ComfyUI-KreaReason

ComfyUI-KreaReason

A single node that turns Krea 2's text encoder (Qwen3-VL-4B) into a built-in prompt enhancer and image-reference tool. Krea 2's "CLIP" is a full vision-language model with a working LM head, so it can rewrite a short prompt into a rich one, or read a reference image and write the prompt from it, before conditioning — all on the exact model that conditions, no second LLM.

Why this gives you more control than the standard image-reference flow. The usual path — Krea's own describe-and-encode, or embedding-based adapters — hands you whatever the model decides is important (usually the person or main subject) as a black box you can't steer, and you take it whole. Krea Reason breaks the reference into three visible, editable steps: it describes the image, lets you surgically remove exactly what you don't want (people, the main subject, everything except the background/style/colors/composition…), then combines what's left with your prompt — with your prompt leading. You see the exact text that becomes your conditioning and can keep, drop, or rewrite any part of it. So instead of "here's the whole image, roughly," you get "my subject, in that reference's lighting and setting, and nothing else." Pair it with an abliterated encoder (below) and nothing refuses or quietly softens your intent — total control end to end.

Krea Reason vs a standard image reference

Same reference image (a cowboy by a campfire at desert sunset) and the same prompt — "a victorian magician releasing a Ferrari F40 out of a magical box, ethereal." Left (Krea Reason, people + main subject (keep the scene)): the magician and the Ferrari, dropped into the reference's desert-sunset setting — the cowboy is gone, exactly as asked. Right (standard image reference): the reference's cowboy is dragged into the output and fights the prompt, so you get a gunslinger instead of a magician.

Install

ComfyUI-Manager (once it appears in the registry index): open Manager → Custom Nodes Manager, search "Krea Reason", click Install, then restart ComfyUI.

Manual (git clone): drop it into your custom_nodes folder and restart —

cd ComfyUI/custom_nodes
git clone https://github.com/shootthesound/ComfyUI-KreaReason

There are no extra Python dependencies — it rides on ComfyUI's built-in Krea 2 support. Restart ComfyUI (a full restart, not just a browser refresh) after installing, and the node appears under ShootTheSound/KreaReason. Drag workflows/Krea Reason vs Standard Image Ref.json onto the canvas to get going.

Models (Krea 2)

Load the encoder with a CLIPLoader set to type krea2. Everything below goes in the usual ComfyUI model folders:

Slot	File	Folder
Text encoder	abliterated Qwen3-VL-4B (recommended — see below) or stock qwen3vl_4b_bf16 / qwen3vl_4b_fp8_scaled	models/text_encoders/
Diffusion model	krea2_turbo_fp8_scaled.safetensors	models/diffusion_models/
VAE	a Wan 2.1 VAE, e.g. wan_2.1_vae.safetensors (see Notes)	models/vae/

The Krea 2 model files are on Comfy-Org/Krea-2; the recommended abliterated encoder is linked in the section below.

Node: Krea Reason (expand prompt + encode) (category ShootTheSound/KreaReason)

Text mode — type a prompt and it expands it into a detailed one, then encodes:

• mode = expand (default): the rewritten prompt is what gets encoded (impactful).
• mode = think: keeps your prompt and appends the generated text as <think> context (subtler).
• instruction: system instructi