Anyone else having issues with model merges creating unusable files?

Oh man, model merging can feel like alchemy sometimes! One thing that's super easy to miss is the model format compatibility. That `"model_a": "model_a.safetensors"` line in your command - are you absolutely sure both your base models are the *exact* same architecture? Like, merging a SD 1.5 model with a SDXL model, or even two different flavors of 1.5, will absolutely give you a corrupted or silent file. The webui won't always stop you.

Also, that `"ratio": 0.5` is a perfect 50/50 split, which in my experience is a real coin toss for usable results. Try starting with something much more weighted, like 0.8 or 0.2, to gently blend a bit of one model's "seasoning" into another. I'd recommend doing tiny, iterative merges and testing each one rather than going for a big 50/50 blend right away. The crash on load definitely sounds like an architecture mismatch, though. What models are you trying to combine?

Welcome to the sometimes frustrating world of merging! user700 gave you great advice about model compatibility and ratios.

Before you even get to merging, there's one more crucial check: always verify the hashes and file sizes of your source models in the webui's model list. A download that got corrupted or interrupted will absolutely cause silent failures in merges. It's a common, easy-to-miss culprit.

I'd also suggest checking the community guidelines when you're ready to share your results - we've got some specific rules about posting merged models to respect original creators. Good luck, and I hope your next attempt works!

That ratio's your first real trap! Starting with a straight 0.5 is like tossing two paint cans together and hoping for a nice color. It rarely works.

Beyond what others said, you're probably just seeing the webui's basic merge. For any real control, you need to look at the *merge method* dropdown. "Weighted sum" is the default, but try "Add difference" - it's way more predictable for blending styles. Also, double-check that "Interrogate CLIP" option isn't ticked unless you know why you need it; it can silently fail and leave you with a dead file.

What tutorial are you following? Some older ones skip the critical step of pruning the merged model, which leaves a bloated file that can crash on load. Give `scripts/merge_models.py` a look in a terminal - the extra flags give you way more feedback than the UI.

Oh wow, the "pruning" step is something I've never heard mentioned in any of the beginner guides I've watched. That could definitely be my crash issue.

So when you say to check the merge method dropdown and try "Add difference," is that generally more stable for a first-timer than "Weighted sum"? And is pruning something the script does automatically, or is that a separate step I need to learn?

Thanks for pointing me to the script, too - getting more feedback than the silent UI would be a lifesaver. I'll go look for it.

Yeah, the architecture mismatch you mentioned is a big one. I ran into that early on trying to merge a model trained on specific anime styles with a general purpose one, both SD 1.5. Even that subtle difference gave me completely silent outputs. How do you reliably check compatibility beyond just the base model version listed on Civitai? Are there certain tags or metadata fields in the safetensors file that flag a "flavor" mismatch?

You're probably running into a silent architecture mismatch. That JSON block doesn't show the actual model names, so you could be trying to merge a 1.5 base with a 1.5-pruned or a finetune with a different CLIP skip baked in.

Don't just trust the filename. Load each model separately and generate a basic test image first. If one works and the other fails, you've found your bad source. If they both work individually, then move on to checking the merge method as others said. The default "Weighted sum" at 0.5 is the most likely to produce garbage.