image
Invoke when a user shares a reference image and wants new images with the same "feel" — composition, color palette, lighting, style, or mood — but different content. Analyzes the image across 10 aesth
SKILL.md.claude/skills/image/SKILL.md
GitHubHub Origin: Convexed from MiniMax Hub skill
image-remix(version 0.7.9). Original Hub-SKILL.md is preserved atscripts-originals/SKILL.md.hub, original meta.yaml atscripts-originals/meta.yaml.hub. All Hub-specific paths (e.g.~/.hub-global/skills/image-remix/) translated to Hermes-equivalent references inreferences/.
Image Remix — Image Inspiration Remix Assistant
The user sends a reference image; analyze what makes it "good," then generate images that retain the core feel but with entirely new content.
Core Philosophy
- Origin-Driven Generation: First determine which model/camera/hand-drawn the original came from, then directly call the corresponding model + matching prompt style — MJ uses keyword stacking, Flux uses natural language long sentences. Origin, tool, and writing style are unified as one
- 10D -> 3D Collapse: Internal 10-dimension analysis, collapsed into 3 user-visible dimensions: Form / Aesthetic / Mood, precisely locating the "soul dimension"
- Preserve Feel, Replace Content: Remix = swap content, keep the soul
- Ultra-Lightweight Operation: Send one image + say one sentence -> get output, results first then iterate
Conventions
- Intermediate artifacts are stored in
./.image-remix/{session_name}/(session_name derived from image filename, lowercased with special characters removed) - No collages/grids: Each generation call produces only 1 image. Never include "grid," "collage," "2x2," etc. in prompts
- ⚠️ MJ Four-Grid Note: MJ always outputs a 2x2 four-grid; the user crops it on the canvas themselves
- ⚠️ Do NOT pass reference images to text-to-image tools: The reference image is only used for Phase 1 analysis; generation relies entirely on text prompts to avoid copyright risks
Pipeline
Phase 0: Input → Phase 1: Decompose → Phase 2: Quick Select → Phase 3: Generate → Phase 4: Iterate → Phase 5: Output
Phase 0: Input
Supported: local path / URL (download with import_images) / search description (use image_search).
Multiple images: analyze separately -> extract each image's soul dimension -> merge into a combined prompt (e.g., "Image A's composition + Image B's color palette") -> inform the user then generate.
File structure:
.image-remix/{session_name}/
├── decompose.json # Phase 1 analysis results
├── generation_log.md # Generation log
└── outputs/ # Generated images v1_01.png …
Phase 1: Decompose
1a. Origin Tracing
Use read_media to analyze the original image's origin, cross-referencing visual fingerprints in references/model-fingerprints.md:
read_media(
file_path=reference image path,
question="Determine which AI model / camera / hand-drawn this image most likely comes from.
Cross-reference the visual fingerprints in model-fingerprints.md one by one.
Focus on: skin rendering, highlight reflections, edge blending quality, color tendencies, overall AI feel.
Output: model_guess, confidence(0-1), clues(at least 3 specific pieces of evidence),
alt_guess, is_ai_generated, is_photo, is_illustration"
)
1b. 10-Dimension Analysis
read_media(
file_path=reference image path,
question="Deeply analyze this image across the following 10 dimensions, providing for each:
description (English technical description), summary (Chinese summary), score (0-1 salience):
1.Composition 2.Art Style 3.Color (including palette hex values) 4.Lighting 5.Atmosphere
6.Texture 7.Subject 8.Narrative 9.Post-processing Style 10.Medium/School"
)
1c. Collapse to Form / Aesthetic / Mood
| Dimension | Underlying Components | Meaning |
|---|---|---|
| Form | Composition + Subject | What is depicted and how it's arranged |
| Aesthetic | Art Style + Color + Texture + Post-processing + Medium | What painting style and tonal quality |
| Mood | Atmosphere + Lighting + Narrative | What feeling and emotion |
Take the weighted average of sub-dimension scores to identify 1-2 soul dimensions (the most distinctive ones).
1d. Write to decompose.json
Merge origin + dimensions + collapsed + soul fields and write to file.
Phase 2: Quick Select
Display Origin Conclusion (mandatory)
📍 Origin Conclusion
Image Source: [model_guess] (confidence [confidence])
Evidence: [2-3 most critical clues]
Soul Dimension: [soul_summary]
🎯 Direct Call: [matched tool] — [one-sentence explanation why]
Prompt Style: [writing strategy for this model]
Fallback: [what to use if tool is unavailable]
Origin -> Tool + Prompt Style Mapping Table (origin directly determines which tool and style to use; the two are inseparable):
| Origin Source | Tool to Call | Prompt Style | Fallback |
|---|---|---|---|
| Seedream | Seedream image generation tool | Realistic long description, precise facial features/clothing materials, refined lighting and atmosphere | all_in_one, keep style |
| Kolors | Kolors image generation tool | Realistic description, emphasize highlight-material contrast | all_in_one |
| Hunyuan | Hunyuan image generation tool | Realistic + Chinese aesthetic vocabulary, concise with breathing room | all_in_one |
| Jimeng | Jimeng image generation tool | Direct description, explicit style keywords, write out saturation levels | all_in_one |
| Wanx / Tongyi Wanxiang | Wanx image generation tool | Concise description, avoid overly dense detail stacking | all_in_one |
| Midjourney v5/v6 | MJ image generation tool | Keyword stacking + style tags, concise and not verbose | all_in_one, keep keyword style |
| Flux.1 | Flux image generation tool | Natural language long sentences, detailed lighting and material description | all_in_one, keep long description |
| Stable Diffusion | SD image generation tool | Supports weight syntax (word:1.3), negative prompt written separately |
all_in_one, remove weights and switch to natural language |
| DALL-E 3 | DALL-E image generation tool | Natural language, straightforward description, avoid over-stylization | all_in_one |
| Ideogram | Ideogram image generation tool | Specify text content and layout requirements clearly | all_in_one |
| Real photograph | High-realism image generation tool | Photography terminology dominant, lens parameters + film stock | all_in_one, keep photography terminology |
| Hand-drawn / Illustration | Illustration image generation tool | Medium + school as anchor | all_in_one, keep medium description |
| Generic AI / Hard to determine | all_in_one_image_generation |
Keywords + natural language hybrid | No fallback needed |
| Completely uncertain | Multiple tools, one image each for comparison | Use each tool's optimal prompt style | — |
Fallback rule when tool is unavailable: Fall back to all_in_one_image_generation, but keep the original tool's prompt writing style unchanged, preserving the style DNA as much as possible.
Generation Path Selection (must ask)
Ask the user via Q&A:
- Text-to-Image (recommended): Pure text prompt, zero copyright risk
- Image-to-Image: Reference image passed as image_ref, higher fidelity, user must confirm copyright risk
Content Direction Confirmation
Directly ask the user "What do you want to create?" — a single sentence is enough to proceed to generation.
If the user has no direction (only says "help me remix"), recommend 3 replacement directions based on the soul dimension for selection.
Quick Commands: Reskin (swap subject only) / Extract Style (preserve Aesthetic) / Extract Composition (preserve Form) / Extract Mood (preserve Mood)
Phase 3: Generate
Branch A: Text-to-Image
Strictly prohibit passing the reference image to the generation tool. The prompt is entirely reconstructed based on Phase 1 analysis.
Prompt Construction:
[New content subject]. [English description of preserved dimensions]. single subject, one image only, no grid, no collage, no text, no watermark
Dimension -> Prompt Focus:
| Preserved Dimension | Prompt Emphasis |
|---|---|
| Form | Camera angle, shot scale, subject position, depth layers |
| Aesthetic | Art school, hex values/color relationships, medium description, color grading reference |
| Mood | Emotion adjectives, light source description, cultural/era references, metaphorical language |
Branch B: Image-to-Image
Reference image passed as image_ref; prompt focuses on describing the new content direction, with style description as supplementary guidance.
Generation Parameters
Ask the user for size: Follow Original (default, measure with ffprobe then map) / 1:1 / 16:9 / 9:16
"Follow Original" mapping rules: ratio > 1.5 -> 16:9; 0.8~1.2 -> 1:1; < 0.8 -> 9:16. Must measure first then write to task_description; guessing is prohibited.
Default generation: 4 variants, each called separately, with slight prompt variations to maintain diversity.
Save output to .image-remix/{session_name}/outputs/v{version}_{index}.png, and synchronously update generation_log.md.
Phase 4: Iterate
After displaying results, ask about satisfaction: Satisfied, finish / Fine-tune / Change direction / Switch model.
Fine-tuning adjusts at the prompt level directly, without going back to Phase 1 for re-analysis:
| Feedback | Action |
|---|---|
| "Make the colors cooler" | Fine-tune color prompt words |
| "Change composition to close-up" | Adjust camera parameters |
| "Style is too different" | Strengthen style description or switch model |
| "This one's composition + that one's color" | Combine dimensions from two images and regenerate |
Phase 5: Output
Delivery checklist: file location / total generated / iteration rounds / soul dimension / origin source / prompt writing strategy
Optional: CDN upload / style library entry (save to .image-remix/style-library/{style name}.md) / continue generating variants
Style library format (see references/analysis-example.md for an example):
# {Style Name}
## Origin
- Image source / tool called / prompt writing strategy
## Soul Dimension
- Soul: {Form/Aesthetic/Mood} — {soul_summary}
## Successful Prompt
{complete English prompt}
## Dimension Snapshot
- Form / Aesthetic / Mood