Regex Weights

Mecha Regex Weights is the "I want a different strength for these weights" node. Where Blocks Mecha Hyper gives you one weight per architecture block, this gives you weights per key pattern: write rules like pattern: weight, one per line, and every model key matching a pattern gets that weight when the merge runs. It's the general case of block weighting - instead of being stuck with the architecture's notion of "block," you can weight by anything you can express as a pattern. Merge attention blocks differently from the rest, boost the mid layer, treat all time_embed keys one way - that's the kind of control that used to mean writing a custom script.

Inputs

• regex_weights - the rules, one per line, pattern: weight. The author's tooltip: <pattern>: <weight>. What counts as a pattern depends on regex_mode:
  • simple (default) - the only special character is *, meaning "any number of characters." a*n matches alliteration. \* matches a literal *, \\ a literal backslash. Perfect for people who flinch at regex.
  • regex - full Python regex. .* matches any number of characters. For real pattern power.
• default - the weight for any key matching nothing (a FLOAT, default 0.0). This is your "everything else" value, so zeroing it out zeroes out unmatched keys - usually you want it set to the same strength as the rest of the blend.
• model_config - which key layout the patterns run against (17 configs, default sdxl-sgm). The patterns match keys in this layout.
• match_full_keys - false (default): a pattern matches if it covers part of a key. true: the pattern must cover the whole key start to finish. false is the forgiving mode - you'll notice if a partial match catches more than you meant.
• regex_mode - simple or regex, as above.

Outputs: recipe (the param MECHA_RECIPE) and recipe_txt (the STRING form of it - handy for seeing exactly what the patterns expanded to).

How it works

Patterns are matched against each key in the model config, in order; first match wins, and matched keys carry that weight into the merge's param slot (typically alpha). Because the result is a param recipe, it plugs into the same sockets as a Float - but with per-key variation instead of one flat number.

Install & gotchas

ComfyUI Manager → search mecha → Mecha Merge Node Pack, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ljleb/comfy-mecha.git
pip install -r comfy-mecha/requirements.txt

Restart; only dependency is sd-mecha==1.1.7.

Where people get burned: the default of 0.0 is a trap - if you write patterns for a handful of keys and leave default at zero, every other key gets strength 0 and the merge collapses to almost nothing. Set default to your base strength first, then layer patterns over it. Second: patterns are matched against the keys as they appear in the chosen model_config - a pattern that worked for SDXL won't necessarily match the same semantic layer in a kohya-layout config, where keys are renamed. And in simple mode, remember only * is special; if you write what looks like regex and it silently matches everything, that's the mode doing its job. Use the recipe_txt output to verify what actually matched before running the full merge.