Hyper Remap Krea2 Edit Slim (CCN)

This is the Slim treatment applied to the image-grounded remapper: the full four-phase remapping pipeline running on a krea2_edit encode, with the text input stripped down to a wire and a single conditioning output. If you read the Hyper Remap Krea2 Edit article and thought "that's exactly what I need, but I don't want five extra sockets," this is the node for you.

The pipeline is identical in spirit: comma , for string replacement, -> for token embedding blend, => for concept direction nudges, ~~ for additive residual deltas, with per-pair (b:X, s:X, t:X, sx:X, tx:X) overrides beating the global blend / sharpness / threshold defaults. What changes when the encode is image-grounded is how the embedding phases behave, and it's worth knowing before you rely on them:

• Concept remap runs in cosine mode - since the grounded sequence carries the vision tokens and chat template, a text-only re-encode can't match its length, so differential weighting is skipped and the nudge is placed by cosine similarity instead.
• Delta remap runs in pooled mode - aux prompts encode text-only (shorter), so each delta is collapsed to a direction vector and broadcast across the grounded sequence.

None of that needs your input; it's the node adapting the math to the vision-bearer's longer sequence. It just means results skew toward "broad directional push" rather than surgical placement.

Inputs, reduced

Wire in a clip, a text string, and the image grounding:

• image - source image for the grounded encode. Disconnected, the node still runs the pipeline on a text-only encode.
• image_b - optional second reference for multi-ref LoRAs (scene first, subject second in training order).
• grounding_px (default 768) - longest-side cap for the vision model; 0 = native resolution.

That's it. No case_sensitive, no debug, no untouched output - normalize_delta is the only extra toggle, default on, and you should leave it there.

Where it fits

This is the version you wire into a workflow that already manages its text upstream - a prompt store, a string builder, a chained neutral-prompt setup - and where the graph needs to stay clean. The cost of the leanness is the same as the other Slim: no A/B reference output to check "did the remap actually move anything," and no debug line when a nudge silently does nothing. If you're actively iterating on remappings, the full Krea2 Edit node's original_prompt/modified_prompt outputs are worth the clutter. Once your remappings are tuned, swap in Slim and forget it.

Setup reality check

The node ships with the pack -

cd ComfyUI/custom_nodes
git clone https://github.com/valkymaera/ComfyCollectorNodes

restart, or ComfyUI Manager → ComfyCollectorNodes - but it's not a self-contained tool. The krea2_edit grounded encode needs the Qwen3-VL text encoder plus the krea2_edit LoRA and its chat template, the same stack ComfyUI-Krea2Edit uses. If that's not already in your workflow, this node has nothing to ground on. And expect the vision-encoder tax: re-encoding through a VLM per remap pair adds seconds to a run. It's a specialist tool for people already living in the Krea2Edit ecosystem - which is exactly who it's for.