Artfat LLM Prompter v2

Artfat LLM Prompter v2

One ComfyUI node that turns a brief plus reference images into a finished image prompt with a resident llama.cpp VLM, and encodes the result to CONDITIONING.

(reference gallery / image_1 / image_2 / text) -> LLM -> prompt -> CLIP -> CONDITIONING

Node: Artfat LLM Prompter v2 (ArtfatLLMPrompterV2), category artfat/llm.

It is a fork of the Artfat Video Prompter pack: the same tabbed panel and reference gallery, with every time-based part removed (task modes, scene length, the shot/cut table, the native MiniMax-H3 conditioning and its media transports) and the image-side features folded back in (composite/batch captioning, prompt-list mode, image_1/image_2 and queue sockets).

Install

1. The node. ComfyUI Manager → Custom Nodes → search Artfat → Artfat LLM Prompter v2 → Install. Not found or blocked by your Manager security level? ComfyUI Manager → Install via Git URL → https://github.com/artfat-creator/artfat-comfyui-llm-prompter-v2 (or clone it into ComfyUI/custom_nodes/ and run python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\artfat-comfyui-llm-prompter-v2\requirements.txt).

2. llama-cpp-python (the LLM engine). ComfyUI Manager runs the node's install.py, which picks the GPU build matching your torch CUDA (Windows and Linux; CPU build as a fallback). By hand: python_embeded\python.exe ComfyUI\custom_nodes\artfat-comfyui-llm-prompter-v2\install.py with ComfyUI closed. If it still runs on CPU, the console prints the exact command; the commands below do the same.

   Windows + NVIDIA, ComfyUI portable with torch cu130 and Python 3.13 (the current portable):

   python_embeded\python.exe -m pip install --force-reinstall --no-deps https://github.com/JamePeng/llama-cpp-python/releases/download/v0.4.3-cu130-win-20261017/llama_cpp_python-0.4.3+cu130-cp313-cp313-win_amd64.whl
   

   Linux + NVIDIA (torch cu130; Python 3.12 shown, use your ComfyUI venv's python and its cpXXX):

   python -m pip install --force-reinstall --no-deps https://github.com/JamePeng/llama-cpp-python/releases/download/v0.4.3-cu130-linux-20261017/llama_cpp_python-0.4.3+cu130-cp312-cp312-linux_x86_64.whl
   

   Other CUDA / Python versions: pick the matching wheel on https://github.com/JamePeng/llama-cpp-python/releases (cuXXX = your torch CUDA, cp3XX = your Python). No NVIDIA GPU: pip install llama-cpp-python (CPU, slow).

Put GGUF models in ComfyUI/models/LLM/ (a matching mmproj-*.gguf next to them for vision). ComfyUI/models/text_encoders is scanned too, because GGUF VL encoders conventionally live there.

Panel

Five tabs over the node's native widgets:

| Tab | Holds | |---|---| | Prompts | final prompt, negative, prompt-list mode | | Shot | brief, rig / size / angle / lens / pose / wardrobe, scene, light, action | | Refs | reference gallery with sets, roles and custom labels | | Presets | format + task + persona presets, their editable text, persona image, identity block, config save/load | | Advanced | model, mmproj, chat handler, context, sampler |

The panel creates no widgets. Every widget is declared in image_nodes.py INPUT_TYPES and lives in that frozen positional order; web/llm_prompter_v2.js only groups, styles and hides them. ComfyUI serializes widget values as a positional list, so a widget created in JS shifts every saved value after it — built this way, that drift is impossible by construction. New widgets go at the END of required, never in the middle.

Modes

composite (default) — every image goes into one call and produces one prompt. The reference roles are handed to the writer as a numbered [REFERENCES] block (<Image 1> = identity, ...).

batch — each image gets its own caption in its own call. Captions are joined in prompt and split across prompt_list. Only one image is in each call, so the numbered role block is not sent; the persona identity block still is.

prompt list (batch_mode) — the LLM is skipped entirely and batch_prompts is read as a list: one prompt per line, or split on numbered 1. / 2. markers so a multi-line prompt stays whole. # lines are comments. Wire positive_list to KSampler.positive and one Queue renders the whole list.

Persona identity

reference_image (socket) or the in-node persona image (Presets tab) is describe-only: it is described once with ref_preset, the text lands in ref_description, and that block is prepended to every prompt. It is never sent alongside the frames. Sending it with each frame made the model re-describe the face per call, so one batch came back with several different faces.

The description is reused while the image and the preset hash the same. Tick refresh_ref to force a re-describe; edit ref_description freely, your text is what gets prepended.

Reference sets

A set is a whole alternate pool, not a filtered view of one: each set numbers from <Image 1> in its own right. Only the starred set reaches the writer. ＋ set (copy) duplicates the open set so you prune rather than refill. Entries with no set belong to set 1.

Type @ in the brief or action field to insert a reference tag from the active set, so the number in the text always matches the gallery.

Outputs

| Output | Notes | |---|---| | positive / negative | CONDITIONING (needs clip; unconnected -> None) | | prompt | the final text | | prompt_list | LIST: one entry per caption in batch mode | | image_1 / image_2 | passthrough | | queue | passthrough, to force execution order between prompter nodes | | clip | passthrough | | positive_list | LIST: one CONDITIONING per batch_prompts entry |

Presets

prompts/*.txt are format profiles (HOW to write); prompts/reference/*.txt are persona-identity profiles. A copy in ComfyUI/models/LLM/prompts/ (and .../prompts/reference/) overrides a bundled file of the same name, so a preset can be tweaked without touching pack files.

Task presets ("WHAT to do") are Python strings in presets.py.

The preset is the authority: picking one overwrites the matching text box. Select Custom to keep your own text.

Running next to the sibling packs

This pack is self-contained and is designed to sit beside Artfat LLM Prompter and Artfat Video Prompter. The three collide in exactly three global namespaces, and each is claimed separately here:

• HTTP routes — /artfat_llm_v2/... (aiohttp raises on a duplicate route)
• the model_management VRAM-cleanup guard — mm._llm_prompter_v2_hooked (each pack owns its own LLMEngine class, so a shared guard would leave the second pack's model resident)
• the folder_paths LLM key — registered additively: the path is appended if missing and the extensions are unioned, so nothing another pack registered is removed

Each pack also answers preset requests out of its own prompts/ directory, which is what its dropdowns list.

Three resident llama.cpp models can be loaded at once if three nodes run with different settings. force_offload on the nodes you are not using, or wire them through queue.

License

MIT. llama_core.py is adapted from lihaoyun6/ComfyUI-llama-cpp (MIT).