Nous Branding — MultiModel

Above: A 21:9 banner generated with this pipeline.

Above: A 21:9 banner generated with this pipeline using the different paths.

Nous Research visual identity, model-agnostic.
Forked from hermes-theia-codex-vision and nous-branding. Works with any image generation backend.

---

What It Does

Capability	How
Image generation	Text-to-image or image-to-image via any API (OpenAI-compatible, local, etc.)
Analog post-processing	14-step imprint pipeline — risograph grain, CRT scanlines, xerox decay, plate wobble, ink bleed, paper texture, palette compression
Image analysis	Describe, QA, and understand images via any vision model
Batch generation	Generate multiple images from a prompt file, with optional shared reference
Identity-preserve	Pass reference images with --image ref.jpg for consistent characters/style

Quick Start

How it works

1. Configure your API — Pick any image generation model (3 examples below)
2. Generate images — Text-to-image, with style references, or batch
3. Post-process — Run the imprint pipeline for analog print aesthetics
4. Composite — For banners: generate panels → composite with PIL → imprint

1. Configure your API

Edit scripts/generate_image.py and set your endpoint + key:

# Example: OpenAI (DALL-E 3)
API_URL = "https://api.openai.com/v1/images/generations"
API_KEY = "sk-..."
MODEL = "dall-e-3"

# Example: Stability AI
# API_URL = "https://api.stability.ai/v2beta/stable-image/generate/sd3"
# API_KEY = "sk-..."
# MODEL = "sd3-large"

# Example: Any OpenAI-compatible endpoint
# API_URL = "https://your-provider.com/v1/images/generations"
# API_KEY = "your-key"
# MODEL = "your-model"

Same for vision analysis in scripts/analyze_image.py.

2. Generate images

# Single image
python scripts/generate_image.py "A portrait at dawn in a forest" out.jpg

# With style reference + auto imprint
python scripts/generate_image.py --image ref.jpg \
  "Same aesthetic, different subject" out.jpg --postprocess

# Multiple references
python scripts/generate_image.py --image face.jpg --image scene.jpg \
  "Place person from image 1 into scene from image 2" out.jpg

# Batch generation
python scripts/batch_generate.py --image ref.jpg prompts.txt ./output

# Post-process existing image
python scripts/postprocess.py raw.png final.png --mode imprint --intensity 0.7

# Analyze an image
python scripts/analyze_image.py photo.jpg "Describe this"

Banner Example

This banner was generated entirely with the pipeline below — 9 panels via image generation, composited with PIL, imprint post-processed:

python scripts/gen_panels.py           # generate 9 panels
python scripts/composite_banner.py     # composite into 21:9 grid
python scripts/postprocess.py \
  output/nous-branding-banner-composite-raw.jpg \
  output/nous-branding-banner-composite-imprint.png \
  --mode imprint --intensity 0.4

Pipeline

prompt + [--image refs] → Image API → raw output → Imprint post-process → final
                                              ├─ risograph grain
                                              ├─ xerox thresholding
                                              ├─ registration offset
                                              ├─ plate wobble
                                              ├─ CRT scanlines
                                              ├─ ink bleed
                                              ├─ paper texture
                                              ├─ palette compression
                                              └─ ... 6 more effects

Requirements

Requirement	Details
Image generation API	Any OpenAI-compatible API, local model, or custom backend. Configure endpoint + key in scripts/generate_image.py
Vision API	Any vision-capable model. Configure in scripts/analyze_image.py
Python libs	Pillow + numpy (pre-installed with Hermes)

Project Structure

.
├── SKILL.md                     # Hermes skill definition
├── README.md                    # this file
├── .gitignore                   # prevents key/image leaks
├── banner.png                   # Example banner output
├── scripts/
│   ├── generate_image.py        # Image gen + --image refs + --postprocess
│   ├── batch_generate.py        # Batch gen with optional shared reference
│   ├── analyze_image.py         # Vision analysis
│   ├── postprocess.py           # 14-step analog print pipeline
│   ├── gen_panels.py            # Generate banner panel images (no text)
│   ├── gen_anime_portal.py      # Generate 4 poster tiles
│   └── composite_banner.py      # Composite panels into 4K 21:9 grid + text
├── references/                  # Style docs, design language, lane audits
│   ├── nous-branding-skill.md   # Upstream Nous branding style guide (v2.1)
│   ├── design.md                # Core design principles
│   ├── design-language.md       # Visual language spec
│   ├── pitfalls.md              # Known gotchas
│   ├── reference-index.md       # Reference image index
│   ├── reference-catalog.md     # Full reference catalog
│   ├── identity-constants-template.md
│   ├── v8-pipeline-notes.md     # Legacy pipeline docs
│   ├── v9-imprint-pipeline.md   # Current imprint pipeline spec
│   ├── v10-*.md / v11-*.md     # Style lane audits & candidates
│   └── x-posts/                 # Nous X/Twitter reference screenshots
├── nous-assets/                 # Style reference images (from upstream)
│   ├── references/              # Character & style reference PNGs/JPGs
│   └── palettes/                # Color palette references
└── output/                      # Final images

Why MultiModel?

The upstream projects require Codex CLI + ChatGPT Plus for image generation. This fork rewrites the execution layer to work with any model:

Original	This Fork
codex exec --image (vision)	analyze_image.py / any vision model
codex exec -s workspace-write (image gen)	generate_image.py → any image API
--mode imprint post-process	Same postprocess.py, model-agnostic
ChatGPT Plus subscription required	Any backend — swap the API, keep the pipeline

License

MIT — forked from plntrprotocol/hermes-theia-codex-vision and plntrprotocol/nous-branding. All upstream credit to Llovos/plntrprotocol.