LLM Writer – Krea2 & Illustrious (LM Studio)
Turn a short idea, or up to 8 reference images, into a ready-to-use prompt with a local LLM running in LM Studio, then generate with Krea2 or Illustrious from the same workflow.
FEATURES
• One switch for Krea2 / Illustrious: the unused branch is muted, and the LLM writes natural language for Krea2 or Danbooru-style tags for Illustrious.
• Up to 8 reference images, each with toggles for what to take from it: Accessories, Action, Background, Camera, Clothing, Colors, Expression, Hair, Lighting, Person, Pose, Style. Only the enabled aspects are used (none on = everything except camera).
• Hair, expression or accessories from one image can be applied to the person from another image without changing their identity.
• Images are described automatically before generating. Each description is tied to its photo and toggles: change the photo or the toggles and it is redone on the next Generate. You can also describe one image at a time, or clear a description with its ✖.
• Output style, camera framing, Clothing and Participants selectors (1–8, Auto or Custom).
• Clothing selector: LoRA (clothing from the LoRA triggers; the LLM decides if there are none), Photo (clothing from the image with the Clothing toggle wins; picked automatically when you turn that toggle on) or Nude.
• LLM passes: 1 pass, or 2 passes where a second call reviews the prompt against all your settings and fixes or improves it. Very short answers are automatically retried.
• ✏️ Modify: ask for a small change (e.g. "make the hair red") and the LLM edits only that.
• 🧹 Clear resets the idea, prompts, images, toggles and all selectors to their defaults.
• LoRA trigger words from Lora Manager are inserted automatically (long character descriptions are blended into the scene instead of pasted).
• Live LLM activity log in the Config node (model load, each description, passes, timings, errors).
• 🎲 Random seed button and a "Random on every run" switch.
• Negative embeddings picker for Illustrious.
• Base image i2i with denoise for both branches. Leave it empty for text to image.
• Size node with aspect ratio + megapixels and a live pixel preview.
• Save Image node with an auto save on/off switch and a manual 💾 Save button.
• The model is unloaded from LM Studio after generating (optional), freeing VRAM for the image model.
• Built-in notes (Setup / How it works) in English or Spanish, following the workflow language.
REQUIRED CUSTOM NODES
• LD PromptWriter (LM Studio) 1.1.3 or newer – search "LD PromptWriter" in ComfyUI Manager, or https://github.com/damidero/ComfyUI-LD-PromptWriter-LMStudio
• ComfyUI-Lora-Manager – https://github.com/willmiao/ComfyUI-Lora-Manager
• ComfyUI-Custom-Scripts (PlaySound, optional) – https://github.com/pythongosssss/ComfyUI-Custom-Scripts
MODELS
• Krea2 UNET: Krea2-Turbo-HD-V1-int8_convrot – https://huggingface.co/wikeeyang/Krea2-Turbo-HD-V1
• Krea2 CLIP qwen3vl_4b_fp8_scaled + VAE qwen_image_vae – https://huggingface.co/Comfy-Org/Krea-2
• Illustrious: WAI Illustrious SDXL v17 – https://civarchive.com/models/827184/wai-illustrious-sdxl
(or Illustrious Realism – https://civarchive.com/models/2354628/illustrious-realism)
• LLM (LM Studio): Qwen3-VL-8B-Instruct abliterated, GGUF Q4_K_M + its mmproj file – https://huggingface.co/noctrex/Huihui-Qwen3-VL-8B-Instruct-abliterated-GGUF
Faster: the 4B version – https://huggingface.co/noctrex/Huihui-Qwen3-VL-4B-Instruct-abliterated-GGUF
HOW TO USE
1. Open LM Studio, start the local server and load the vision model (with its mmproj).
2. Set the server URL in the LM Studio Config node, pick the model and Krea2 or Illustrious.
3. Write an idea and/or load reference images, and choose what to take from each one.
4. Press ✍️ Generate prompt, check the text and queue the workflow.
TIPS
• Use an Instruct model, not Thinking. Reasoning / MTP models often return empty or cut-off prompts. If yours "thinks", turn off Reasoning in LM Studio.
• Pick a fixed model in the Config node instead of "auto", and set max_tokens to 2048.
• Without the mmproj file the model can't see images.
• If the positive box already has text, the workflow uses it as is and does not call the LLM. Clear it to generate a new one.
• Krea2 LoRAs don't work with Illustrious and vice versa. Swap them when you change engines. LoRAs made for other models (Flux, Qwen-Image, Z-Image, SDXL…) fail on Krea2 with a shape error.
• In i2i the LLM doesn't see the base image. To change clothing or pose from it, raise denoise to 0.75–0.9, or load the same photo in one of the 8 reference images with the right toggles.
FEEDBACK
If something errors or works wrong, please leave a comment or feedback here so I can fix it. Screenshots of the error or the LLM activity log help a lot.
Description
v1.0 – First release. LM Studio prompt writer for Krea2 and Illustrious: turn an idea or up to 8 reference images into a prompt, choosing what to take from each one (action, pose, person, clothing, background, style, camera). Includes Modify for quick edits, LoRA trigger insertion, negative embeddings, base image i2i, a Size node and a Save Image node with auto save on/off.
FAQ
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.
