CRM Preprocessor For Poser

Before CRM can imagine the other five sides of your object, it needs to see the object in exactly the format it was trained on. Feed it a raw photo with a busy background and you'll get six views full of hallucinated scenery. CRM Preprocessor For Poser exists to prevent that: it takes your image plus a background mask and normalizes it into the gray-background, square, centered reference that the pose and CCM samplers expect. It's the node that decides whether your mesh looks like the object or like the object's worst fan-fiction.

What it does, step by step

The source is refreshingly readable. It takes your reference_image and reference_mask, combines them into an RGBA image (alpha channel = your mask, so the background goes transparent), then:

1. Scales the subject by foreground_ratio - the content shrinks but the canvas stays the same size, with transparent padding around it.
2. Expands to square - pads to 1:1 on the shorter axis.
3. Composites onto a flat gray background (#7F7F7F) and converts to RGB.

That middle-gray backdrop is the key detail. The CRM / ImageDream models were trained on orthographic renders against gray backgrounds, so the gray is load-bearing: it tells the diffusion "this is the void, object lives in the middle." If you skip the preprocessor entirely the whole pipeline quietly degrades.

Inputs that matter

• reference_image - your IMAGE, straight from whatever did the background removal.
• reference_mask - a MASK marking the subject. In the pack's demo workflows this comes from ComfyUI_essentials' RemBG node (u2net by default) - so yes, background removal is assumed to have already happened. The mask quality is your mesh quality: fine hair and translucent edges survive only as well as the mask does.
• foreground_ratio - FLOAT, default 1, range 0.5–1.0. How much of the frame the subject occupies. 1.0 keeps it full-size; dialing down to ~0.8 shrinks it into the frame, which can help when the subject's edges are clipped or you want margin around it. Step 0.1, so it's a coarse knob on purpose.

Output: a single processed IMAGE that feeds CRM PoserConfig's processed_image input, and that's it - one wire, and the pipeline picks it up from there.

Practical notes

• Square up the image first. The node pads to square but it's cheaper to feed something near-square. If you feed a wildly portrait image, you'll get huge gray side-bars and a tiny subject.
• Gray is intentional. It looks dead, it's supposed to. Do not "fix" it by feeding the RGB image back - that's the format the models want.
• The mask is the real input. Skip the mask (or feed a blank one) and the whole subject turns transparent-ish and you get garbage. If your background removal is weak, fix that stage, not this node.

Installation

Standard pack install: ComfyUI Manager (search "ComfyUI-Flowty-CRM") or clone into custom_nodes and pip install -r requirements.txt. No model downloads for this node itself - it's pure image math - but the demo workflow wants ComfyUI_essentials for the RemBG step, and the diffusion checkpoints still need to be in models/checkpoints. Run the whole graph on one device, and use the low-vram/ split workflows if you're under 16GB VRAM.