🖌️ SDXL Text Encoder Simple

Here's the thing most people don't realize about SDXL: the text encoder doesn't just read your prompt, it reads the resolution. SDXL was trained with aspect-ratio and crop information baked into the conditioning - which is why generating at 1344×768 feels different from 1024×1024 even with the same prompt. Core ComfyUI exposes all that as CLIPTextEncodeSDXL, a node with six size fields and a text box. SDXL Text Encoder Simple is that same shape, renamed, in this pack's menu: same fields, same single conditioning output.

If you've never touched the six size dials, you've been lucky. Let's map them, because they're less scary than they look:

• width / height - the resolution of the image itself.
• target_width / target_height - the resolution the model is actually generating at (the KSampler's latent size).
• crop_w / crop_h - crop-offset conditioning, used by SDXL's inpainting and aspect-ratio bucketing.

The rule of thumb: target_width and target_height should match your sampler's output resolution, and width/height should match the source image when you're doing img2img-style work. Get the target pair right and the rest usually falls into place.

The twist in this node

The defaults. Core ComfyUI's SDXL encoder defaults to 1024×1024 across the board. This one defaults all six to 4096. That's almost certainly a leftover from the author's own upscaling workflows (the pack is a personal grab-bag), and it's a trap if you blindly hit run: the conditioning will tell the model you're generating at 4096×4096 while your KSampler is actually outputting 1024×1024. The mismatch is exactly the kind of subtle thing that produces odd composition and nobody knows why.

So: treat the defaults as wrong until you've set them. Match target_width/target_height to your sampler, keep the others consistent with the image, and this node behaves like the core one.

The rest

The other two inputs are the familiar ones: text (the prompt, multiline) and clip (wire from your checkpoint or a dual-CLIP loader). Output is a single CONDITIONING that feeds the sampler's positive (or negative) input. It's an SDXL-family node - SD 1.5, Flux, and Qwen-Image all use different encoder shapes, so don't reach for this one outside SDXL workflows.

Install

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

or ComfyUI Manager → search ComfyUI-NeuralMedia, then restart ComfyUI. Pack by YarvixPA (the FLUX.1-Fill-dev-GGUF quantizer); if a manual clone 404s, Manager's registry entry is the reliable route.

The honest take

This node is a rename-plus-defaults of functionality ComfyUI already ships. If you know your way around CLIPTextEncodeSDXL, there's nothing here you're missing - except that 4096 default, which is the one thing you should actively distrust. Use it if you're building a self-contained NeuralMedia workflow; otherwise core's node does the same job with saner starting values.