📏 Qwen Image Scale

Every Qwen-VL editing workflow has a quiet mismatch buried in it: the vision encoder and the latent space want different image sizes. Qwen's encoder is happiest around 147K pixels at a divisible-by-32 dimension; your sampler wants something closer to 1.7MP so the edited result has real resolution. If you hand both paths the same image, one of them is working with the wrong size, and you get either mush or wasted context. ArchAi3D_Qwen_Image_Scale is the pack's original answer to that - the V1 of the scaler this repo revolves around.

How it works

It's a smart resize with a split personality. One input (image), and it computes two outputs with matching aspect ratios but different pixel budgets:

• image_vl - sized for the Qwen-VL encoder (vl_target_area ≈ 147K, divisible by vl_divisible_by = 32).
• image_latent - sized for the sampling side (latent_target_area ≈ 1.7MP, also divisible by 32 for VAE compatibility).

The key design choice: by default vl_use_latent_source is True, meaning the VL image is derived from the already-processed latent. That guarantees the encoder literally sees the pixels the sampler is about to denoise - pixel-perfect alignment instead of two independently-resized copies that quietly disagree.

aspect_ratio_mode = auto snaps to the closest of 24 preferred ratios tuned for Qwen (1:1, 2:3, 3:4, 16:9, 21:9, 9:21…); manual lets you force one. latent_area_tolerance (default 0.3) gives the algorithm room to pick dimensions that hit the aspect ratio within ±30% of the target pixel count - that tolerance is what makes exact-ratio snapping possible.

Inputs worth setting

• aspect_ratio_mode / preferred_aspect_ratio - the ratio that both outputs will match.
• vl_use_latent_source - keep True unless you have a specific reason; this is the pixel-perfect guarantee.
• vl_letterbox / latent_letterbox - add black bars to preserve the whole image instead of cropping. The vl_ignore_latent_* toggles exist because the latent is already letterboxed/cropped, so double-processing would be a bug.
• upscale_method options (area, bicubic, lanczos…) - only matter if you're upscaling rather than downscaling; lanczos for the latent is the usual choice.
• mask - optional; resized alongside image_latent with bilinear so a mask stays aligned if you're doing mask-guided edits.
• debug - detailed console log of ratio selection and tolerance checks.

Outputs

image_vl, image_latent, mask, latent_width, latent_height, and debug_text. The width/height outputs feed an empty-latent or conditioning-size node so the sampler knows exactly what to build.

Install and the one thing to know

Same pack, same route:

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 → search "ArchAi3d Qwen" → restart. If you're starting fresh on a region-edit workflow, go straight to V2 instead - it does everything this node does and adds mask-based cropping plus stitch_data for pasting edits back. V1 is the simpler, no-region version, and it's the one you want if your whole image is the edit target. Both are free personal / paid commercial under the pack's license.