RP Regional Detailer (Qwen)

What it is

The Qwen variant of the regional detailer, and it's honest about being one: RP Regional Detailer (Qwen) delegates entirely to the Z-Image detailer. Qwen-Image and Z-Image share enough of the latent/VAE family that the same 16-channel crop-and-inpaint pipeline works for both. The one structural difference is a quiet piece of engineering - Qwen's VAE can return a 5D tensor shaped [B, T, H, W, C], which the node normalizes to 4D [B, H, W, C] before it hits the sampler. If you've ever seen a "5D tensor" error from other Qwen nodes, this one already handles it.

What you'd set

Because it inherits the Z-Image pipeline, the model-specific advice shifts a little:

• steps: 15~20, not Z-Image's 6~10. Qwen isn't Turbo-distilled the way Z-Image is.
• cfg: 1.0 - recommended for Qwen's distilled variants. The usual distilled-model rule applies: CFG fights the model.
• denoise: 0.35–0.55, same as the other detailers.
• shift: the AuraFlow sigma shift, for Qwen.
• Everything else - bbox_model, detect_threshold, scale_to_pixel, use_base/use_common - mirrors the Z-Image node, including the YOLO file in models/ultralytics/bbox/ and the same optional divide_mode/divide_ratio inputs.

The wiring

Standard RP chain: RPPromptParser → RPRatioParser → a plain KSampler (fine for Qwen) → this node with model/clip/vae plus a negative CONDITIONING. Outputs are image and debug_image. Region prompts are encoded as COMMON + BASE + DIV, same as every detailer in the pack - no scene-narrative merging.

The honest review

If you only run Qwen, this is the node you reach for, and the delegation means it stays maintained in lockstep with the Z-Image detailer. It's also a tidy illustration of how the pack thinks about architecture: one YOLO assignment and one inpaint pipeline, re-skinned per model family with the right latent shape, step count, and CFG. Don't expect Qwen-specific tuning beyond that - the value is that the whole regional stack now has a working Qwen path at all. Dependencies are the same trio as Z-Image: ultralytics, opencv-python, onnxruntime.