Local AI Infrastructure Notes (15/15) — Migrating the Blog Image Pipeline to ComfyUI

Why FLUX Local Execution Failed and How Tailscale-Based GPU Sharing Works

ํ•ต์‹ฌ ์š”์•ฝ

  • What this post covers: the three axes to evaluate when moving a blog image generation pipeline from API to a local GPU (memory · disk · quality), and the design pattern for incorporating another machine on the same LAN as "local."
  • Measured failure points of running FLUX.1 Schnell directly on Mac (memory, disk, quality, speed) with empirical numbers.
  • How to reuse a GPU managed by a separate agent as shared infrastructure via ComfyUI HTTP API + Tailscale virtual IP.
  • Trade-offs (availability · queue latency · fallback policy) that come with moving from $0.08/image API cost → $0/image, and Mac memory footprint 25 GB → 0 GB.

What You Can Take From This Post

  1. Checklist for running local image generation models (FLUX family) on Apple Silicon — memory contention with co-resident models, disk headroom relative to model weight size, quantization quality loss, inference throughput.
  2. Using ComfyUI as an HTTP backend — REST call endpoints, queue model, explicit parameter control (width/height).
  3. Pinning a stable address with Tailscale virtual IP — a fixed agent-to-agent call path independent of physical IP or NAT changes.
  4. Treating fallback as an operational policy rather than a code path — the hidden cost of maintaining two routes simultaneously.

Background: The API Cost Curve

Measured cost for Nano Banana 2 (Gemini 3.1 Flash Image): $0.08/image. At an average of 30 images per blog post, that is ~$2.40 per post; at 20 posts per month, ~$48/month (≈ ₩67,000). The absolute amount is small, but recurring call costs accumulate linearly until a stop decision is made. If equivalent quality can be produced locally, the justification for keeping the API disappears.

The core question is: which machine counts as "local"?

FLUX.1 Schnell on Mac — Decomposing the Failure

Test configuration: Mac mini M4 32 GB, Gemma-4 26B resident (~25 GB occupied), FLUX.1 Schnell loaded via mflux Q4 quantization.

Item Measured Result
Memory Gemma 25 GB + FLUX 7 GB → OOM Crash
Disk FLUX model 33 GB vs external SSD free space 22 GB Download failure
Quality Q4 quantization output visibly degraded vs Nano Banana 2 Below threshold
Speed ~130 s/image Slower than API

Mechanistic Interpretation

  • Memory: FLUX.1 Schnell requires ~7 GB of resident memory even at Q4 quantization. Sharing unified memory with a 26B-class LLM makes OOM a predictable outcome.
  • Disk: Downloading and caching the original pre-quantization weights requires 30 GB+. If free space on the target drive is smaller than the model size, the process fails at initialization.
  • Quality: Under identical prompts, Q4 lightweight models show measurable gaps in detail and color stability compared to commercial cloud models. Thumbnails depend on first impressions, so the quality floor is relatively high.
  • Speed: 130 s/image is in the "technically runs but cannot be inserted into a batch pipeline" range.

One-off Pilot Feasible; Continuous Operation Is Not

By suspending other agents to free memory to ~9.5 GB and redirecting the cache path to an external SSD, a single pilot run completes. However, this path is incompatible with continuous operation — it requires interrupting other work on the main machine each time. Mac standalone execution is therefore excluded from the design.

Alternative Architecture: Localizing Another Machine's GPU

A separate Windows PC with RTX 3080 is already running ComfyUI (SDXL) for the video production pipeline. Reusing that infrastructure for blog image generation requires zero new resource investment. Three factors make the reuse viable:

  • Protocol: ComfyUI exposes an HTTP API. Any caller can access it via REST regardless of language or OS.
  • Address stability: Any machine running Tailscale gets a 100.x.x.x virtual IP. The address persists through router reboots, LAN reconfiguration, and NAT traversal.
  • Existing call pattern: _try_comfyui is already implemented in the video pipeline's image_generator.py. The request payload and response parsing are validated and can be ported directly into the blog publish skill.

No new infrastructure to build — only one additional caller in an existing structure.

Post-Migration Pipeline

Post writing complete
  → SDXL image request to ComfyUI(<TAILSCALE_IP>:8188) with explicit width/height/prompt
  → Insert returned image into post body
  → Publish via Blogger API

Comparison:

Item Nano Banana 2 ComfyUI Remote
Cost per image $0.08 $0
Mac mini memory footprint 0 GB 0 GB
Average generation speed ~5 s/image ~8 s/image (observed)
Aspect ratio control Prompt only Explicit width/height
GPU host offline Continues Publish held

Limitations and Forward Direction

Availability: GPU Host Offline

Design decision: "hold publish." No automatic fallback to the API. Two reasons:

  1. Maintaining two routes doubles maintenance scope. On failure, branch analysis is required to determine which path is the source.
  2. A fallback implicitly undermines the "$0 cost" target. If exception calls accumulate, measured cost is no longer zero.

→ Publish delay is acceptable. Dual-path routing is not.

Quality Validation: SDXL vs Nano Banana 2

SDXL produces stable output in both photorealistic and illustrative styles — validated by existing video pipeline production. A direct comparison pilot for blog thumbnail use will be run separately. If output falls below threshold, prompt templates, model selection, or post-processing will be adjusted.

Concurrency: Two Agents Sharing One GPU

ComfyUI is queue-based; concurrent requests are processed sequentially. If the video pipeline holds priority, blog image requests queue in the background. Contention becomes latency, not a collision.

Extracted Principles / Applicable Patterns

  • Redefining "local": Local-first design does not mean single-machine execution. A GPU run by another agent within the LAN + Tailscale boundary can be incorporated as "local."
  • Fallback as policy, not function: Resolving it as a code branch doubles maintenance cost. An operational rule — "if unavailable, hold" — is more explicit and does not undermine cost targets.
  • Linear accumulation of recurring costs: Even at ~$48/month, the amount accumulates linearly for as long as the decision is deferred. The lower the unit cost, the harder the stop decision becomes — this must be factored into design choices.

Open Questions

  • Can SDXL-based thumbnails match or exceed Nano Banana 2 in perceived quality in a blog context — which layer (prompt template, model selection, post-processing) drives the gap?
  • To what extent should fairness policies (priority, max wait, cap) be applied to a ComfyUI queue serving multiple callers? The current two-caller structure is sufficient, but if the number of callers grows, queue policy becomes the bottleneck.

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 »