Krea2 Visual Reference Debug

ComfyUI conditioning is a black box. You wire a CONDITIONING tensor into the sampler, the image comes out wrong, and you have no idea whether the text got mangled, the feature width is off, or the reference image silently took over. Krea2VisualReferenceDebug is the flashlight for that box, purpose-built for the Krea 2 visual-reference pack.

It takes exactly one input - a CONDITIONING - and spits out a text report. That's the whole job, and it's quietly the most useful node in the kevinzilin/ComfyUI-Krea2-VisualRef pack, because this pack is experimental and things will go wrong while you're learning it.

What it reports

Wire it after Krea2VisualReferenceEncode's output and it prints, in plain text:

• how many conditioning items there are (the encode node can produce one or two - a text-only item and a separate visual-token item)
• tensor shape, sequence length, and feature width
• whether that width matches Krea 2's expected 30720 (that's 12 selected Qwen3-VL layers × 2560 hidden dim, and if it doesn't match you loaded the wrong CLIP)
• the attention-mask shape
• all the metadata Krea2VisualReferenceEncode stamps onto the conditioning: style_text_mode, the final unified style_text, the raw per-reference style_text_raw plus its items, the computed palette_hint and per-reference items, style_text_fusion, style_image_count, strength, and effective_strength

The output is a STRING you can view in the UI - or feed into a Save Text node if you want a record. It's not an output node in the save-image sense; it's there to read.

How to actually use it

The README's validation workflow is the right mental model. Keep your prompt, seed, sampler, steps and checkpoint fixed, change only the reference, and read the debug report looking for two things:

1. Does style_text_raw describe style, or does it describe the subject? If the extracted text names the reference's objects, people, text, logos, pose or layout, your style lock is poisoned - even the text-only mode will leak content. This is the single most common failure and the debug node is the only fast way to see it.
2. Does palette_hint match the reference image's actual palette? If Qwen3-VL misread the image, the hint reveals it immediately.

You can also use it to sanity-check the mechanics: is the feature width 30720, is the image-token span present in visual modes, did batching split your images the way you expected (style_image_count)?

One honest limitation

If the conditioning didn't come from this plugin, the debug node still reports shape information - conditioning items, tensor shape, sequence length, feature width - but it can't reliably infer the image-token span, and it will tell you so. That's fine: the shape checks work for any Krea 2 conditioning, which makes this a decent generic Krea 2 CLIP sanity check even before you use the encode node.

The pack is brand new (0.1.0, a single commit) and there's no community mileage behind it yet, so when something looks wrong, this node plus the README's VALIDATION.md checklist is the whole troubleshooting apparatus you've got. Use it. It beats staring at output and guessing.