Skip to content

CHANGELOG.md

Latest commit

 

History

History
121 lines (90 loc) · 13.4 KB

CHANGELOG.md

File metadata and controls

121 lines (90 loc) · 13.4 KB
Raw

Changelog

All notable changes to the Nano Banana Prompt Generator skill and plugin are documented here.

The format follows Keep a Changelog, and this project adheres to Semantic Versioning.

[1.2.0] - 2026-08-05

Maintenance audit against Google's official Gemini API docs (ai.google.dev/gemini-api/docs, fetched from the canonical .md.txt sources on 2026-08-04). This is a correctness-critical release: the model IDs shipped in v1.1.0 have been shut down by Google.

Fixed

  • Dead model IDs removed. gemini-3.1-flash-image-preview and gemini-3-pro-image-preview - both recommended by v1.1.0 - were shut down on 2026-06-25. Replaced with the GA IDs gemini-3.1-flash-image and gemini-3-pro-image (both GA 2026-05-28). A skill that emits a shut-down model ID produces a 404 for the user; this was the reason for the audit.
  • Default model recommendation reversed. v1.1.0 recommended gemini-2.5-flash-image as "the only stable model" and the production default. That model is now deprecated with shutdown on 2026-10-02, and Google's own guidance is to migrate off it. The default is now Nano Banana 2 (gemini-3.1-flash-image), matching Google's "should be your go-to image generation model" wording in the image-generation guide.
  • "Preview" framing dropped throughout. All four models are Stable/GA; the caveats about preview instability and restrictive preview rate limits no longer apply.
  • C2PA claim corrected. v1.1.0 stated the docs never mention C2PA. Nano Banana 2 Lite's model card explicitly lists "SynthID (Always On) + C2PA". The claim is now per-model instead of blanket-denied.
  • Pro reference-image capacity corrected from "up to 6 objects high-fidelity" to the documented 5, and the invented per-character count for Pro removed (Google does not publish one).
  • Text-rendering language claim replaced with the official list. v1.1.0 said the docs do not enumerate languages. They now do: 15 best-performance locales. Polish is not among them - the skill states this plainly and gives the proof-render / text-layer-in-post fallback, which matters for Polish-language deliverables.

Added

  • Fourth model: Nano Banana 2 Lite (gemini-3.1-flash-lite-image, GA 2026-06-30) - sub-2s latency, cheapest per image ($0.0336 at 1K Standard, ~$0.017 Batch), 1K only, 10 aspect ratios, function calling supported, no search grounding, explicitly not optimised for multiple references or long multi-turn edit chains.
  • Video-to-image section - gemini-3.1-flash-image accepts a public YouTube URL or an uploaded video file as multimodal context and generates a still from it (launched 2026-05-28, exclusive to this model). Includes thumbnail / poster / summary-infographic prompt patterns and the caveat that output is synthesised art direction, not a frame extraction.
  • Two-step text workflow from the official best-practices list: generate the copy in one turn, then ask for the image with that copy - instead of one shot.
  • Deprecation timeline table in "Verification & Sources": GA dates, -preview shutdown, Lite GA, sampling-parameter deprecation, Imagen shutdown, legacy 2.5 shutdown. Gives the next audit a dated baseline.
  • response_format mechanics - aspect_ratio and image_size are parameters, not prose. Documented with the note that stating "16:9" in prompt text does not set the canvas.
  • Grounding limitation - on Nano Banana 2, Search grounding does not support real-world images of people from web search.
  • Exact pixel dimensions for common 1K ratios on Nano Banana 2 (16:9 = 1376x768, 9:16 = 768x1376, 4:5 = 928x1152, 1:1 = 1024x1024).
  • 4 new eval cases covering video-to-image, model/cost selection, the 404-on-preview-ID scenario, and Polish diacritics.

Changed

  • Frontmatter split into description + when_to_use (the when_to_use field is now supported and is appended to description in the skill listing, sharing a 1,536-character cap). Trigger phrases and model codes moved out of description, which now states what the skill does. Combined length 1,019 characters.
  • Capability matrix rebuilt as four columns and extended with video input, multi-reference capacity, function calling, Flex/Priority availability, and token limits, all from the per-model cards.
  • Resolution and aspect-ratio tables now per-model (Lite is 1K-only; the extreme banner ratios 1:4/4:1/1:8/8:1 are exclusive to Nano Banana 2).
  • references/pricing.md rewritten for four models across Standard / Batch / Flex / Priority, with the note that Flex and Priority are unavailable on Nano Banana 2 and that legacy 2.5 costs more than Lite for a worse result.
  • Sampling parameters - noted temperature, top_p, top_k were deprecated API-wide on 2026-07-21, so prompt advice must not lean on them.
  • Access points - the image docs are now written against the Interactions API (client.interactions.create, response_format) rather than generate_content.

Source verification notes

Every fact above traces to one of: the image-generation guide, the four model cards, the pricing page, the API changelog, or the deprecations page - all on ai.google.dev, fetched 2026-08-04 as raw markdown. No third-party reviews were used. Where Google does not publish a number (Pro character count, total supported language count, C2PA on models other than Lite), the skill says so instead of estimating.

[1.1.0] - 2026-05-21

This release was rewritten against Google's official Gemini API documentation at ai.google.dev/gemini-api/docs, fetched via curl directly from the canonical markdown sources on 2026-05-21. The earlier draft relied on third-party reviews and contained factual errors corrected below. All facts in this entry are sourced from official Google pages - URLs cited in SKILL.md "Verification & Sources" section.

Added

  • Three-model lineup with canonical model IDs - gemini-2.5-flash-image (Nano Banana, Stable), gemini-3.1-flash-image-preview (Nano Banana 2, Preview), gemini-3-pro-image-preview (Nano Banana Pro, Preview). Each row includes status, last-updated date, and best-use guidance.
  • Per-model capability matrix - thinking, search grounding, batch, caching, structured outputs - sourced from each model's spec page on ai.google.dev.
  • Image Search Grounding section - documented for Nano Banana 2 (3.1 Flash adds Image Search alongside Web Search per the model spec).
  • references/pricing.md - per-image cost across Standard / Batch / Flex / Priority tiers, plus token math footnotes copied verbatim from the official pricing page. Verified 2026-05-21.
  • evals/evals.json - 12 should-trigger + 7 should-not-trigger + 2 edge cases for the skill-creator framework.
  • .claude-plugin/marketplace.json - the repo now serves as both plugin source and marketplace catalog. Marketplace name: maciejdzierzek (matches kling-ai marketplace for cross-skill consistency).
  • Verification & Sources section in SKILL.md - links every major claim back to its docs URL on ai.google.dev.
  • Explicit "what this skill does NOT claim" section - calls out unverified numbers (the "100+ languages" figure, C2PA, the 4-stage Veo pipeline) so future authors do not reintroduce them.
  • Storyboards for Video Workflows section - documents the actual handoff to Veo, referencing the Veo guide instead of inventing a multi-stage pipeline.
  • CHANGELOG.md (this file).

Changed

  • Frontmatter description - extended to cover all three models and feature surfaces (in-image translation, search grounding, multi-character consistency limits). Length within character cap.
  • Quick Start example switched from "4K" default to "2K, 3:2" to match a realistic billing decision (4K is 2.3x more expensive than 2K on Nano Banana 2; thumbnail use cases need 1K).
  • Marketplace install command standardized to /plugin install nano-banana-prompt-generator@maciejdzierzek - shorter marketplace name, consistent with kling-ai-prompt-generator.
  • plugin.json description updated to mention all three model IDs and key features.
  • README.md - Features list rewritten, Models table refreshed against per-model spec pages, install commands corrected.
  • Aspect ratio list in technical specs - corrected from the partial 4 ratios in the earlier draft to the full 14 documented in the image-generation guide (1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9, 21:9).
  • Capacity claims split per-model. Earlier draft mixed Pro and NB2 capacity into a single misleading row. Now each model has documented limits sourced from the "Use up to 14 reference images" table in the image-generation guide.

Fixed (corrections from official source verification)

  • Nano Banana Pro is gemini-3-pro-image-preview, NOT gemini-3.1-pro-preview. Earlier draft conflated the image-generation Pro model with the text/multimodal Gemini 3.1 Pro Preview - these are different products on different pricing tiers. The image Pro model is gemini-3-pro-image-preview per its own spec page.
  • Nano Banana (original) is the Stable mainstream model, not "Simple quick edits". Per the gemini-2.5-flash-image spec page: "best for high-volume generation, conversational image editing, and low-latency creative workflows that require native multimodal understanding." Repositioned accordingly in the Model Lineup table.
  • "Default: Nano Banana 2" softened. NB2 is a Preview model and may change; production defaults should be gemini-2.5-flash-image (Stable). The skill now distinguishes "production default" from "feature-richer Preview option".
  • Multi-character / multi-object capacity per model corrected. Earlier draft claimed "up to 14 reference images, 6 with high fidelity" with mixed-model attribution. Actual per-model numbers from the official "Use up to 14 reference images" table:
    • Nano Banana 2: up to 10 objects high-fidelity, up to 4 characters consistency, 14 total
    • Nano Banana Pro: up to 6 objects high-fidelity, up to 5 characters consistency, 14 total
    • Original Nano Banana: best with up to 3 input images
  • Aspect ratios full list - was 4 ratios, actual is 14 per the Python code example in the image-generation guide.
  • Per-image pricing - added verified token math for each model. Previously omitted; now sourced from pricing-page footnotes.

Removed

  • "Supported for 100+ languages" - third-party claim. Official docs do not enumerate a language count. The Spanish example in the in-image-translation section is the only language-coverage hint in the guide. Replaced with "verify against current docs if a specific script matters."
  • "SynthID + C2PA Content Credentials" watermark claim - C2PA is not mentioned in the Gemini API image-generation docs. Only SynthID is documented. C2PA reference removed entirely.
  • "4-stage video production pipeline: Nano Banana + Veo + Lyria + Flow" - third-party narrative. Official docs only link to the Veo guide at the end of the image-generation page. The "keyframe handoff" prompt template remains (as Storyboards for Video Workflows) but is no longer presented as a documented Google pipeline.
  • "Default access via Gemini app, Flow, AI Studio, Vertex AI, Gemini API, Gemini CLI" - access list trimmed to surfaces explicitly named in the API docs (Gemini API, Google AI Studio, Vertex AI). Consumer surfaces mentioned in passing but no longer presented as exhaustive.
  • "Gemini 3 Pro Image was deprecated March 9, 2026" (was in README) - false. gemini-3-pro-image-preview is the current Preview Pro model and remains active.
  • Stale version: 1.0.0 placeholder.

Source verification notes

This release was researched against:

Fetched via curl https://...md.txt from the canonical Devsite markdown sources, not via WebFetch summaries.

Things NOT verified and therefore softened or omitted:

  • The "100+ languages" text-rendering count (no enumeration in official docs - skill says "improved i18n text rendering" per the 3.1 Flash release blurb).
  • C2PA support (not mentioned in image-generation docs - skill says SynthID only).
  • Default consumer-app model (skill scopes to the API surface).
  • Lyria audio integration with image generation (not documented as a connected workflow).

Verify against ai.google.dev/gemini-api/docs before quoting any specific number to a client.

[1.0.0] - 2026-03-14

Added

  • Initial release of the Nano Banana prompt generator skill.
  • SKILL.md with model overview (3 models), prompt structures for generation/editing/thumbnails/infographics, Golden Rules, technical specs, common issues.
  • references/prompt-examples.md - template library by category.
  • examples/ directory - generation, editing, thumbnails, infographics samples.
  • Plugin manifest (plugin.json), MIT license, README with installation instructions for Claude.ai, Claude Desktop, and Claude Code.

Source basis: third-party reviews and community discussion (no canonical Google docs verification at the time). Corrections incorporated in v1.1.0.