ComfyUI-Flux2Klein9Mod

No-training "RefMod" reference adapters for FLUX.2 and Klein9.

RefMods pre-encode reference images and datasets into compact, reusable .safetensors adapters stored in models/refmods-klein9/. Rather than loading and VAE-encoding raw 4K/1024px image files on every generation run (which adds 4096+ tokens and heavy attention latency), RefMods inject pre-computed reference latents directly into the FLUX.2 / Klein9 2D RoPE sequence with instant loading, zero VAE overhead, and full multi-character scaling.

---

🔗 Links & Resources

• More Info & Documentation: Hugging Face Dataset (klein9)
• RefMods Library Browser: RefMods Browser on Hugging Face Spaces

---

📦 Features

• Pre-encoded Latent Storage: Stored in ComfyUI/models/refmods-klein9/.
• Zero VAE Generation Latency: Latents are loaded directly from disk without runtime VAE encoding.
• Configurable Token Budgets: Native resolution or spatial pooling to target budgets (e.g. 512, 1024, 2048 tokens).
• Multi-Subject Scaling: Load up to 8 distinct RefMods simultaneously (mod_1 through mod_8) with independent strengths.
• Manifold Blending: Smooth weakening and strength attenuation using spatial low-pass filtering along the latent manifold instead of noise degradation.
• Universal Compatibility: Works with any standard ComfyUI CONDITIONING graph for FLUX.2, Klein 4B, Klein 8B, and Klein 9B models.

---

🚀 Installation

1. Navigate to your ComfyUI custom_nodes folder:

   cd ComfyUI/custom_nodes

2. Clone or place this repository:

   git clone https://github.com/malcolmrey/ComfyUI-Flux2Klein9Mod.git

3. Ensure the model folder exists:

   ComfyUI/models/refmods-klein9/
   

4. Restart ComfyUI.

---

🧩 Nodes

Node Name	Category	Description
Load Klein9 RefMods (FLUX.2) (Klein9RefModsLoader)	model/conditioning/klein9	Loads 1–8 RefMod .safetensors files from models/refmods-klein9/ with individual strengths (0.0–3.0) and repeat counts.
Apply Klein9 RefMods (FLUX.2) (Klein9RefModApply)	model/conditioning/klein9	Injects the loaded RefMod bundle into standard ComfyUI CONDITIONING (reference_latents).
Extract Klein9 RefMod (FLUX.2) (Klein9RefModExtract)	model/conditioning/klein9	In-graph VAE encode & token compression of input images directly into a saved .safetensors file in models/refmods-klein9/.
Load RefMod Folder (Images) (Klein9RefModFolderLoader)	model/conditioning/klein9	Batch loads an entire folder of reference images for extraction.

---

🛠️ Graph Wiring

[ CLIPTextEncode ] (Prompt) 
         │
         ▼
[ Apply Klein9 RefMods (FLUX.2) ] ◄── [ Load Klein9 RefMods (FLUX.2) ]
         │ (conditioning out)          (Selects models from models/refmods-klein9/)
         ▼
[ CFGGuider / SamplerCustomAdvanced ]

---

💻 CLI Extraction Tool

You can also pre-encode image datasets using the CLI:

python custom_nodes/ComfyUI-Flux2Klein9Mod/extract_klein9_mod.py \
  --name fk9_billieeilish_v1_refmod \
  --dataset-dir C:/Development/ai-toolkit/datasets/billieeilish \
  --concept-type identity \
  --token-budget 1024 \
  --output-dir C:/Development/ComfyUI/models/refmods-klein9/