Krea2 Token Attention Weight

Prompt weighting on Krea 2 is a known pain point: the model uses a Qwen3-VL text encoder plus visual conditioning, and the usual (word:1.2) trick works inconsistently because it only nudges the text tokens. UC_Krea2TokenAttentionWeight goes further - it monkeypatches Krea 2's attention layers so you can weight individual phrases as attention odds, and it does it while re-encoding the full multimodal conditioning (text plus reference images). (arms:1.5) genuinely raises the attention those tokens command across the model's blocks, not just the emphasis on the token embedding.

If you've been fighting Krea 2 to obey a specific detail - "arms crossed", "red jacket", "the painting in the background" - this is the node for that fight.

How it works

You connect your Krea 2 model and the clip (the Qwen3-VL text encoder), write the prompt, and give attention_weights a space-separated list like (arms:1.5) (painting:0) (photo:2). The node patches the attention forward pass so those tokens' attention is scaled by the given odds - a value of 0 suppresses a phrase hard, 2 roughly doubles its pull. strength is a global multiplier on the whole effect, and per the tooltip it compounds over all blocks, so a little goes a long way - start at 1.0 and dial down.

The full input list is long because this is the heavyweight multimodal encoder: vlm_resolution (semantic image path, 384 default), vae_resolution (structural path via the VAE), ref_latent_mode (whether reference latents are appended to the main stream or run in a separate "parallel" stream to avoid overriding semantics), system_prompt, formula, padding_method, multiplier (a final overall scale on the conditioning vector), and the autogrowing image_inputs for your reference images. The optional visual_fusion_config input accepts a config from this pack's UC_VisualFusionConfig node to blend visual blocks spatially; when that's disconnected, padding_method and formula become the fallback controls.

Outputs: a patched MODEL (route this into your sampler) and a CONDITIONING (route into positive/negative conditioning).

What to know before you click

This is a genuine power tool with a real learning curve. The attention_weights syntax is (phrase:odds) pairs, space-separated, non-negative - and if a phrase in the weights doesn't appear in the prompt, it silently does nothing. The author also notes numbered multi-image inline placement is intentionally not supported here (use the fusion config path instead), and fusion only accepts image_input_fusion or image_input_1 for its single visual slot.

It's the non-TokenFusion variant of the node. This one does the attention weighting during a single conditioning encode; the ...TokenFusion sibling applies the weights after a token-first visual fusion pass. For most single-image, single-prompt work this is the one to reach for.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart (or ComfyUI Manager → search "ComfyUI-UtilsCollection"). Needs the full Krea 2 stack - checkpoint, Qwen3-VL text encoder (~8GB), Qwen-Image VAE - plus the pack's own opencv-python / typing-extensions. On a 16GB card you'll want to keep vlm_resolution at the Fast setting and watch VRAM; this node loads and patches real attention weights, it's not a string utility.