main-image-multi-model

name	main-image-multi-model
description	Run the same Amazon main image prompt across multiple Higgsfield models (GPT Image, Nano Banana Pro, Flux, Soul, Seedream) for side-by-side comparison. Use when running `main-image-multi-model {ASIN}` or when client/team explicitly wants to evaluate model output differences. Default workflow uses best-only single model — use this skill to override.

main-image-multi-model — Multi-Model Concept Comparison

By default the skill library uses best-only models (GPT Image / Nano Banana Pro). This skill is the explicit override when you want a side-by-side: same prompt, 4-5 different models, pick the winner.

Prerequisites

• Higgsfield MCP

Invocation

main-image-multi-model {ASIN} --concept="{prompt}"
main-image-multi-model {ASIN} --models=gpt-image,nano-banana-pro,flux,soul,seedream

Output

/tmp/cro-content/{ASIN}-multi-model-{date}.md with 4-5 images side-by-side + scoring.

Phase 1 — Concept Setup

Single concept prompt (don't compare different concepts — same prompt for fair model comparison). Pull product reference image.

Phase 2 — Generate Across Models

Higgsfield's Multi-Model mode supports running the same prompt across:

• GPT Image (latest) — creative angles, dramatic lighting
• Nano Banana Pro — product fidelity, clean Amazon look
• Flux Kontext — packaging variations, text overlay
• Soul — when character/person is in shot
• Seedream 4.0 — alternate perspective

Run all 4-5 in parallel.

Phase 3 — Score

Same 6-dim rubric as main-image-concepts:

• Fidelity / Background / Scroll-Stop / Compliance / Creative / Quality

Score each model's output. Compare averages.

Phase 4 — Output

# Multi-Model Main Image Comparison — {ASIN}

**Concept:** {prompt}

| Model | Image | Fid | BG | Stop | Compl | Creat | Qual | Avg |
|-------|-------|-----|-----|------|-------|-------|------|-----|
| GPT Image | ![](gpt.png) | 9 | 9 | 8 | 10 | 9 | 9 | 9.0 |
| Nano Banana Pro | ![](nbp.png) | 10 | 10 | 7 | 10 | 7 | 9 | 8.8 |
| ... | ... | ... | ... | ... | ... | ... | ... | ... |

## Recommendation

**Winner:** {model} at {avg} — {why this model nailed this concept}.

For future runs of similar prompts: default to {model}.

Reference Files

• ~/.claude/skills/main-image-concepts/SKILL.md — the standard (single-model) workflow

When to Use

• You explicitly want a model bake-off
• A particular concept isn't working — swap models to see if any handle it better
• Client wants to see "what's the model space look like for our product"

When NOT to Use

• Default content planning — use main-image-concepts (single best-only)
• Time-pressured workflows — multi-model is slower

Auto-Triggers

• User asks "compare models for this prompt" / "what does X model do here"

v0.1.12 — Read brain.image_strategy before generating

Before drafting any Higgsfield prompt, read brain.image_strategy from working memory:

• If null → pause and offer image-strategy first (~10 min, makes every future render bespoke). If user proceeds without, flag concept cards with "No category research used — generic aesthetic."
• If set + fresh (<90d) → inject prompt_adjustments.scene_keywords into prompt section 1, palette_keywords + mood_keywords into section 4, and do NOT include {anti_patterns_csv} near the end. Cite the strategy on each concept card.
• If stale (>90d) → flag and recommend image-strategy --refresh.

This brand-level strategy comes from skills/image-strategy.md's top-15-bestsellers analysis. Same adjustments apply across every render for this brand.

v0.1.13 — Visual verification gate + base64 embedding (NON-NEGOTIABLE)

Florence does NOT present an image she hasn't actually looked at, AND she does NOT use raw Higgsfield URLs in artifact HTML. Two blocking rules added in v0.1.13:

Verification gate

After Higgsfield returns each generated image, BEFORE adding to any artifact OR sending to Pinion:

1. Load the image into Florence's multimodal context — paste the URL in chat so Cowork's multimodal Claude loads it natively. Narrate: "Looking at #{N} before I include it."
2. Visually inspect against: aspect ratio (1:1 main+listing / 16:9 A+), product fidelity vs reference, technique landed (visibly), background appropriate, no clipart leak (anti-clipart rules), no text-on-main-image (TOS), no model faces.
3. If anything fails → re-prompt with specific fix + regenerate. Cap at 3 attempts per concept; surface honestly on attempt 4.
4. Only after ALL concepts pass → proceed to artifact emission.

The eye trumps the score. Even if the rubric said 100/100, if visual inspection finds a wrong product or off-aspect output, it fails the gate.

Base64 embedding

Higgsfield URLs are temporary AND Cowork's artifact iframe sandbox blocks external image loads in many builds. For every verified concept:

1. HTTP GET the Higgsfield URL → fetch image bytes
2. Detect MIME type from response headers (typically image/png)
3. Base64-encode the bytes
4. Substitute {{image-src}} (or {{winner-image-src}} for tests) with data:image/png;base64,<encoded> — NOT the raw Higgsfield URL

This makes the artifact self-contained — survives sandbox + URL expiry. ~1-3 MB per image is fine for Cowork.

The live product image ({{product-image-url}} in concepts.html hero strip / dossier head / cockpit product cards) stays as the live Amazon CDN URL — that's permanent and not affected.