ULtraReal 2k
1280x1600 native Latent Size (1536x1536 Square)
2400px up-scaling image to image workflow
EVERY example image is first seed (No seed hunting)
This model uses FP32 CLIP these commands should be used, this will not slow down your IT's unless you have very low system RAM
Comfy UI --fp32-text-enc
Forge/Auto1111 --clip-in-fp32
VAE in FP32 is recommended
Description
FAQ
Comments (21)
Ok, how brutally should I put this to the test with my lora stack or would it be better suited to test low/no lora first?
I had 1 lora that caused a bad failure out of 50 with my own trainings, but I have far far less lora's then you
@Felldude I'll attempt to keep the stack small for now, its standard pony with the Clip L&G otherwise right ?
@A_Friendly_Spider It is all 32bit - with the base PONY CLIP not SIM CLIP-G, but I nearly always downcast the unet to BF16
@A_Friendly_Spider I did do a write up showing how much detail FP32 UNET adds on ULtraReal Supermodel ED, https://civitai.com/articles/10935/comparing-full-fp32-and-bf16fp32-pony-ultrareal-2
@Felldude Preliminary testing thus far is, strange, results to follow.
@A_Friendly_Spider I see the gradient pattern from some of the LORA's and not others, Im not sure if this is a prodigy or ADAfactor issue....the lora's that seem to work with everything still work with everything though lol
Still getting the random overblown exposure that I was getting in the standard version, but still can't figure out how to avoid it. Some prompt/LoRA combos are 100% extreme contrast, while some give alternating good contrast then bad contrast from the same run.
Little difficult to work with but can produce some amazing results when it wants to!
Thanks. I have had a few images with poor exposure, I do use FP32 VAE and CLIP. But that may not be the reason.
@Felldude There's not a VAE already baked-in for your UR 2k and non-2k models?? When running a model without any VAE I always get 100% washed-out, faded results (completely lacking in both contrast and saturation), but this is always increased contrast. Sometimes nearly black-and white, sometimes nearly perfect, all in the same batch. I'll try loading an SDXL VAE with it and see what it does.
Could you recommend an FP32 VAE I could try? Tried searching here and at HF but didn't come up with anything other than standard SDXL/Pony VAEs.
@AFD_0 The FP32 VAE is baked in but if your not using the FP32 VAE command it would be downcast
@Felldude Downcast as in functioning as an FP16 VAE or just not loaded at all?
I'm using Easy Diffusion and haven't noticed any VAE-specific settings other than the section to load a custom VAE. It's a fairly basic and outdated UI, so I'm actually surprised your FP32 models work with it at all to be honest. Since I don't have a separate FP32 VAE to load, I'll try an FP16 VAE and see if anything changes.
@AFD_0 Cast to FP16, other then forcing BF16 UNET I can't think of anything else that would be different - All the workflows are for comfy though
Does LCM exponential or uni_pc sampler help?
@Tootle My software is old and doesn't support either afaik, just DDP++ SDE, 2M SDE, Euler A and a few others. I've had this same contrast issue on most of the recent models by @Felldude - except for his 3xThreat SDXL/PONY/ILLU model, which doesn't exhibit this problem at all.
@AFD_0 Easydiffusion was helluva good. WAS! for 1.5
If you want to use xl or pony and normal speed, use Forge, Fooocus. Fooocus is the easiest.
@2u0v645oz5683 Seems fine for SDXL/Pony/Illu. Not the most comprehensive selection of samplers and options and lacks support for any advanced workflows, but it's still usable for 99% of those models (a newer Flux version is in beta). But yeah, for an old, basic and somewhat limited tool, I'm still impressed at what it's able to do. And oddly enough, Felldude's 3xThreat FP32/2k model doesn't exhibit the same random contrast issues I get with his ULtraReal FP32/2k model.
@AFD_0 Sadly you limit yourself with Easy. I tried very long ago XL models on it and i always got out of memory, and the speed was slow with 8gb Vram. I dont know, what card do you have, but with Forge you will get big speed boost. If you have 16gb card, than its ok.
@2u0v645oz5683 Using an old 2080 Ti with 11GB VRAM for now. I just leave it on the "low VRAM" setting and it hasn't given me any issues, even with >12GB FP32/2k checkpoints ("balanced" usually works fine too for 6GB models). Was planning on trying ComfyUI or Auto1111 whenever I get around to building a new PC, but not really worried about changing software until then. I'll take a look at Forge and Fooocus as well. Appreciate the suggestions!
I was able to see an example onsite of an overblown image, I was able to resolve it by lowering the CFG and steps. I would not use high cfg on any sampler but euler,
Pruned fp16 6.5GB please.