Goddess Raw (Schnell)

@Felldude Just a small update and I'll refrain from pumping another 40 images to the model (unless you really want them)
But I did another 20 in which I skipped the "LLM prompting" So the "They hold a sword in one hand and a shield in the other hand." was always at the top of the prompt, then all the random prompt generation and stylizing (which I fully randomized this time).
A total of 6 of 20 were holding the sword incorrectly or not in a natural way. 1 was missing the sword altogether.
With the LLM taking a pass at the prompt, it was just 2 out of 20, but 2 were missing the sword, one of which was totally nothing like anything it was more a pattern etc.

I thought, why would the LLM re-wording the data have an impact. So I tested something.
I wrote a one sentence as it would appear in a book. "As they held their sword in one hand, they swung down, they held up their shield with the other hand to protect against the coming attack." Removed all the prompt generation and styling and just pushed it through 20 seeds. It was similar to the LLM taking a pass. Three "weird" looking sword holding.

I wonder if the schnell base needs more 'context' to remove randomness. It very clearly has a higher 'weirdness' factor when you just throw a few elements you want at it and it has to work out the rest. But when there is additional "fluff" that the LLM adds it, for lack of a better word, understands the intent, rather than just making sure the 'objects' are in the image.

Sorry for the long post, I'm going to test raw schnell probably tomorrow to see if that's the case.

@Felldude My little Schnelly tried to give me an unprompted lewd the other day, I was so proud 🥹

There is little difference in VRAM usage between Q6 and Q8. If you can afford to run a Q6, you can run the Q8 just fine. From all the data tables I've seen the difference in data loss between Q6 and Q8 is "little" but it's "twice" as much loss as the Q8 from FP16. You should be good with Q8. It's under 12G, I'm guessing a Q6 would be like under 1G smaller. Still too large for our 8G brothers and sisters, but Q8 should be just fine for 12G VRAM and more than perfect for the 16G VRAM users :)

Thank you Felldude for the Q8 :)

Everyone is different for this. For me NF4 does good images, but they are so much more variance over F16 that bothered me. Q8 does way less variance over the F16 model. I guess I kind of like that more. It's not that NF4 is bad, it's probably my OCD LOL

@SencneS I know I can use Q8 or FP8 with 16GB VRAM, but no, it's no perfect, unless I want to wait a lot, away from your computer.
Q6 is the sweet spot if with 16GB you want to use loras, change T5... or just use your computer while it's working on its picture. ComfyUI is more efficient and predictable with VRAM, but I prefer to use Forge.

@eFeRBe I guess I've never noticed any real difference in generation time between the versions and I've tested a lot. You're right I am using Comfyui and that may be the difference. Honestly I just tested the difference between a Q6 and Q8 for me was 3.82s (faster for Q6) This was 1 sample. It might be Comfy vs Forge thing. Generation for me from start to finish are between 40 and 90 seconds depending on what I'm doing. The 90 second mark is me passing the prompt through a separate text generation LLM. So I can get up there in generation times if I really have the ultra complex workflow, but not enough to consider "excessive generation". But if you're use to like 10 second generations I guess 40seconds seems like a lot.

Edit :- Clarifying note I'm not generating with Schnell, pretty much always Dev with 25+ steps.
Because schnell is under 10 seconds for me pretty consistently.

@Felldude I just tested your FP8 with my 4060Ti. I get 2.3s/it, it seems normal (and reassuring). But sometimes I get a warning saying I have no VRAM left. I thought that the option --clip-in-fp32 will load the TE in RAM to be used by the CPU, but it seems that at one moment, I must have the size of model+clip free in VRAM. Am I wrong or did I misunderstood that it's possible to load the text encoder only in RAM ? (and thank you for your previous answers)

@Felldude I think it's me and not Forge :) FP8 is 14.5GB and it's easy to use the 1.5GB left, especially if you have several web browsers opened and a 4K display, and changing the T5 size doesn't change anything.
And I ran out of RAM (without V) too with 32GB with another Flux model. When it appends it's hard to diagnose precisely because the mouse pointer moves 1cm / minute...
I ordered 32 more (I knew it will append one day).
And thx for the pix btw.