EmAySee Empty Qwen Image Layered Latent noRS

If you're building a Qwen-Image workflow, you need a latent to start from, and Qwen-Image's latent isn't the same shape as SD/SDXL's. It's a 16-channel latent (same channel count as Flux and Wan), it downsamples 8× per side, and - the part that trips people up - it carries a layers dimension, because Qwen-Image generates and edits in layers. This node gives you that correctly-shaped blank latent in one step. It's the "Empty Latent" node for Qwen-Image, minus the reference-sampling machinery.

What "noRS" means

The pack also ships Qwen encode nodes (EmAySee_TextEncodeQwenImageEdit_noRS and a Plus variant) whose names share the noRS suffix. RS here is reference sampling - ComfyUI's built-in Qwen-Image empty-latent nodes support feeding in a reference latent for guided editing. This family drops those extra inputs entirely. So what you get is the simplest possible start: width, height, layers, batch_size in, LATENT out. No hidden reference-latent wiring to confuse a fresh graph.

How it works

The node allocates a zeroed tensor of shape [batch_size, 16, layers + 1, height // 8, width // 8] on ComfyUI's intermediate device. Reading that shape tells you everything about Qwen-Image's latent space:

• 16 channels - four times SDXL's 4-channel latent, which is why Qwen-Image holds fine texture through editing passes that would smear older models (the same channel-count ladder that Flux and Wan sit on).
• ÷8 spatial - a 640×640 input becomes an 80×80 latent per layer.
• layers + 1 - the extra layer is where Qwen-Image's multi-layer editing happens; layers defaults to 3, which is what most editing workflows want.

The noRS suffix is worth knowing because the author's own README doesn't document it - the README is AI-generated, admits being stale, and predates this node.

Inputs and outputs

• width (INT, default 640, min 16, step 16) and height (INT, default 640, min 16, step 16) - must be multiples of 16; the step enforces it.
• layers (INT, default 3) - layer count for the latent stack.
• batch_size (INT, default 1) - how many latents to allocate.

Output: LATENT. Wire it into the pack's Qwen TextEncode/Edit nodes (or ComfyUI's native Qwen-Image conditioning nodes) and then into a KSampler tuned for Qwen-Image.

Install

Part of ComfyUI_EmAySee_CustomNodes (Manager → search "EmAySee", or git clone https://github.com/EmAySee/ComfyUI_EmAySee_CustomNodes into ComfyUI/custom_nodes, then restart). No requirements or model downloads - the Qwen-Image checkpoint itself is separate and you load it with your usual checkpoint loader. Community users run this pack's Qwen nodes on 16GB cards, so it's not a VRAM monster, but the model files are; budget for a Qwen-Image checkpoint download.

The honest take

If you already use ComfyUI's built-in Qwen empty-latent node, this is a lateral move - same shape, fewer inputs. Where it wins is the author's broader Qwen workflow: their noRS encode nodes expect this exact blank-latent shape, and having the whole family from one pack means no version-skew between encode and empty-latent nodes. Grab it when you're following an EmAySee Qwen workflow; otherwise ComfyUI's native node is fine.