **nanobanana**, or **GPT image**; later edits of that picture reuse the channel | **Channel lock** — Family map same row only (`t2i` ↔ `i2i.hd.lite`, `t2i.hd` ↔ `i2i.hd`, `t2i.hd.gpt` ↔ `i2i.hd.gpt`). An unrelated new text picture returns to plain `t2i` unless you name a model in that message | | **Delivery that continues** — after the picture, buttons are **Adjust ratio** then **Upscale** | **Progressive disclosure** — craft lives in `references/prompts/` and `references/edit-prompts.md`, read only for this call | | **Plain maker voice** — one short “I’m making it” line, then the picture; no pipeline jargon | **Marketplace-ready** — clear triggers, defaults, and tool contracts | ## Pipeline | Stage | Output | | --- | --- | | **Intake** | Lock reply language. Auto-match **orientation + aspect_ratio**. Never ask ratio before generate | | **Mode** | No source → **t2i** (omit `input_urls`). Upload or last result → **i2i** (required `input_urls`). Images only → ask intent and wait | | **Designer** | Become the matching professional — stylist, poster designer, retoucher, product photographer, illustrator — without announcing the role | | **Plan** | Who / Where / To what / Keep. Unclear person or region → ask once. Plan is not dumped to you | | **Look** | English prompt from the plan, using this call’s model file. People → anatomy + whole-figure style locks. On-image words confirmed, never invented | | **Model** | Default **General image**: plain `t2i`, or `i2i.hd.lite` when editing. Poster-like with words and no pick yet → buttons **General image / nanobanana / GPT image**. More than 2 uploads → `i2i.hd.lite` unless a model is already chosen | | **Generate** | One `generate_image`. No pixel cap unless you named a size. t2i → `generated_images`. i2i → `edited_images` | | **Self-check** | `recognize_image` vs the plan. One silent fix, same tier, then deliver. Never described to you | | **High-res** | Only if you asked 2K / 4K / sharper / upscale → `upscale_image` after a pass (`scale` 2 or 4) | | **Sequel** | Tweak the last accepted picture, or redo from the source being edited. Same ratio and model unless you change them. A new unrelated text picture starts again on plain `t2i` | ## Production stack | Parameter | Default | Notes | | --- | --- | --- | | **Output** | **One still** | Photo, illustration, portrait, scene, product, poster, pictorial, cover, flyer, banner, avatar. Overflow edits are follow-ups, not a video | | **Engine** | `generate_image` · t2i or i2i | Family does not change tier. Forbidden: video, audio, documents, mixing `input_urls` with `t2i*` | | **Generate resolution** | **Native** | Omit `resolution`, `width`, and `height` unless you named a size. Never treat 2K / 4K as the `.hd` channel | | **Image orientation** | **Auto-match** | Five pairs only: 16:9 · 9:16 · 4:3 · 3:4 · 1:1. Later calls reuse the last pair. Change only via **Adjust ratio** | | **Image refs** | User uploads preferred | First edit → that upload. Accepted tweak → last successful image (+ original as identity ref if the face must stay). Unhappy → the pre-edit source, not the rejected result | | **Ref order** | Base first, then extra refs | Prompt names each source. Unclear person / clothing / scene / style → ask once. 5+ dumps → ask which 2–4 matter | | **Models** | **General image** | t2i `t2i` (Krea 2). i2i `i2i.hd.lite` (Nano Banana). Buttons: `General image`, `nanobanana`, `GPT image`. Aliases banana / GPT still map. Adult NSFW → plain `t2i` / plain `i2i` only, no `.hd*` | | **Raise ladder** | **Only when allowed** | t2i “make it better” → i2i **same tier** (`t2i` → `i2i.hd.lite`). i2i “make it better” → `i2i.hd.gpt`. Unhappy i2i → `i2i.hd.gpt`. Unhappy t2i → `.hd` first, then `.hd.gpt`. GPT image never raises, and does not drop except on an unrelated new text picture | | **Family map** | **Same row** | `t2i` ↔ `i2i.hd.lite`, `*.hd` ↔ `*.hd`, `*.hd.gpt` ↔ `*.hd.gpt`. Raising is the ladder, not the map. Plain `i2i` only when named or this call is adult NSFW | | **New unrelated picture** | **Plain `t2i`** | Drops the previous lock and `user-chose`. A model named in that same message wins | | **Background** | **Follow, else match or keep** | Named background is used. Text-to-image with none → match one to subject, action, and mood. An existing picture keeps its background until you ask to change it | | **Designer lock** | **Matching craft** | Vague asks get concrete cut / color / lighting / layout inside the existing change. Do not override what you specified. Do not restyle unmentioned regions | | **Identity** | **Face stays** | Face-swap only with a face ref **and** an explicit ask. Same-person face edits change only the named features. Product / logo un-morphed except what you asked | | **Art style** | **Whole figure** | Each character keeps their source style on face, hands, hair, clothes, skin. Clothes / pose / background are not a style change | | **Anatomy** | **Anatomically sound** | Two arms, two hands, five fingers. No extra or fused limbs. High-risk on pose / combine / camera | | **On-image text** | **Confirmed only** | Exact spelling, language, placement. No invented slogans. Poster-like + no model yet → pause for the three buttons. GPT image with text adds the lettering check | | **Self-check** | **On** | After every generate except upscale-only. One silent fix, same tier, then deliver. Do not loop | | **Talk** | **One line** | `Okay, I'm making it for you now.` (or the same meaning in your language). No tool names, channels, URLs, or a `Locked:` line | | **Confirm** | **After delivery** | First picture auto-runs when intent is clear. Pause only for the ask-once list: intent, target, contradiction, face ref, text, age / body / major hair, crop, upload roles, poster, or Adjust ratio | | **Guidance** | **Adjust ratio → Upscale** | Only these two, in this order. Dissatisfaction does not change the buttons | | **Series** | **Same-thread edits** | Continue tweaking the last accepted still, or undo to the original upload | | **Language** | **User-typed lock** | Replies in your language; generate prompt is English | ## Modes | Mode | You bring | Agent delivers | | --- | --- | --- | | **Full auto-run** | Nothing — or “make a picture of …” | Scene from your words → auto-match ratio → native-size plain `t2i` → self-check → still. No named background → one matched to the subject | | **Idea-first** | One-line premise (“editorial portrait”, “product on marble”) | Matching designer fills the look → full pipeline | | **Script-first** | Verbatim on-image words | Confirm spelling / language / placement → poster ask if needed → generate with exact copy | | **Reference-driven** | Photos + “change the clothes / background / pose” | Bind uploads; do not redraw what you did not ask to change | | **You in the picture** | Face or full-body ref + “use this person” | Keep identity; optional clothing / scene / style refs named in the prompt | | **Combine refs** | Person + outfit + room (or more) | Ask roles if unclear; more than 2 uploads → `i2i.hd.lite` unless you already named a model | | **Named model** | “nanobanana” / “GPT image” / “General image” | Lock that channel (`user-chose`). GPT image stays on GPT for later redos of that picture | | **Unrelated new picture** | A new text subject, not a tweak of the last still | Plain `t2i`. Previous model lock is dropped unless you name a model in this message | | **Make it better** | “Better quality”, same subject | t2i → i2i the last result, **same tier**. i2i → `i2i.hd.gpt` (stay if already GPT image) | | **Unhappy redo** | “Not good” / redo / reject | Re-edit the **image that was being edited**, not the failed result. i2i → `i2i.hd.gpt`. t2i → `.hd` first, then GPT image. GPT lock → same GPT channel only | | **Undo** | “Go back to the original” | Original upload, or previous t2i if there was no upload | | **Adjust ratio** | The post-success button | Five labels, wait, i2i the last still into the new pair | | **Upscale** | The post-success button, or “2K / 4K / sharper” | `upscale_image` on the checked or last successful picture. Do not regenerate | | **Crop / zoom** | “Keep only …” / “enlarge the face in frame” | i2i generate with no pixel cap (not upscale). High-res after only if you also asked | ## Example prompts **Full auto-run** > Make a cinematic portrait of a woman in a linen suit standing by a rainy window. **Script-first** > Magazine cover. Headline exactly "NIGHT SHIFT". Keep my uploaded face. Portrait 3:4. **Reference-driven** > This photo is the person. This photo is the jacket. Put the jacket on them. Do not change the face or the drawing style. **You in the picture** > Use this selfie. Change the background to a quiet bookstore. Keep my face and clothes. **Named model** > Use GPT image. Poster, flyer layout, the line exactly "SUMMER SALE" at the top. **Unrelated new picture** > Forget that poster. Draw a red bicycle on an empty street. **Make it better** > Make this a bit better — same person, same scene. **Unhappy redo** > This isn’t it. Redo the clothes change. **Series** > Keep this picture. Only change the hair to a short bob. **High-res** > Upscale this to 2K. Don’t redesign it. ## When to use - New stills from text: photo, illustration, portrait, scene, product shot, poster, pictorial, magazine cover, flyer, banner, avatar - Uploaded images that need clothes, decorations, recolor, pose, background, retouch, beautify, remove, lighting, camera, hair, makeup, face, combine, outpaint, enlarge, crop, zoom, or upscale - Uploads with no edit text yet (the agent asks intent first) - Follow-ups on this skill’s last picture, including unhappy redo or undo - A new text picture that is not the last one (it starts again on plain `t2i`) - Anyone who named **nanobanana** or **GPT image** - Adult NSFW stills, which generate only on plain `t2i` or plain `i2i` > **Not the right tool when** you need video, voiceover, music, copywriting, scripts, research, or small talk; when you only ask what the picture is or whether it looks good; when the source is a library asset, pasted URL, or canvas image that is not your upload and not this skill’s last result (plain-text t2i is still allowed); when the request is sexual content involving minors or CSAM; or when you want a storyboard grid or a six-clip stitch — this skill does not make those. ## Invoke ``` /image_edit ``` Or: *"Turn this into a still — generate or edit, face locked, style locked, one picture, then Adjust ratio or Upscale."* **Your brief · Your refs · One coherent still.**

@official

37714 21 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/dev/admin/199/2026-10-09/yvrahgc6.gif

Digen teilen

Image Bot · öffentlicher Open Skill

Image Bot