@JoeLink not all ppl have free unlimited traffic

@JoeLink So what the reason to upload full 5Gb version if 95% cases needs 2Gb pruned version?

@Unicom because it is more accurate and better, if your video card want 2 gig you can prune it afterwards, the opposite if only 2 gig available in public, you cannot do the reverse to return accuracy, worse even full precision mode on such pruned sh*t gives NaN issues on my video card. Mind you that I also pay ludicrous download expense + surprise download failed in the middle so alternatively you may just wait for auto bot to collect it elsewhere, there is a bot project to archive everything including deleted content, but expect such secret mirror to be sued for taken down just like IA

@Unicom now back to your question why not 7 gig? such as 5 full + 2 pruned?

answer, because traffic is expensive, better upload 5 gig rather than 7 gig

@Memeater The people constantly berating full unpruned models do not fully appreciate their broader usefulness. In other threads (not this original poster) others will claim its all just extra junk data, etc...and numerous other users will thumbs up. Those of us training or mixing these models regularly know that isn't true. Granted, there can be junk data if there is data merged from broken or incompatible sources - which does happen, but an unpruned model with all its original functioning data isn't just better for the highest quality output, it's far more useful and I find provides more reliably stable results when merging - if the model itself is stable overall to begin with. I think the people providing pruned as a default are providing less of a complete product, than unpruned. Then again, if one preferred their model only be thought of as a finished product and not meant for merging, they might prefer releasing the pruned. I could understand it in that case. Personally I would always rather have unpruned.

@callie What about those of us who aren't merging models? Is there a significant benefit to 3x size (3x fewer models we can store)?

@callie there are cases (not this model!) like MEGASH*T models by asmrgaming/ollobrains in which the diagnostic and analysis tools gave reports + manual testing gave conclusion of junk data. This is objectively (not matter of hatred or being jihadist as that poster accused me of, way months ago) proven due to testing, say unpruned with Kandinsky gave exact same output with pruned after Kandinsky mix, implying the change appeared of model unmixed vs post mixed is due to the damage/corruption in the process (they said 2x or 4x mixing process, whatever).

however, there are other uploaders too, who mix/merge such as the pruned = unpruned, but sigh, at least without the anti-criticism mentality, locking/erasing comment and playing tricks with revision upload to erase bad reviews, unlike that unspeakable-who-shall-not-be-named poster mentioned before.

forgive me for my rant, but yes, I agree, if there is an objective method to determine junk or if the extra weights is truly never invoked, then of course I also happy with pruned ones.

I hope in next few years lower end consumer cards will have baseline of 16-20 GB due to the overly hyped necessity for AI from voice, video, language models, and whatnot. If the server oriented ones can reach hundreds GB of VRAM soon, it seems getting such lower baseline not too farfetched. Not to mention SDXL and beyond which merges could cough a whopping 30-50 GB soon, requiring collabs.

@Memeater Yes that is what I mean by incompatible source material, without pointing at anyone. I think there is a value to someone trying whatever, more power to them, but broken or incompatible data that is never going to be accessed isn't worth retaining and with the current state of tools would, as far as I know, require pruning the entire model in order to address it in any meaningful way. Ironically a severely broken unpruned model might even function better when pruned. I'm only referring to stable models with content that has been repeatedly tested and plays nice in the same blend, and agree that unfortunately there are, as you point out, stolen, reuploaded, renamed, etc etc approaches to "gaming" the site. It should never be about any of that and all that stuff does is harm the work of everyone else. There are unpruned models that have stable (or mostly stable data) and those are the ones I would always prefer be unpruned. After a doing this a while one can almost tell the models that are going to work well from their sample pics, provided they aren't manipulated with LORAs, img2img, etc. Not just the quality of an image - the stability of the model. My comment wasn't to suggest no one should post pruned or they aren't relevant, but I think most professional mixers want as much (usable) data as possible when merging. I agree about video cards too - and we may likely see a time when a subset of cards are marketed for this purpose specifically. The 3060 12GB is a good example of a early card that already exists that, for what it is, is well designed (and at an attractive price point) for this kind of tech for people struggling with low ram issues and on limited resources.