AI Character Lookdev (Bible) v13.0 - Akki

In the AkkiNodes "script to screen" pipeline, this is the Character Art Department. You've already written the screenplay, broken it into shots, and picked a character with the Asset Selector. Now you need a prompt that actually renders that character the way a character artist would design them - not "a woman in a coat" but a full art-directed description with props, costume, and physical detail you can drop into a checkpoint. That's the job of AI Character Lookdev (Bible) v13.0.

What it actually does

It's a three-stage LLM pipeline with a deterministic Python tail - the author's "hybrid multi-stage AI + Python enforcer" architecture. Stage 1 is a "Creative Artist" that drafts a rich character description from the bibles and the story. Stage 2 is a "Ruthless Editor" that trims the fat. Stage 3 is a "Technical Assembler" that formats the final output. Then shared utility code extracts the tagged main output so you get a clean prompt even when the model rambles.

That's the philosophy of this whole pack in one node: let the LLM be creative, then use boring, reliable code to make sure the result is actually usable. You'll see a full_llm_process_log output that shows you the whole conversation - genuinely handy the first time the prompt comes out weird, because you can see exactly which stage did what.

The inputs that matter

Most of the required inputs are STRING wires from earlier Akki nodes: world_bible, character_bible, story_or_script, shot_list_csv, and selected_character_name. The llm_model comes from a loader node (GGUF via the pack's own LLM Loader, or LM Studio).

The interesting ones are the optional drop-downs: ethnicity, age_range, body_type, hair_color, hair_style. They default to "Default (From Bible)" and are populated from wildcard files in the pack, so you can override a specific look without editing the bible. Leave them alone unless you have a reason; the bible is usually enough.

Sampling-wise: temperature 0.7, top_p 0.95, max_tokens 2048. Crank temperature if you want more variety across generations, drop it if the character keeps drifting.

Outputs and where they go

• character_prompt - the finished lookdev prompt. Feed this into a text encoder / prompt node, or save it to disk with a Generic File Saver.
• character_name - which character it built (handy in loops).
• full_llm_process_log - the raw pipeline output for debugging.

The character_prompt output is what you'll actually render from. It's designed to be stable enough that repeated renders of the same character stay recognizably the same person - which, as anyone who's fought with character consistency knows, is the whole game.

Installing and running it

The pack installs as a normal custom node:

cd ComfyUI/custom_nodes
git clone https://github.com/routhakash/AkkiNodes-LLM-Suite-for-ComfyUI

Or search "AkkiNodes LLM Suite" in ComfyUI Manager. On Windows, right-click install.bat and run as administrator once - it activates your ComfyUI venv and installs llama-cpp-python 0.3.9 and openai. Then drop a GGUF into ComfyUI/models/llms/, restart, and plug an LLM Loader into this node.

Where people get burned: this node does nothing without an upstream bible and CSV, so if you're testing it in isolation you'll get nonsense. And a small GGUF will produce weak lookdev - the quality of these agents is capped by the model you feed them. Budget a 7B+ model if you want prompts that don't read like a drunk first draft.