DR34MY, AIN'T IT?
===========================================================================
D1-Q4_0 GGUF - 5/13 - (dev-1) - Here's a quick Q4 (GGUF) quant for y'all to enjoy while we finish up the other v1 variants! Stay cool. =)
Note: We may not release Full/Fast versions unless we get much interest as our focus (and VRAM) is focused elsewhere for the next few months. :)
D1-fp8 - 5/10 - (dev-1) -The first of the first, the best of the worst. We've got high hopes for HI-DR34Mz, which can be thought of as a sister model to C4PACITOR.
===========================================================================
Note - try our beta custom, fine-tuned CLIP-G (aka Dirty-G) if you intend to use the model primarily for NSFW purposes. Feel free to employ with other models that use Clip-G too.
Dirty-G may introduce the occasional watermark or cooked image, we're working on an improved version but wanted to share it for fun regardless. We're working to balance pose issues and text problems with more balanced resolution training in future epochs.
Full and Fast checkpoints coming soon. NF4's too!
We look forward to seeing what y'all make - please share below!
===========================================================================
Models are created by DR34MSC4PE with all the trappings you've come to expect as well as some cool bonuses:
Enhanced realism and photo-realistic concepts, trained on high quality datasets with the latest techniques.
Specific anime/illustration tuning to introduce some lost artistic concepts
NSFW tuned and capable of realism and artistic/anime images. Female anatomy is well represented with additional fine tuning in the works.
Exceptional performance with character/other/stacking Lora
Like our work? Buy us a coffee: https://ko-fi.com/dr34msc4pe
===========================================================================
DR34MSC4PE is
@c0ur4ge
@ERA5ER
===========================================================================
Recommendations by base model:
dev -
good: uni_pc + normal / ~28-42 steps
best: ClownSampler + Beta52 / ~28-??
Description
Test. Test. Is this thing on?
FAQ
Comments (27)
Love your work on the flux models and looking forward to testing this out along with the clip G.
Thanks for the kind words - let us know what you think!
Having fun. The diversity is great compared to one other fine tune I've seen. Been getting great results from the clip G and the other changes in the text decoding.
PLEASE add Q2 GGUF of all versions !!!!!!!!!!!!
I could probably make this happen - I wasn't super duper happy with the quality honestly but I'll see what I can do!
Thank You. Great Work.
hidream is faster or flux dev ? if say we do 8 steps ?
I believe the recommendation for DEV is 28 steps but you should experiment!
Getting good results at 20 steps but still struggle a bit with shift and cfg. Cfg because some dev variants gave me good results on cfg 3 for the dev model and shift seems to be a matter of preference
@Kaleidia Yeah - I've noticed this myself, particularly as it relates to sampler choice. One thing that helps a fair bit IMO is using ClownsharkBatwing's RES4LYF custom node suite. There's some fantastic stuff in there like the RES2M/Beta57 combo that works well for FLUX and I assume would be decent here too.
Odd, when I search Hidream, this page doesn't come up. Just so you know
Hmm - I wonder if that has to do with defaults or safe search. A good call out though since that's what I have the model typed as!
I'm not sure. Tho I know other things have fallen into the cracks when searching. I just try my best to follow Hidream models and yours isn't the only one showing up for what should be a simple search.
I don't know what I'm doing wrong but this model gives me comparably worse results than base HiDream Dev Q3 GGUF. Colors are washed out, and sometimes anatomical problems also crop up. I have tried both uni_pc/normal and beta57 with res 3 and res 2 at 30 steps. I have also tried various clip and T5/llama combinations, the base ones and the cutomized clip G along with long VIT, VIT-L14 and base clip L. Llamas I tried were base Q4 Llama 3.1, and Darkdol version. But I just cannot get good results from this model. Do I need to use shorter prompts?
What is your shift value set to? I've noticed some of this when the settings were misaligned and I've also noted that the GGUF versions behave a little differently as well so I've not tested it too much. How long are the prompts being used? I'd start by seeing if its improved by shortening them.
@c0ur4ge Thank you for responding. I have tried shift values between 3 and 3.5 in SD3 sampling, should I try higher values? I have tried varying prompt lengths, both longer ones and simpler shorter ones that use tags like old SD 1.5 models. The shorter prompts work better but I'm still seeing weird anatomy (extra hand, missing hand specifically), and sometimes (rarely) distorted anatomy. I also tried even more samplers like Runge Kutta but still getting the same results. I tried step values between 20 and 30, and I did notice much better results with lower step values oddly.
Also if it helps my system specs are lower/mid-range but runs HiDream Q3_K_M GGUF fine. Also btw, the Clip G works really well with both Dev and Full base models. My specs are an old Intel 12th Gen Core i5 12400F, 4070 GPU (12 GB VRAM) and 16GB system RAM. Not a very powerful system but runs both Q3 variants of Dev and Full just fine with both standard and customized clips.
@silverlinings29991791 Thanks for the details - let me check my settings when im back later and get back to you. The future of this checkpoint may be that I just simple extract our training differences into a LoRA and release that for ease of use moving forward given how unwieldy the Hi-DREAM checkpoints are but still evaluating. I'll take a peek after a bit here!
@silverlinings29991791 I guess the next question is probably which resolutions are you generating at primarily - anatomy issues almost always come up when the model is trained on a concept but is undertrained at a specific resolution. This can be compounded a bit when using "Dirty G" but I'd probably start with 1024x1024 on the same prompt and move up from there to see if you can push the problem into a corner.
@c0ur4ge Yeah I generated only at 1024x1024, and tried 768x1360, 1360x768. I can try more if you want.
A little bit of an update - I finally managed to get it working after reducing the T5 and Llama down to Q2, seems like memory was the culprit in this case. Q2 isn't great for quality but its working. What's weird is that GGUF is only slightly bigger than the Q3KM for HiDream Dev, that is also around 9GB and works fine with Q5 T5 and Q5 Llama.
@silverlinings29991791 Thanks for the update!
Hi, if I were to fine tune this model myself, can I then be able to use it for full commercial including the images for resale? Thanks
Can I run the pruned nf4 model with 8 gb vram?
It's definitely not going to fit completely in VRAM, I'm afraid. Hi-DREAM is mega heavy but we've certainly got variants of Flux/Chroma at Q4 you could have a jolly 'ol time with! :)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.