Flux Attention Seeker Generator

If you've looked at GR85_FluxAttentionSeeker2 or GR85_FluxAttentionSeeker3 and thought "I am not typing twelve, or worse twenty-four, comma-separated floats by hand" - this node is the answer. It generates those value strings for you, seeded so the results are reproducible, instead of you hand-crafting a plausible-looking distribution.

Quick framing before the mechanics: "attention seeker" isn't a term with community writeups behind it - this looks like the pack author's own experimental idea for varying Flux's text-encoder attention per layer, and this generator is the tool that makes iterating on it practical rather than tedious.

How it works. seed drives a reproducible RNG - same seed, same generated values, every time. distribution_scale (default 0.5, range 0–1) controls how far the generated values spread from a neutral baseline - low values keep the output close to 1.0 across every layer (a gentle, conservative variation), high values allow bigger swings between layers. random_generation_count (default and max 36) caps how many of the total 36 layer slots (12 for CLIP-L plus 24 for T5-XXL) actually get randomized - set it lower and the rest presumably stay at their neutral default, giving you a partial rather than full randomization. The optional combined_values input lets you feed in an existing value string to build on or reuse instead of starting from pure random.

Outputs. clip_l_values is a ready-to-paste string for GR85_FluxAttentionSeeker3's clip_l_values_str (or as a reference if you're using Seeker2's individual sliders - you'd read the values off and set them by hand). t5xxl_values is the matching string for either Seeker node's t5xxl_values_str input, identical format in both. combined_values bundles CLIP-L and T5-XXL together into one string - handy for saving or logging a full "recipe" you liked in one place. average_value gives a single number summarizing the overall distribution, a quick way to compare two generated sets without reading every value.

Installing it. Search "comfyui_gr85" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/veighnsche/comfyui_gr85

Restart ComfyUI. No model downloads or heavy dependencies - this node only produces numbers, it doesn't touch a model itself.

Where it fits, and the one thing to check. This node doesn't do anything to your generation on its own - it has to feed GR85_FluxAttentionSeeker2 or GR85_FluxAttentionSeeker3, which are the nodes that actually apply the values to a CLIP object. If you generate values here and nothing changes downstream, confirm the string outputs are actually wired into one of those two nodes and not left dangling. And because this is exploratory by nature - there's no established "good" distribution_scale to reach for - expect to try a handful of seeds before you find a combination that does something you like, the same way you'd explore any other random-seed-driven parameter.