shootthesound/comfyui-kreareason — explained in plain English
Analysis updated 2026-05-18
Expand a short prompt into a detailed one before generating an image with Krea 2
Extract only the background or style from a reference image, not its subject
Combine your own prompt with elements pulled from a reference photo
Swap in an uncensored encoder to avoid refusals when describing images
| shootthesound/comfyui-kreareason | aarav90-cpu/fetchit | adammartinez271828/x4-analysis | |
|---|---|---|---|
| Stars | 10 | 10 | 10 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | easy | easy |
| Complexity | 3/5 | 2/5 | 2/5 |
| Audience | designer | developer | general |
Figures from each repo's GitHub metadata at analysis time.
This is a custom add-on node for ComfyUI, a visual tool many people use to build image generation workflows with AI models. It works specifically with a model called Krea 2 and adds one extra building block, or node, to a workflow. Krea 2's text encoder is actually a full vision and language model, meaning it can both read and understand images and generate text, not just turn words into numbers the way a typical text encoder does. This project takes advantage of that by letting the encoder rewrite a short, simple prompt into a longer, more detailed one before that prompt is used to generate an image. It can also look at a reference image, describe what it sees in words, and then use that description as part of the prompt. The image reference feature works in three visible steps. First, the model describes the reference image in detail. Second, the user can choose to remove specific things from that description, such as people or the main subject, while keeping the background, style, or color scheme. Third, whatever remains gets combined with the user's own prompt, with the user's prompt taking priority. This gives someone much more control than simply feeding a whole reference image into an image generator and hoping it captures the right details. Installing the node involves either using ComfyUI's built in custom node manager, or manually cloning the repository into ComfyUI's custom nodes folder and restarting the program. Beyond that, using the node requires downloading specific text encoder, diffusion model, and VAE model files and placing them in ComfyUI's model folders. The README also recommends using a modified version of the underlying language model that has had its built in safety refusals removed, since the standard model can sometimes refuse to describe an image or quietly change what it writes, which would interfere with generating the requested output. This is presented as giving the user more direct control over what gets generated, rather than the model second guessing the request.
A ComfyUI node that lets Krea 2's vision language encoder rewrite prompts or describe reference images before generation.
Mainly Python. The stack also includes Python, ComfyUI, Krea 2.
Setup difficulty is rated moderate, with roughly 1h+ to a first successful run.
Mainly designer.
This repo across BitVibe Labs
double-check against the repo, no cap.