Maxomarai Pack for ComfyUI

A small suite of ComfyUI tools by Maxomarai, designed to work together.

One repo, one install — two tools:

Tool	What it does
The Last Model Switcher	One-click model switching with auto-configured CLIP/VAE/resolution/sampler, plus AI-powered model identification and prompt enhancement.
Wayfinder	Frontend-only visual aid: type-colored drag links, connected-node highlighting, breathing glow on palette-colored nodes.

---

Install

Option 1 — Git clone (recommended for now)

cd ComfyUI/custom_nodes
git clone https://github.com/maxomarai/maxomarai-pack.git

Restart ComfyUI. Both tools activate on next browser load.

Option 2 — ComfyUI Manager

Open Manager → Install Custom Nodes → search for maxomarai-pack → Install. Restart ComfyUI.

Option 3 — ComfyUI Registry

comfy node registry-install maxomarai-pack

---

The Last Model Switcher

One node replaces the usual Load Checkpoint → CLIPTextEncode × 2 → EmptyLatentImage → KSampler plumbing. Pick a model; CLIP, VAE, resolution, and sampler settings auto-populate. Outputs are ready to wire directly into a KSampler and EmptyLatentImage.

Supports SDXL, SD 1.5, Flux 1 (Dev / Schnell), Flux 2, SD 3.5, Chroma, HunyuanVideo, and more — auto-detected from the safetensors tensor-key fingerprint.

Outputs

Output	Wire to
model	KSampler
vae	VAE Decode
positive	KSampler (positive) — FluxGuidance is baked in for Flux models
negative	KSampler (negative) — empty for Flux, safe to keep connected
width / height	EmptyLatentImage
steps / cfg	KSampler
seed	KSampler

AI features

TLMS ships with two AI-assisted flows. Configure the provider once in Settings (see below); the node calls it on demand.

• 🔎 Identify Model (Model toolbar): sends the model's filename + tensor-key fingerprint to the AI and receives suggested sampler / scheduler / steps / CFG / guidance / negative-prompt support. Shown as a diff preview in the info panel — Apply commits to presets.json, Discard throws it away. Nothing is saved without your confirmation. A warning appears when using local/custom providers or when AI confidence is low/medium, since smaller models can hallucinate settings.

• ✨ Enhance Prompt (Prompt toolbar): rewrites the positive prompt in your chosen style (enhance / detailed / concise / creative / fix) or per a custom instruction. Shown as a preview; ✓ Apply replaces the prompt, re-clicking Enhance overwrites the preview.

Supported providers:

Provider	Needs API key	Notes
Anthropic (Claude)	✓	Get one at console.anthropic.com
OpenAI (GPT)	✓	Get one at platform.openai.com/api-keys
Ollama (local)	—	Default URL http://localhost:11434/v1/chat/completions. Run ollama serve and ollama pull <model>
Custom OpenAI-compatible	depends	LM Studio, vLLM, anything with /v1/chat/completions

Toolbars on the node

Four compact pill-shaped segmented toolbars replace stacked full-width buttons:

• Model — 🔎 Identify · 📡 Scan
• Prompt — ✨ Enhance · ✓ Apply · ⌄ Options · ⛶ Editor (Apply only appears when an enhanced prompt is pending; Options expands enhance_style + enhance_instruction)
• Seed — 🔀 Random · ↺ Reuse last
• Admin — ⚙ Settings · 🔌 Test · 📄 Presets · ↻ Reload

Editor

⛶ Editor opens a modal with big textareas for the positive and negative prompts, plus a read-only view of the info panel. Esc to cancel, Ctrl+Enter to save.

Settings

Exposed in ComfyUI Settings → Maxomarai → The Last Model Switcher:

Setting	Type	Default	Purpose
AI Provider	combo	anthropic	anthropic / openai / ollama / custom
AI Model	text	empty	Provider-specific model name — e.g. claude-haiku-4-5-20251001, gpt-4o-mini, qwen3:14b, llama3.2
API Key	text	empty	Stored in plaintext in comfy.settings.json. Not needed for ollama/custom.
Base URL	text	empty	For ollama/custom. Default Ollama URL: http://localhost:11434/v1/chat/completions

⚠️ Reasoning models (qwen3, deepseek-r1, qwq) work but consume many tokens on internal "thinking" before producing the JSON response. TLMS bumps max_tokens to 2000–3000 for AI flows to accommodate this.

---

Wayfinder

Pure frontend tweaks that make large graphs more navigable. No nodes. No workflow changes. Just better visual feedback.

Settings

Exposed in ComfyUI Settings → Maxomarai → Wayfinder:

Setting	Type	Default	Purpose
Type-colored drag link	toggle	on	While dragging a new connection, color the line with the socket's type color instead of ComfyUI's default green
Drag highlight	toggle	on	While dragging: dim unconnected nodes, glow source/target in their own color, pulse the primary link
Idle color glow	toggle	on	Palette-colored nodes get a subtle breathing glow even when you're not dragging
Node glow intensity	slider	1.0	Scales glow brightness/size (0.3–2.0)
Pulse speed	slider	1.0	Breathing pulse multiplier (0.3–3.0)
Override glow color	toggle	off	Use a fixed glow color below instead of each node's own palette color
Glow color	color	#64b4ff	Used when override is on
Override untyped link color	toggle	off	Use a fixed color for links/drag-lines with untyped sockets (STRING, INT, etc.)
Untyped link color	color	#99AA99	Used when override is on

---

Development

cd ComfyUI/custom_nodes
git clone https://github.com/maxomarai/maxomarai-pack.git
cd maxomarai-pack
# hack on nodes/the_last_model_switcher.py or web/js/*.js
# restart ComfyUI (Python change) or F5 (JS change) to see results

The frontend uses Preact + htm + Lucide icons loaded from esm.sh at runtime — no build step required.

---

Philosophy

Clean UI by default, power accessible on demand. Widgets you rarely touch (like enhance_style, custom instructions, LoRA stacks) collapse behind chevrons. Widget preferences persist in localStorage. Destructive AI suggestions always preview-then-confirm before touching config files.

---

License

MIT © 2026 Maxomarai

Changelog

See CHANGELOG.md.