In-Context-LoRA persona

In-Context LoRA is the technique behind a lot of the "consistent multi-panel" images you've seen floating around - a comic strip, a set of story beats, a product shown from several angles - all generated together so the style and subject stay locked across every panel, instead of drifting the way separate generations would. Getting an LLM to write good prompts for that kind of multi-scene generation is a different ask than writing a single-image prompt, and ic_lora_persona is a ready-made system prompt tuned for exactly that job.

Like this pack's other persona nodes, it doesn't generate anything itself - it produces the instructions an LLM node uses to generate the actual scene-by-scene prompt.

How it works

You tell it how many scenes you want and what the overall subject is, and it builds a system prompt that asks the downstream LLM to write that many coordinated scene descriptions - consistent characters, consistent style, distinct beats - rather than one generic caption per scene written in isolation.

The inputs and outputs that matter

• scene_num (default 4, minimum 1) - how many panels/scenes you want the LLM to write prompts for. This should match however many panels your actual image workflow is going to generate downstream - set it lower than that and you'll be short scenes; higher and you'll have prompts nothing consumes.
• theme (default: "comic book") - the overall subject and format. Swap this for whatever you're actually building.
• is_enable - standard bypass toggle.

Output: system_prompt_input (STRING) - wire it into this pack's LLM node's system-prompt slot. The LLM's actual response is what carries the per-scene prompts.

How to install it

Via ComfyUI Manager: search comfyui_LLM_party, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/heshengtao/comfyui_LLM_party.git

Then pip install -r requirements.txt from inside the pack folder using ComfyUI's own Python, and restart. This node itself needs nothing beyond the base install - no models, no keys - since it only returns text. Getting from that text to actual consistent multi-scene images is a separate step: you still need an In-Context LoRA (or equivalent multi-panel technique) wired up in your image-generation chain to make the visuals stay consistent, not just the prompts.

Common issues & troubleshooting

The LLM's scene count doesn't match scene_num. This node sets up instructions, it doesn't enforce them - a weaker or less instruction-following model may drift from the requested count. If that keeps happening, check what LLM you've got wired up before assuming the node is broken.

Prompts are coordinated but the resulting images don't look consistent. That's expected if you haven't got real In-Context LoRA (or similar) machinery downstream - this node only solves the prompt-writing half of multi-scene consistency. Matching visual style across scenes is a property of your image-generation setup, not of the text this node produces.

Output feels generic despite a specific theme. Same as with this pack's other persona nodes: results scale with how well the LLM you've wired up follows detailed instructions, not just with what you typed into theme.