Awesome work and thanks, but why are the Bobs significantly smaller?
I think I speak for everyone when I say we greatly appreciate you doing the conversion.
Had a question though: Why are the Bob variants so much smaller than the FP8s? Int8 ConvRot shouldn't be over 5GB smaller... if anything, they should be slightly larger. It must be stripping something out, and I think that's probably why you're seeing it be the fastest. I'm just not sure what we're losing. Have any idea?
Hi great question. I looked quickly a few days ago at Bob's recipe for exclusions and it is different than the flag in the CTQ. Its much more surgical and it doesn't exclude the entire audio_vae and vocoder blocks. I haven't looked at what else he does. I haven't had time. Decisions like row or tensor might be in play as well for example. I actually think he does tensor not row which is the same as the CTQ settings I used, that was just an example.
While the CTQ variants should be closer to BF16 because they exclude more I couldn't tell you how much by and I think it ventures into territory where its a bit subjective. Gets into a weird area where people are comparing the output of generations and I've seen those discussions turn pretty wild. The limited amount of testing I did, I found a seed that looked ok in bf16 in a basic 2 stage workflow, first frame input only, 20 seconds and ltx generated sound. Outputs all looked pretty even to me, sounded ok, as good as ltx sounds go, but others may have different opinions of course. All up I ended up making and testing around 8 variations.
In short those made from Bob's node are to me great quality and the fastest out of the four posted. I have found another variation that is faster but I want make more comparisons before I put it up.