I was recently trying some stuff with Civitai generator and had hard times to replicate from Comfy in Civitai. Here are my notes on it, it should be useful the other way around too.
Testing was done primarily with SD1.5 but results should be universal.
Yell if you have something to add or complain.
TLDR
Didn't find a way to pixel perfect replication. Subtle difference remains.
Enable Mature Content filter, unless you have a reason not to.
Token weights are different from comfy.
Hi-res fix is probably only img2img at 1.5 scale.
Still no luck? Run a batch of four.
Still no luck? Forget it.
Mature Content Filter
I think this filter does filter prompts or adds an embedding to prohibit mature content, probably even both. I realized this when having "sexy" in prompt and women looked really tame, enabling the filter solved the problem. Simply testing with "a tree" the impact of the filter seems fairly strong. (See cover image.)
I always thought the filter is there to give the hint for moderation, but in hindsight that's quite stupid. The filter is there trying to guarantee to never show nasty stuff.
Token Weights
Token weights don't work the same across all user interfaces, as explained here. Not sure how civitai does token weighting but it's not the comfy version.
Schedulers or Secret Sauce
Schedulers aren't listed, e.g. I didn't find the scheduler in comfy to perfectly replicate an image, but euler/normal seems fairly close. Either I am missing something or there may be some other embedding that tries to tackle illegal content. The difference here is small and probably negligible.
Sub-Seeds
Are sub-seeds are a thing? What I mean are the seeds used for batches. They are not incremental in comfy and this is nothing new. But it adds to the complications to replicate works from civitai. So if an image is off, run a batch of four to have the chance to find it.
Lost Workflow
This should be somehow obvious. Civitai doesn't store or show the full workflow. Any settings, loras, prompts etc. may be lost as long the work isn't really straightforward. So if all of the above didn't help to get close to the result, this is the problem that you won't fix easily, if at all.
Hi-res Fix / Face Fix
Not sure what's happening here, but is odd. Upscale and Face fix does what it is supposed to do. But all the hi-res stuff offers only 50% size increase and is not repeatable? I think hi-res fix is just a new sampler round with 50% size increase. Quite a false flag operation in terminology as I would expect sd upscaling.