🔡 Boogu Edit TextEncoder

This is a text encoder built for instruction-based image editing, not for plain text-to-image. If you've used Flux Kontext or Qwen-Image-Edit, you know the pattern: you hand the model a reference picture and a sentence describing the change, and it re-emits the image with that change applied. The whole family of those models works by pulling the reference image into the conditioning alongside your prompt rather than treating it as a separate control signal. Boogu Edit TextEncoder is the node that packs both halves - your instruction and your reference images - into a single conditioning bundle the sampler can use.

Mechanically that's exactly what it does. It takes your prompt, runs it through the clip you feed it, and separately runs your reference images through the vae to encode them into latent space, then concatenates the two so the model sees the picture and the instruction as one context. That's why it wants both a CLIP and a VAE - a normal text encoder only needs CLIP, but an edit encoder has to VAE-encode the reference too. The images input is a growing slot (add one, another appears), so you can stack multiple references. maxsize caps how large those references get resized to before encoding, which matters more than it looks: edit models are fussy about input resolution, and letting a big reference through unscaled is a common way to get skewed, offset output. There's a translate picker if you're writing your instruction in another language, and a seed that feeds the pack's dynamic-prompt support.

The outputs are the giveaway that this is a one-stop encoder: conditioning, width, height, and latent. You wire the conditioning and the latent straight into a KSampler - the latent is already sized to match your references, so you don't bolt an Empty Latent onto the front. The width and height come out as plain ints in case you want them for a later resize or a save path.

Where you'd reach for it: any time you're doing reference-driven editing and you're tired of stringing together a separate text encode, a VAE encode, and a reference-latent node by hand. It folds that little cluster into one box. The honest caveat that comes with every model in this class, worth knowing before you build around it: these editors re-emit the entire frame, so pixels you didn't ask to change drift a little, and faces are the first thing to wander across a chain of edits. That's not this node's fault - it's how in-context editing works - but it's why the standard heavy workflow still masks the region and stitches the edit back, rather than trusting a full-frame pass to leave everything else untouched.

Install rides with the pack. ComfyUI Manager, search SDVN_Comfy_node; or cd ComfyUI/custom_nodes && git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node, then pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt from your ComfyUI root and restart. The node itself is just glue - the actual editing weights (whatever base model this encoder is paired with, plus its CLIP and VAE) you load and download separately, the same as any Flux or Qwen edit workflow.

Troubleshooting is mostly a matching problem. The CLIP and VAE you feed it have to be the ones for the edit model you're driving; a mismatched VAE gives you garbled latents and muddy output. If your edits come out geometrically shifted or the proportions look off, pull maxsize down to a sane value and control your reference resolution - that's the same input-size gotcha the whole editor category shares. And if the conditioning simply doesn't take, confirm the images input actually has a reference plugged in; with no image, there's nothing for the encoder to edit against.