Bonus: Flux "Negative Prompt" at CFG 1.0
Added 2024-08-09 09:33:08 +0000 UTCMany of us have become used to using a negative prompt when generating images using Stable Diffusion, so going back to CFG 1.0 and not using a negative prompt can feel a bit like losing a favourite tool of some sort
There is a newer version out which I haven't actually tried as yet, but hopefully it (or others) can do much the same thing!
Anyway, the thing I've been testing is with using Automatic CFG as it has interesting controls for sigmas. I'd found ending somewhere around 0.8->0.9 typically did the trick, but then I also noticed the "For magic" preset could also be used and wasn't such a large node as the advanced one :)
You might have seen people using a CFG > 1, so the advantage with using magic here is that we don't get such a massive speed loss.
But does it work? Well... sort of ? I've tested a number of simple words like as hair colours and such like, and there is certainly some change. Single words along the lines of "red" or green" appears to drastically reduce those colours.
Examples
This negative is "blonde hair"
This negative is "brown hair":
This negative is "brown hair", but with the "magic" node disabled - i.e. no negative:
Performance-wise, when using the "negative", it's only slightly slower, with around 1s difference:
Node
More info on the Automatic CFG node is here:
"Older one": https://github.com/Extraltodeus/ComfyUI-AutomaticCFG
"Newer one": https://github.com/Extraltodeus/Skimmed_CFG
Workflow
JSON file attached to this post as usual :)
Comments
Loras need to be the same architecture, so for flux you'll need to use flux loras - https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md
Nerdy Rodent
2024-08-11 11:20:38 +0000 UTCWhaaaaaa? Loras? Like the sdxl loras work? New ones...I did not hear about that. Tried my old ones but they seemed to have no effect. Thank you for the reply, by the way. Check out the Flux Animated stuff from Oliver Sarikas. This might be interesting for you also. Cause it is very impressive.
Asyouare
2024-08-11 10:08:30 +0000 UTCWe've got loras and controlnet already, with ipadapters being trained :)
Nerdy Rodent
2024-08-11 10:02:09 +0000 UTCLooks interesting :)
Nerdy Rodent
2024-08-11 10:01:27 +0000 UTCDear Rodent(?) I'll call you MrR, How long do you think until the flux model in integrated as any other model and do you have any idea why it does not respond to many of the features available for larger models now? I mean it started in the "unet" folder which is already something we did not have much of before...have you heared of anyone working activly on furthering the model? The civtai guys must be on this like politicians on the "Reimbursement of expenses" forms.
Asyouare
2024-08-11 09:38:20 +0000 UTCAnyone else’s using ComfyUI-N-Sidebar? Pretty cool guys, maybe a good video idea
Lord Banzai
2024-08-11 02:56:45 +0000 UTChi, just got this error "Error occurred when executing CLIPTextEncodeFlux: CUDA error: out of memory CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions." any idea what could cause it? [SOVED] just updated comfyui, now flux is integrated
Christian G.
2024-08-09 19:37:05 +0000 UTCIn my limited testing there isn't a difference in speed choosing FP8 vs. FP16. Or if any it's negligible.
Richard Gates
2024-08-09 14:52:15 +0000 UTCVery little difference - https://x.com/NerdyRodent/status/1819406126389383367 - fp16 does handle text and hands/fingers better though.
Nerdy Rodent
2024-08-09 13:50:38 +0000 UTCAs always, great work! Have you done any testing using the same seed between FP16 and FP8 to see how much of a loss of quality there is? I'm able to do 3 images a minute with this workflow at 832x1216.
Richard Gates
2024-08-09 12:38:41 +0000 UTC
