| name | ai-image-recreation |
| description | AI recreation editing style transfer enhancement and variation of user-uploaded images via Grok Imagine image_edit and image_gen. Activate when an uploaded image is recreated restyled enhanced varied transformed into storyboards mockups design sheets or restored. Grok 4.5 orchestration with multi-pass refinement Identity Lock handoff and pre-video plate prep for SuperGrokPro and Cinematic Studio. |
AI Image Recreation v3.7.1 (Grok 4.5 · Imagine Image)
Scope: User-uploaded (or path-provided) reference images → faithful recreation, style transfer, enhancement, restoration, variation, and production-ready plates.
Sibling skills: generated-image-editor (session Grok outputs by ID/path) · i2i-cinematic-refiner / i2i-refiner (multi-pass cinematic / explicit polish) · imagine (core tool craft) · character-dna-extractor (identity lock from refs)
Model Layer (Grok 4.5 · studio v3.7.1)
| Layer | Slug | When |
|---|
| Orchestration (default) | grok-4.5 | Intent parse, multi-pass planning, QA, handoffs |
| Long-context (opt-in) | grok-4.3 | Huge multi-ref series / long memory banks only |
| Grok Build CLI | grok-4.5 · grok-build | Skills / coding (≥ 0.2.93) |
| Imagine Video | grok-imagine-video / 1.5 | 1.0 cost · 1.5 native audio |
| Imagine Image | grok-imagine-image / quality | Stills / hero plates |
Prefer stable prompt_cache_key (project slug) on multi-turn grok-4.5 loops. Reasoning high for identity-critical multi-pass and DNA handoffs; medium for routine style transfers; low only for trivial routing. Image spend is image_edit / image_gen (and video tools), never chat models as generators. Registry: tools/models.py · references/agents/MODEL_LAYER_v3.7.1.md · models verify.
When to Activate
- Uploaded or path-referenced image + recreate / edit / restyle / enhance / vary / transform / restore
- Cinematic storyboard frames, product mockups, character design sheets, environmental immersion
- Archival / low-res restoration and commercial retouch
- Pre-video hero plates from user photos before
image_to_video or Studio pipeline
- User says:
ACTIVATE AI_IMAGE_RECREATION, RECREATE THIS IMAGE, STYLE TRANSFER, ENHANCE THIS PHOTO, VARIATIONS FROM UPLOAD
When NOT to Use
| Situation | Use instead |
|---|
| Pure text-to-image, no reference | image_gen + imagine skill |
| Iterate on Grok-generated session assets | generated-image-editor |
| Multi-pass cinematic polish with strength curves | i2i-cinematic-refiner (SFW) or i2i-refiner (explicit) |
| Forensic DNA + long-sequence identity | character-dna-extractor → Identity Lock |
| Basic crop/resize/mux only | ffmpeg / cinematic-ffmpeg |
| Full multi-agent film production | grok-imagine-cinematic-studio / Studio Director |
Activation
ACTIVATE AI_IMAGE_RECREATION
Begin work with: "Initiating AI Image Recreation Protocol v3.7.1 (Grok 4.5)…"
Tool Map (Grok Build Imagine)
| Goal | Tool | Notes |
|---|
| Recreate / restyle / enhance from upload | image_edit | Primary path. Pass absolute path, attachment token, or data:image/...;base64,... |
| New composition loosely inspired by description only | image_gen | Only when pixel reference is not required |
| Multiple variations | Multiple parallel image_edit calls | No n/count param — one call per variation |
| Still → motion | image_to_video | After plate is approved; prefer 6s shots |
| Multi-ref motion | reference_to_video | When user needs multi-image style/content blend |
| Named real person | image_edit + real reference | Never pure image_gen for real likeness |
Rules aligned with imagine skill:
- Prefer
image_edit whenever a source image exists.
- For real people: reference-first; never non-consensual sexualized or minor-involving likenesses.
- Describe what to change and what must stay; front-load subject and intent.
- On moderation block: stop, do not paraphrase to evade; offer a different direction.
- Save outputs under
artifacts/ (or session artifacts path) with descriptive names.
image_edit contract
- prompt — desired outcome + fidelity anchors (2–6 sentences; natural prose)
- image — one or more refs (prefer one clean primary for reliability)
- aspect_ratio — multi-image edits / reframe; single-image edits often preserve source AR
image_gen contract
- prompt — full scene description
- aspect_ratio — e.g.
1:1, 16:9, 9:16, 4:3, 3:4, auto
Core Principles
- User intent first — faithful copy, style transfer, targeted edit, enhancement, or exploratory variations.
- Fidelity anchors — composition, subjects, poses, camera angle, lighting direction, identity markers unless the user wants change.
- Multi-pass by default for “pro / hero / delivery” — do not burn quota on one-shot hopes for hard jobs.
- Consistency tokens — for series work, lock face/body/style phrases; escalate to DNA Extractor when characters will recur in video.
- Artifact hygiene — descriptive filenames, paths reported, ready for DNA / i2i / i2v handoff.
- Safety — follow platform safety; adult content only when clearly requested by adults; escalate cinematic NSFW sequences to ErosForge.
Step-by-Step Workflow
1. Identify the source
Confirm accessibility:
- Attachment / render token (e.g. session image reference)
- Absolute path (
…/artifacts/upload.jpg, /home/workdir/artifacts/…)
- Prior recreation path for chained passes
If unclear, ask once for the path or re-upload.
2. Classify the job
| Mode | Intent signals | Default tool plan |
|---|
| Faithful | “exact”, “same”, “restore”, “upscale detail” | 1–2× image_edit fidelity + polish |
| Style transfer | “as Ghibli / oil / cyberpunk / …” | Fidelity anchors + style layer; optional 2nd polish |
| Targeted edit | change bg, outfit, expression, remove object | Describe only changes + preserve list |
| Enhance / restore | low-res, old photo, noise | Restoration templates; light second pass if needed |
| Variations | “3 looks”, “outfit pack” | N parallel image_edit with shared anchors |
| Env immersion | place subject in new world | Subject lock + environment rewrite |
| Design sheet | turnaround, expressions | Multi-call grid of views from same DNA anchors |
| Storyboard plate | pre-video, cinematic still | Photoreal cinematic template + AR for shot |
| Product / mockup | e-commerce, packaging | Studio lighting + material accuracy templates |
3. Craft the prompt
Layer in order:
- Intent lead — recreate / restyle / enhance / vary…
- Fidelity anchors (from
references/prompt-library.md §1)
- Creative layer — style, mood, environment, product brief
- Quality boosters — detail, sharpness, anatomy, lighting (avoid empty keyword spam)
- Hard constraints — “preserve text exactly”, “anatomically correct hands”, “identical eye color and freckle pattern”
Load ready templates from:
references/prompt-library.md — full library (styles, restore, storyboard, product, adult intensity tiers)references/adult-cheat-sheet.md — fast R-rated intensity + category keywords
4. Execute
image_edit(
prompt = "<layered prompt>",
image = ["<source path or attachment>"],
aspect_ratio = "<if reframing / multi-ref>"
)
For variations, fire distinct prompts in parallel (one call each).
For multi-pass, chain: pass-1 output path → pass-2 image_edit input.
5. Multi-Pass Refinement Protocol (v3.7.1)
Trigger when user asks high quality / hero / pro grade, or first pass fails QA.
| Pass | Focus | Prompt bias |
|---|
| 1 · Structure | Composition, pose, identity, major lighting | High fidelity anchors; allow intended style change |
| 2 · Identity / material | Face, hands, materials, text, product edges | “Preserve pass-1 layout; refine only …” |
| 3 · Polish | Sharpness, color harmony, film/cinematic grade, artifact cleanup | Minimal structural change |
Pass count guidance:
- Simple enhance or mild style: 1 pass
- Style transfer + identity: 2 passes
- Hero plate / design sheet / pre-video: 2–3 passes
- After 3 fails on identity → hand off to
character-dna-extractor + i2i-cinematic-refiner
6. Consistency Engine (series & cast)
When producing ≥2 related images:
- Write 3–7 consistency anchors (face shape, hair, scars, wardrobe color, lens feel).
- Inject the same anchor block into every prompt.
- Prefer edit-from-best-plate rather than fresh
image_gen for each frame.
- For recurring characters entering Studio video:
ACTIVATE Character DNA Extractor → Identity Lock → then more recreations.
7. Review & QA (before presenting)
8. Save & report
Prefer:
artifacts/recreation_<mode>_<slug>_<passN>.png
End with a short Recreation Report (see Output Format).
Mode Playbooks (quick)
Faithful recreation / enhancement
Ultra-faithful recreation of the reference. Exact composition, subjects, poses, camera angle, lighting direction. Only improve sharpness, detail, dynamic range, and remove noise. No content changes.
Style transfer
Exact scene and layout of the reference, reimagined in [STYLE]. Preserve subjects, poses, spatial relationships. [style descriptors]. Photoreal/illustration quality as appropriate.
Environmental immersion
Keep this exact subject identity and pose; fully replace environment with [WORLD]. Match lighting direction on subject to new environment; realistic contact shadows and color bounce.
Character design sheet
Produce separate image_edit calls (or sequenced): front, 3/4, side, back, expression set — shared DNA anchors every call. Optional grid composite via later edit if needed.
Cinematic storyboard keyframe
Cinematic film still from the reference subject, [shot size], [lens feel], subtle grain, color grade for [genre], motion-ready clarity for image-to-video. Lock identity and wardrobe.
Product / commercial
Commercial product plate: accurate materials, controlled studio light, clean reflections, catalog-ready, preserve logo/text legibility.
Reasoning & Cache (Grok 4.5)
| Task | Reasoning | Cache |
|---|
| Trivial single restyle | medium | optional |
| Multi-pass hero / identity | high | prompt_cache_key = project slug |
| Design sheet batch | high | same key across batch |
| Long multi-ref memory | opt-in grok-4.3 only if context > ~400k | long bank key |
Integration Chain
User upload
→ AI Image Recreation (this skill)
→ optional Character DNA Extractor (recurring cast)
→ optional i2i-cinematic-refiner / i2i-refiner (polish)
→ Imagine Prompt Master (if entering full Studio prompts)
→ image_to_video / Sequence Director / Studio Director
→ Reference Asset Curator (hero vs draft tier)
| Downstream need | Activate |
|---|
| Motion from approved still | image_to_video or ACTIVATE I2V_SPECIALIST |
| Full production | ACTIVATE STUDIO_DIRECTOR / cinematic studio |
| Explicit sequence design | ACTIVATE EROSFORGE (after still recreation if needed) |
| Quota planning | workflow-quota-optimizer / sfw-batch-orchestrator |
| Docs / pitch | pptx / pdf with before/after artifacts |
| Imagine API / grok.com paste | imagine-execution-bridge / Agent Mode Handoff |
NSFW / Adult Protocol
- Apply adult templates only when the user clearly requests adult content.
- Use intensity tiers (Light / Medium / Hardcore) from
references/adult-cheat-sheet.md.
- Keep fidelity anchors so identity does not drift under style escalation.
- For multi-clip intimate video pipelines, hand off to ErosForge + NSFW sequence tools; this skill remains the still/upload recreation specialist.
- Never process sexual content involving minors; refuse clearly.
Output Format
AI IMAGE RECREATION COMPLETE · v3.7.1
Mode: <faithful|style|edit|enhance|variation|sheet|storyboard|product>
Passes: <n> | Reasoning: <medium|high>
Source: <path or attachment>
Outputs:
- artifacts/...
Consistency anchors: <3–7 bullets or n/a>
QA: <pass / notes>
Next recommended: <iterate | DNA | i2i | i2v | Studio Director | done>
References (load on demand)
| File | Use |
|---|
references/prompt-library.md | Templates: fidelity, styles, restore, variations, storyboard, product, adult |
references/adult-cheat-sheet.md | Fast intensity + category keywords |
~/.grok/skills/imagine/SKILL.md | Global image_gen / image_edit craft rules |
Studio MODEL_LAYER_v3.7.1.md | Canonical model stack |
Quality Bar
- Prefer one strong pass over five vague retries; escalate structure when stuck.
- Prefer edit chain from best plate for series consistency.
- Prefer DNA + Identity Lock before expensive video on a face from an upload.
- Never invent tool parameters; never treat Imagine image/video slugs as chat models.
AI Image Recreation v3.7.1 — Grok 4.5 orchestration · Imagine image_edit / image_gen · SuperGrokPro + Cinematic Studio compatible
