ComfyUI-Spellcaster

ComfyUI-Spellcaster

Architecture-aware nodes for AI image generation. Auto-detects your model architecture (SD 1.5, SDXL, Illustrious, ZIT, Flux Dev, Flux 2 Klein, Chroma), loads the correct CLIP and VAE, enhances prompts via local LLM, and samples with optimal settings — all automatically.

Part of the Spellcaster ecosystem. ONE SOURCE OF TRUTH: the same architecture definitions power the GIMP plugin, Darktable plugin, and Wizard Guild.

Nodes

Spellcaster Loader (Auto-Arch)

Drop in any model and it figures out the rest. Detects architecture from the filename, loads the right CLIP encoder(s) and VAE automatically.

• Checkpoint models (SD 1.5, SDXL, Illustrious, ZIT): single-file load
• Separate loaders (Flux Dev, Klein, Chroma): auto-selects UNET + correct CLIP type + correct VAE
• Klein CLIP auto-detect: picks qwen_3_8b for 9B models, qwen_3_4b for 4B models
• Flux dual CLIP: loads both clip_l and t5xxl with correct type
• Override inputs: force architecture, CLIP, or VAE if auto-detect gets it wrong

Outputs: MODEL, CLIP, VAE, arch_key (string for downstream nodes)

Spellcaster Prompt Enhance (LLM)

Sends your prompt to a local LLM (KoboldCpp, Ollama, or any OpenAI-compatible server) for rewriting, tuned per architecture.

• Booru tag style for SD 1.5 / SDXL / Illustrious: comma-separated tags with quality boosters
• Natural language for Flux / Klein / Chroma: descriptive sentences, no tag spam
• Architecture guidance: the LLM receives expert rules specific to each model
• Graceful fallback: if the LLM is offline, returns your original prompt unchanged

Inputs: prompt text, arch_key from Loader, LLM URL, enable/disable toggle

Spellcaster Sampler (Auto-Config)

One sampler node that works for every architecture. Auto-selects the correct sampling pipeline and fills in optimal defaults.

• Standard KSampler for SD 1.5, SDXL, Illustrious, ZIT, Flux Dev
• SamplerCustomAdvanced + CFGGuider for Flux 2 Klein (automatic)
• Auto-defaults: steps, CFG, sampler algorithm, scheduler all populated from architecture config
• Override anything: steps, CFG, sampler, scheduler, denoise all accept manual overrides (0 = use default)

Spellcaster Output (Privacy)

VAE decode + save with metadata stripping enabled by default.

• Decodes latent to image
• Strips generation parameters from PNG metadata (no prompt leaks)
• Clean timestamped filenames
• Toggle metadata stripping on/off

Install

Universal installer one-liner

If you have Python 3.10+, the fastest path is the LaboratoireSonore Universal Installer:

python install.py

install.py at the repo root is a thin shim for the LaboratoireSonore Universal Installer. It fetches the latest installer, picks the right artifact for your platform, and drops the node pack into your ComfyUI custom_nodes/ directory.

ComfyUI Manager (recommended for ComfyUI users)

Search for "Spellcaster" in ComfyUI Manager and click Install. ComfyUI Manager pulls directly from the ComfyUI Registry listing.

Comfy CLI

comfy node install comfyui-spellcaster

Manual (git clone)

cd ComfyUI/custom_nodes
git clone https://github.com/laboratoiresonore/ComfyUI-Spellcaster.git

Restart ComfyUI. No pip dependencies required.

Via Spellcaster Installer

If you're using the Spellcaster installer, the node pack is installed automatically.

Workflow Templates

Drag these into ComfyUI to get started:

• example_workflows/spellcaster_txt2img.json — text-to-image with all 4 nodes
• example_workflows/spellcaster_img2img.json — image-to-image with prompt enhancement

Supported Architectures

| Architecture | Loader | Sampler | Negative Prompt | Default Steps | Default CFG | |---|---|---|---|---|---| | SD 1.5 | checkpoint | KSampler | Yes | 25 | 7.0 | | SDXL | checkpoint | KSampler | Yes | 30 | 6.5 | | Illustrious | checkpoint | KSampler | Yes | 28 | 5.5 | | ZIT (turbo) | checkpoint | KSampler | Yes | 6 | 2.0 | | Flux 1 Dev | unet+clip+vae | KSampler | No | 25 | 3.5 | | Chroma | unet+clip+vae | KSampler | No | 25 | 3.0 | | Flux 2 Klein | unet+clip+vae | CustomAdvanced | No | 4 | 1.0 | | Flux Kontext | unet+clip+vae | KSampler | No | 25 | 3.5 |

License

NOPE - Free Software

---

Part of the Laboratoire Sonore ecosystem

These nodes are extracted from Spellcaster — the GIMP / Darktable / DaVinci Resolve / chat-UI plug-in suite that hides 24 ComfyUI custom-node packs and 8 architectures behind 69 one-click tools. Same architecture definitions power both the plug-in and these nodes — one source of truth, two deployments.

Le Laboratoire Sonore maintains three public projects:

• 🪄 Spellcaster — AI image generation, hidden behind one menu
• 🔧 ComfyUI-Spellcaster — these architecture-aware nodes (this repo)
• 🎚️ BeatWeaver — DJ overlay tool for non-musicians (BPM + key auto-detect, MIDI controller integration, bundled offline neural voice via Piper TTS)

Open-source when we can, always local.