Sewer Image Fisher
ArcSelf/sewer-image-fisher
Busca, processa e gera imagens com renderização pointilista no terminal
Ver no GitHubStudio de imagens terminal com 5 modos: busca SEO-ranqueada via DuckDuckGo, pré-processamento para datasets SDXL LoRA, geração local com Hugging Face diffusers, renderização pointilista 24-bit, e modo empírico com análise de luminância. Skill nativa para Claude Code.
Instalação
npx convez-ai ArcSelf/sewer-image-fisher1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Preparar datasets de imagens para fine-tuning SDXL LoRA
- Gerar imagens localmente com modelos SDXL e LoRAs customizadas
- Visualizar imagens em ASCII/pointilismo no terminal com análise de cor
Veredito

Ferramenta especializada para pipelines de geração de imagem locais com foco em preparação de dados e visualização terminal. Requer GPU e ecossistema ML, melhor para quem domina SDXL e LoRA.
README
A tradução para o português deste README ainda está na fila.
sewer-image-fisher
A four-stage image studio for the terminal: fish images from the open web (SEO-ranked), preprocess them into SDXL LoRA-ready datasets with Kohya caption sidecars, generate locally via HuggingFace diffusers (SDXL + LoRA), and view any result as terminal pointillism — palettes dots / ascii / impasto with optional jitter, or live denoising animation via --stream.
A fifth, lightweight mode (empirical) treats each pixel as a Humean impression: fetches a single image, renders it in true 24-bit color, logs per-row luminance and chromatic complexity as sparklines, and paints the terminal's own HSV gamut wheel below for reference.
This is a Claude Code skill — drop it into ~/.claude/skills/image-fisher/ and Claude will auto-invoke it when you ask things like "fish me images of a taco" or "generate me an SDXL of a rusted vespa". Standalone usage from the CLI works just as well.
What's in the box
fish.py DuckDuckGo image search, SEO-ranked (alt text, og:title, schema,
filename slug, dimensions, domain quality) via BeautifulSoup
page verification. Optional --preprocess hand-off.
preprocess.py SDXL aspect-ratio bucketing (9 native buckets), perceptual-hash
dedup, EXIF orientation + RGB normalization, PNG re-encoding,
Kohya-format caption sidecars (.txt per image).
generate.py Local SDXL via HuggingFace diffusers. Auto-detects CUDA → MPS →
CPU. Optional LoRA injection. With --stream, decodes mid-flight
latents through the VAE every Nth step and paints pointillism in
the terminal with ANSI cursor overdraw — animated denoising.
view.py Terminal pointillism engine. Three palettes (dots / ascii /
impasto). --organic enables character, bucket, and color
jitter so the result reads as a painted canvas. Also supports
unicode half-blocks (24-bit) and iTerm2 / Kitty inline graphics.
empirical.py Lightweight single-image mode. One image from DDG → 24-bit
half-block render → per-row luma + chromatic Δ sparklines →
dominant-colors log with live ANSI swatches → HSV gamut wheel
for terminal capability reference.
Requirements
Python 3.10+ and a truecolor-capable terminal (iTerm2, WezTerm, Ghostty, Kitty, Alacritty, or modern macOS Terminal.app with COLORTERM=truecolor set).
Core dependencies:
pip install -r requirements.txt
For local SDXL generation (generate.py only), install the ML stack separately so the rest of the skill stays light:
pip install diffusers transformers accelerate safetensors
pip install torch # Apple Silicon → MPS by default; Linux/CUDA → use the pytorch.org selector
Quick start
Fish + preprocess in one shot (SDXL LoRA pipeline)
python3 fish.py "rusted vespa" -k 20 --preprocess
Generate locally with streaming pointillism
python3 generate.py \
--prompt "a rusted vespa parked at golden hour, cinematic" \
--stream --stream-interval 5 --stream-palette impasto \
--preview-palette dots --seed 42 --steps 30
View any image as pointillism
python3 view.py ./photo.jpg --palette impasto --organic
Single-image empirical mode (no disk writes)
python3 empirical.py "rusted vespa golden hour"
python3 empirical.py --only-wheel # just the terminal's color gamut
python3 empirical.py --image ./photo.jpg
Using as a Claude Code skill
Drop the folder into ~/.claude/skills/image-fisher/. Claude will auto-invoke when you ask things like:
| You say | What runs |
|---|---|
| "fish me images of X" | fish.py |
| "prep these for a LoRA" | preprocess.py |
| "generate / draw / synthesize / render me an SDXL of X" | generate.py |
| "show me X as pointillism" / "paint it in the terminal" | view.py |
| "empirical render of X" / "test the terminal's color" | empirical.py |
See SKILL.md for the full auto-invocation contract and flag reference.
Aesthetic note
The whole skill is built around an idea: the terminal is a canvas, not just a control surface. Pointillism is the bridge — each cell is a discrete dab of paint, each dab is a discrete empirical impression. Half-block 24-bit rendering is the most pixel-faithful mode any terminal can produce. --organic adds Seurat-style stochastic variation. empirical.py is the Humean slice: row-by-row, pixel-by-pixel, the terminal's own gamut shown beside the image as the reference canvas.
Brilliantly archaic, modernly precise.
License
Licensed under the Apache License 2.0. See LICENSE and NOTICE for third-party attributions.
Contributing
PRs welcome. By submitting, you agree your contributions are licensed under Apache 2.0 including the patent grant (Section 3).