BD Qwen T2I Custom

Most ComfyUI text encoders are a black box: prompt in, conditioning out, no opinions allowed. BD Qwen T2I Custom is the exception - it's a Qwen-Image text-to-image encoder that hands you the entire system prompt to rewrite. If you've ever wanted to tell the model "describe what I asked for in this very specific way before you draw it," this is the node.

What it is

Of the four Qwen encoders in the BrainDead pack, this is the odd one out and the simplest: no reference images, no VAE, no latent output. Just clip + prompt + system_template in, one CONDITIONING out. It's built for pure text-to-image with a Qwen-Image checkpoint, where you want to inject custom behavior into how the model interprets your prompt.

How it works

Qwen-Image is an autoregressive-ish diffusion model that works like a chat: your generation instruction lives inside a chat template, and the system message shapes how the model reads the user message. The default template here is instructive:

Describe the image by detailing the color, shape, size, texture, quantity,
text, spatial relationships of the objects and background:

So by default the model is primed to be a careful, descriptive renderer - you type a rough idea and it fills in the specifics. That's genuinely useful for short prompts: the template is doing the prompt-engineering for you.

But the whole point is that the template is yours. Want an art-direction frame? "You are a cinematic lighting director; always specify golden-hour lighting and shallow depth of field." Want a consistent style across a batch? Bake the style guide into the system message. The prompt field is where your actual generation text goes; the template is where the personality lives.

The inputs

• clip - your Qwen-Image CLIP (or a Qwen-Image-Edit CLIP, which also works for plain generation).
• prompt - the actual thing you want generated.
• system_template - the multiline system message. The default Qwen generation format is <|im_start|>system … <|im_end|><|im_start|>user {}<|im_end|><|im_start|>assistant, and you should keep that skeleton intact when you edit - change the content, not the delimiters, unless you enjoy broken conditioning.

One output: conditioning, which feeds straight into a KSampler. That's the entire graph.

Where it fits

Honestly? This node makes the most sense inside a bigger BrainDead pipeline - say, a character workflow where you generate a base image with a custom template, then hand it to BD Qwen Character Edit or Identity Lock for the edit pass. Used alone it's a nice Qwen-Image front end for people who like their system prompts verbose. If you're happy with ComfyUI's built-in Qwen-Image CLIP text encoder, this won't change your life - but if you're tired of re-typing the same art-direction boilerplate into every prompt, the template saves it for you.

Installing it

ComfyUI Manager: search "BrainDead" → install. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/BizaNator/ComfyUI-BrainDead
cd ComfyUI-BrainDead
pip install -r requirements.txt

Restart and it's under 🧠BrainDead/Character. The Qwen-Image model weights are on you - this is an encoder, not a checkpoint. One tip: because Qwen templates are chat-like, keep your instruction in the system message and the subject in the user message; models tend to follow that split much more reliably than a wall of prose in one field.