Blur be gone? ZIT solution that helps a LOT
I posted this as reply in another post but it has been a definite game changer in using Hidream-O1 for me, i thought i put it up here:
I found a simple solution in a workflow from civitai, that works remarkably well. It takes the output image from hidream and puts it through a z-image turbo run with 4 steps (i use mostly 6) and 0.35 denoising. While copying the same prompt from the hidream image in. The results are pretty good even for the blurry or shiny skin. Even with disabling the upscaling step at the end (which takes longer than the z-image run)..
i changed it for myself to avoid another custom node (and add my own ones :P) but this is the principle:
the og civitai workflow site: https://civitai.com/models/2629261/hidream-o1-dev-2604-z-image-turbo-refiner?modelVersionId=2952028 It is nothing new but it works well here.
Although i'm fairly sure the negative prompt doesn't do anything on a cfg1 distilled model like ZIT.
before and after (6 step z.image, end-upscaling disabled):
as usual with these denoising tricks, you can trade of zit improvement vs more pixel perfect replication. With more or less denoising. (lower than 035 less improvement but less change, higher then 0.35 Z-image makes it more and more it's own thing). But 0.35 keeps the details, composition, subjects and so on pretty much perfectly intact . But even .20 or something is a dramatic improvement already.
of course hidream o1 makes things so low res and blurry, that Z-image HAS to invent some bits and pieces to fill the hires gaps.
I still need the SeedVR2 upscaler to get it really sharp/crisp (I used ZIT 0.4 denoise).
that workflow did some upscale model upscaling before the zit run, but i found that made it look worse. And isn't much faster than seedvr.
It's a bit of a shame Hidream o1 needs so much pampering. That is unbelievably low res and blurry for a "2Mpx model".
that workflow did some upscale model upscaling before the zit run, but i found that made it look worse.
For images with no humans or non-photorealistic it may be good - but for that we already have a lot other image models where the editing/ reference images don't lose details (flux2klein or qwen edit).
It's a bit of a shame Hidream o1 needs so much pampering. That is unbelievably low res and blurry for a "2Mpx model".
Yes, I thought HiDream-I1 was an alternative (including that NSFW alternative, which at least offered male parts, but HiDream-I1 was mostly ignored, got no LoRA love), but HiDream-O1 is just too raw/unfinished to inspire me. And it's rather slow.
qwen edit yes. It really is the solid workhorse still. It's also always ever so slightly blurry, but notwhere near as bad, and it is far more powerful otherwise. Flux klein? mhm... i'm always disappointed with how slightly NOT similar a referenced subject looks and end up running flux2 dev (fp8+turbo lora) which is a LOT better in that regard (keeping the person identical). I know there are consistency loras for 9b, but most community loras are trained 95%on woman not general purpose. and i'm not.
I think o1 keeps people better than klein9b out of the box. as long as no finer details vanish in the blur. But yeah o1 is kind of lost in the competiton. Also as a pure image model, however much you fluff the prompt up, it isn't very creative, diverse and it really loves to clone people
One thing it has going for it (and i kindof disagree on the speed for the model size) It trains REALLY fast. Both in tokens/s and in how quickly it picks stuff up. Sadly these loras dont work in the default comfy templates (that use the dummy checkpoint trick). they only work for me with the Saganaki22 Hidream-o1 nodes which uses a non fudged model (blur issue is the same though).