🎨 Qwen Encoder Simple V2

The "Simple V2" encoder is the pack's answer to the two most common encoder complaints at once. First: the other encoders expect your images pre-sized for the VL encoder, which is one more node and one more failure mode. This one takes images as-is - "no resizing" is literally in the README's description. Second: most people never bother writing a system prompt, so this node ships with one already filled in, telling Qwen to describe the input and then apply your instruction while keeping consistency where appropriate. You can overwrite it, but you don't have to think about it.

The panel

Required:

• clip - the Qwen-VL CLIP.
• prompt - your edit instruction.

Optional:

• system_prompt - has a sensible English default baked in ("Describe the key features of the input image... then explain how the user's text instruction should alter or modify the image..."). Replace it with your own or clear it.
• vae - needed for the ref-image latents.
• vl_image1 / vl_image2 / vl_image3 - up to three images for the vision encoder, no resize.
• ref_image1 / ref_image2 / ref_image3 - up to three images VAE-encoded as reference latents.

That's it. No strength dials, no image labels, no debug toggle. Two required inputs and everything else optional.

Outputs

Two, and this is the one real gotcha: conditioning (ready for the sampler) and formatted_prompt (the final ChatML prompt for debugging). Note there's no latent output - the other encoders surface the image1 latent for VAEDecode; Simple V2 doesn't. If your workflow expects a latent wire coming out of the encoder, use a different encoder or grab the latent elsewhere.

How it works

Same dual-encode pipeline as the rest of the family - VL images for semantic understanding, VAE-encoded ref images for appearance - wrapped in ChatML. The default system prompt does the "describe, then modify, while keeping consistency" framing so that a bare instruction like "turn the room into a modern kitchen" still produces a proper edit rather than a hallucinated replacement. The trade-off for skipping resize is that the node relies on your images being already at a sensible resolution - it won't warn you if you feed it something huge or weirdly shaped.

When to reach for it

The README's workflow list has this as the "multi-image direct input" version - it's the encoder for mood-board work and reference-based design where you're dropping in several images and a short instruction and want it to just run. It's also the best first encoder for a beginner in this pack: the default system prompt covers the mistake beginners make most (not telling the model to preserve the original where appropriate), and the no-resize behavior removes the sizing node from your first workflow.

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen && pip install -r requirements.txt

Or ComfyUI Manager → "ArchAi3d Qwen". Restart, find it under ArchAi3d/Qwen. You need the Qwen-Image-Edit checkpoint and a Qwen-VL CLIP (quantize the 20B model on consumer GPUs). One note: while it skips resizing internally, the model still has a preferred input range - if your source images are wildly off, run them through the pack's Image Scale node once before this encoder rather than feeding it a 4000px monstrosity.