Unblend Scene Book (JSON prompts)

Once you've got the Unblend two-character pipeline working, the next thought is always "now I want forty of these." Copy-pasting tagged prompts into the Dynamic Sampler and re-clicking char_points by hand for every scene is a great way to spend a day. UnblendSceneBook is the node that makes that a data problem instead of a clicking problem: your whole prompt collection lives in one JSON file, and the node hands you a scene by index.

How it works

The file is a scene book: one shared prefix (your [QUALITY]/[STYLE] block, the constant you don't want to retype) plus a list of scenes. Each scene is a body prompt, its own char_points, and an orientation hint:

{
  "prefix": "[QUALITY] masterpiece, best quality, very aesthetic\n[STYLE] detailed hair, very aesthetic",
  "prompts": [
    { "name": "beach midday",
      "prompt": "[SCENE] beach under a parasol\n[CHAR1] ningguang...\n[CHAR2] shenhe...\n[INTERACT] resting head on shoulder",
      "char_points": "0.3,0.45; 0.7,0.45",
      "horizontal": true }
  ]
}

The node reads it, pulls out the scene at index, and returns the pieces separately: prefix, prompt (scene body only), full_prompt (ready to feed a sampler), char_points, name, count, and horizontal - a boolean for whether the scene is meant to be landscape (true) or portrait (false). That last one is the nice touch: wire it into a latent-size switch and the composition node changes aspect ratio per scene automatically.

You don't have to keep everything in a file, either. The inline_json input lets you paste the JSON straight into the node, which is great for testing a scene before committing it to disk. And the file parser is forgiving about naming - prompts/scenes, prompt/body, horizontal/landscape, or a "16:9"/"9:16" ratio string all work, so you can generate the book with a script without fighting the schema.

The inputs that matter

• json_path - the file path (or leave it and paste into inline_json).
• index - which scene, 1-based. This is where the Line Counter plugs in: put its line_number here with wrap on and every run advances to the next scene, wrapping around when the book ends.
• wrap - true wraps the index by the book's length (for an incrementing counter); false clamps to the first/last scene.
• default_horizontal - the fallback orientation for scenes that don't set one.

Install

Same pack, same two options:

cd ComfyUI/custom_nodes
git clone https://github.com/LatentDesireAI/ComfyUI-Unblend
pip install -r ComfyUI-Unblend/requirements.txt

Or ComfyUI Manager, search "Unblend", Install, restart.

When things go wrong

The info output is a tidy i/N: scene name [horiz|vert] line - if it doesn't say what you expected, your index is wrong, not the book. An empty book throws "no prompts in the JSON", which usually means a typo in the key or a missing prompts array. And remember index is 1-based while the Line Counter starts at 0 by default - misalign those and you'll perpetually be one scene off. If a scene comes back with no char_points, that scene's entry is a plain string (allowed, but it leaves the anchors empty) or you forgot to set them on that dict.