Local AI Infrastructure Notes (13/15) — Auto-Generating Blog Images with Nano Banana 2

Implementation notes on integrating Gemini 3.1 Flash Image (Nano Banana 2) with the Blogger API to auto-generate and insert images into blog posts.

Key Takeaways

  • Nano Banana 2 (Gemini 3.1 Flash Image) generates images via the generate_content API — a different call structure from Imagen 4.
  • The pipeline parses blog HTML to automatically determine per-section image insertion positions.
  • Actual cost including thinking tokens is ~$0.08/img, higher than the documented reference price of $0.067/img.

Background

The Blogger API auto-publish pipeline was already in place. The goal was to add image generation as a pipeline stage. Text-only posts hurt readability, and manual image creation becomes inefficient at scale. The solution: insert an image generation step into the existing API pipeline.

Body

1. Model Selection — Imagen 4 vs Nano Banana 2

Imagen 4 (imagen-4.0-generate-001) was evaluated first. It uses the generate_images API at approximately $0.02/img. However, text rendering inside images was frequently corrupted, and diagram expressiveness was limited.

Nano Banana 2 (Gemini 3.1 Flash Image, gemini-3.1-flash-image-preview) uses the generate_content API. Its SDK call structure differs from the Imagen family, and it consumes thinking tokens alongside output tokens — a factor that materially affects cost.

Item Imagen 4 Nano Banana 2
Model ID imagen-4.0-generate-001 gemini-3.1-flash-image-preview
API generate_images generate_content
Reference price (docs) ~$0.02/img ~$0.067/img
Actual price (thinking included) — ~$0.08/img
Text rendering Frequent corruption Clean English text
Image quality Simple illustration Precise diagrams

Cost note: The documented $0.067 does not account for thinking token overhead. With thinking tokens included, the measured cost is $0.08/img. Use $0.08 as the planning figure for bulk generation.

2. API Call Structure

Nano Banana 2 uses generate_content, and its response structure differs from Imagen 4.

import google.generativeai as genai
from PIL import Image
import io, base64

genai.configure(api_key=API_KEY)
model = genai.GenerativeModel("gemini-3.1-flash-image-preview")

response = model.generate_content(
    contents=prompt,
    generation_config=genai.types.GenerationConfig(
        response_modalities=["IMAGE", "TEXT"]
    )
)

for part in response.candidates[0].content.parts:
    if hasattr(part, "inline_data") and part.inline_data:
        img_bytes = base64.b64decode(part.inline_data.data)
        img = Image.open(io.BytesIO(img_bytes))

Imagen 4's generate_images retrieves the image from response.generated_images. Nano Banana 2 requires iterating candidates[0].content.parts and extracting parts that carry inline_data. When using both models in the same codebase, image extraction logic must be branched per model.

3. Prompt Engineering — The Korean Text Trap

Passing Korean blog titles directly into prompts causes Korean characters to render corrupted inside the generated image. In some cases, meaningless character strings resembling CSS code appeared in the output.

Resolution strategy:

  1. translate_title_to_concept maps Korean technical terms to English visual concepts:
  2. "์—์ด์ „ํŠธ" → "AI agent"
  3. "๋ฐ์ดํ„ฐ ํ๋ฆ„" → "data flow"

  4. "์˜จํ†จ๋กœ์ง€" → "ontology"

  5. strip_korean forcibly removes any remaining Korean characters after the mapping step.

  6. Explicit language constraint in the prompt: All text must be in English only. No Korean, no Japanese, no Chinese characters.

With both steps applied, the probability of Korean characters appearing in the image converges to zero.

Automatic visual theme detection:

Keywords extracted from section content are injected as visual hints into the prompt.

def build_visual_hint(section_text: str) -> str:
    if "flow" in section_text or "ํŒŒ์ดํ”„๋ผ์ธ" in section_text:
        return "Connected nodes with data flowing between components"
    if "architecture" in section_text or "์•„ํ‚คํ…์ฒ˜" in section_text:
        return "Modular system architecture with interconnected blocks"
    return "Clean technical diagram on white background"

4. Structural Analysis — Automatic Image Insertion Positioning

Rather than placing all images at the top, the pipeline parses HTML structure to select contextually appropriate positions.

Parsing:

from bs4 import BeautifulSoup

def extract_sections(html: str) -> list[dict]:
    soup = BeautifulSoup(html, "html.parser")
    sections = []
    for tag in soup.find_all(["h2", "h3"]):
        content = []
        for sib in tag.next_siblings:
            if sib.name in ["h2", "h3"]:
                break
            content.append(sib.get_text())
        sections.append({
            "title": tag.get_text(),
            "body": " ".join(content),
            "length": len(" ".join(content))
        })
    return sections

Placement rules:

  • Header image: 1 image always inserted at the top of the post.
  • Section images: priority given to sections with 200+ characters of body text containing technical keywords (structure, pipeline, architecture, flow, etc.); maximum 2 images.
  • Total image cap: 3 images per post.

5. Rate Limit Management

Nano Banana 2 Tier 1 limits:

Limit Value
RPM (requests per minute) 100
TPM (tokens per minute) 200,000
RPD (requests per day) 1,000

Each image consumes approximately 1,300 tokens (prompt + output + thinking), placing the TPM ceiling at roughly 153 images per minute. The actual bottleneck is RPM 100. With a safety margin applied, intervals are set to 2 seconds between images and 3 seconds between posts.

Bulk processing within the RPD 1,000 limit is achievable within a single day.

6. Blogger API Integration

Generated images are base64-encoded and embedded directly in the Blogger HTML. This eliminates the need for external image hosting and bypasses Blogger's image upload constraints.

def insert_image_to_html(html: str, img_bytes: bytes, position: str) -> str:
    b64 = base64.b64encode(img_bytes).decode()
    img_tag = f'<img src="data:image/png;base64,{b64}" style="width:100%;max-width:800px;" />'
    soup = BeautifulSoup(html, "html.parser")
    if position == "header":
        soup.body.insert(0, BeautifulSoup(img_tag, "html.parser"))
    else:
        target = soup.find("h2", string=position)
        if target:
            target.insert_after(BeautifulSoup(img_tag, "html.parser"))
    return str(soup)

When updating via the Blogger API in batch mode, a 10-second interval between requests is mandatory to avoid rate limit errors.

Lessons Learned

Imagen 3 deprecated: Initial attempts used Imagen 3, which had already been discontinued. Always verify model availability against the latest API documentation before selecting a model.

generate_images vs generate_content confusion: Within the same Google AI SDK, Imagen-family and Gemini-family models have different call structures. Response parsing logic must be branched per model family.

Residual Korean text: Korean terms absent from the translate_title_to_concept mapping table pass through to the prompt and produce corrupted text in the image. The strip_korean function was added as a second-pass safeguard.

Underestimated thinking token cost: Budgeting against the documented price ($0.067) leads to cost overruns against actual spend ($0.08). Thinking token overhead must be factored into budget planning.

Conclusion

Three things drive the Nano Banana 2 auto-image generation pipeline.

  1. API structure — generate_content and generate_images differ in call signature and response parsing.
  2. Prompt engineering — Two-pass Korean removal and automatic visual theme detection control image quality.
  3. Cost planning — Budget against the measured price ($0.08/img, thinking tokens included).

The pipeline applies identically to new post publishing. The section analysis logic and prompt mapping table are extensible to accommodate additional post types.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »