Original Huggingface repo for model at lodestones/Kromoss
Ask questions on the Huggingface repo if anything.
No comments will get responded to.
Kromoss is a full fine-tune of MOSS-Image-Base at
1024 px. It uses the same architecture, inference pipeline, text encoder (Qwen3.5-9B) and VAE (FLUX.2 small
decoder) as MOSS-Image-SFT, and this repo has the same
layout, so it is a drop-in replacement. The fine-tune mainly fixes the Base model's blurriness.
The repo also ships 8-steps-dmd-lora.safetensors, an 8-step turbo LoRA (data-space DMD distillation of
MOSS-Image-Base): 8 steps, CFG 1, no negative branch.
Description
FAQ
Comments (46)
Nice, that was super fast!
Downloading now
Does this improve the quality of male genitals (realistic)? The replacement penis loras I've seen so far suggest that Moss is very weak on male anatomy.
only 6gb to test lol. If its not yet it hopefully will be. and it's not weak at all. It's actually quite good, can do almost anything you give it with enough steps and resolution. however the penis variety is quite lacking. try out lodestone's turbo lora, make sure your outputs are around 2MP, 20 steps of res2m, and it's night and day.
first 6 or so tests have shown no apparent downgrades, and some slightly increased varience. looks like a plain upgrade
Most models are bad a male anatomy. The config space for penises is huge, its possibly similar to that of hands. The tiny sample size for this on base models is very limited as best, and nearly all earlier generations of NSFW models would usually have captions largely limited to simply "penis" and maybe size and less so some of the larger primary differences.
Penises are working Kromoss, but not that well by any standard. It will take time.
@WithoutOrdinary I have found, now that I have been able to do some testing, that both the base and Kromoss are better for male anatomy than some of the recent models but it takes a lot of steps. The penis loras for MOSS so far look worse than what I can get with base or Kromoss. Maybe it's because people are so fond of using too-few steps with turbo.
You know @Lodestone has their account on here too, right? Only a matter of time.
silver is one of core members of lode and original Chroma team. they did all the chroma flash loras etc.
The depth of field, bokeh, or blur that by default applies to photos is unbearable.
It seems to be a structural problem with the original model.
Civitai should add this to the generator
Moss doesn't have data on male organs. Do you have any of that here?
It generate only total crap if just replace unet-moden on kromoss. Do you have example workflow for comfy?
Try to increase CFG scale to 2-3. The distilled lora actually reduces steps amount but CFG depends on topic. Or try to use without distillation (just apply a dmd lora with weight of -1)
gguf quants?
For what? To get lower quality and slower generation speed? Get Safetensors
@would_r then buy me a new GPU
@KDA stop being poor
@would_r actually everything is fine with the quality.
@KDA you made a civitai account in 2022.. and you still haven't thought about upgrading all the way to 2026. The fault is on you. And good luck with this years boost in pricing. Your fault. Cope and seeth with inability to run a 5-6GB model. Spend less money on SaaS next time. SaaS jockey.
@axicec Most people here view AI as a fun hobby rather than an investment. Some struggle to make ends meet with the little money they earn (even less in third-world countries) and can’t afford to invest in a new GPU, which won’t put food on the table for them or their families. Maybe if you step outside the bubble of a rich, first-world kid who can buy whatever they want, you’ll see that there are people around you who can’t afford it now or in the future which is why they’re looking for alternatives that work with their existing setups.
@KDA it is a waste of time, trying to explain to a snob an existence of other human beings in this world. Deaf can't hear, blind don't see.
The issue of image blur has not yet been completely resolved.
I wonder if this is a side effect of Moss being a small model.
@ss9999 i think it is a side effect of realism. Almost any real image have blurred parts here and there... and model just reproduce this "effect". Especially when prompt tend on realism. (chromatic aberrations, depth of field, bokeh effect and so on..)
yea this is what bothers me with this model, everything else about it seem good, just that images sometimes come out as blurry.
@vdf433 I'd rather have blurry than overly sharpened.
I'm very curious about how the Chroma family is always able to effortlessly break through the limitations of the new model and create complete men.
I have been getting worse adult male eyes with this than with the base model. It seems to improve realism in some ways and reduce it in others.
@ss9999 i have fixed some of the eye issues by bumping the resolution 0.1 of a MP until it resolves. it add less time than other models, and generally fixes most of the weird eye artifacts i've gotten with kromoss (mostly just the eyes not getting proper pupil/iris's). it's a weird model
Why not a Chromoss? It is not Krea any more. :)
Is it possible to use qwen3.5_9b_int8 instead of qwen3.5_9b_bf16 for full diffusion model and is it worth it?
Int8 convrot is overrated. It has higher perplexity than Q8_0 as far as I have seen, which doesn't make it a winner in my book. I especially don't see much value in using a quant for a text encoder (nor a VAE).
@ss9999 it goes faster :), but for the encoder I think the only reason you should want the bf16 is if you're going to train with it, Like any time, for the most part
@makiaeveli Any hard data to demonstrate that?
@ss9999 demonstrate what? that int8 convrots go through the math faster than the comparison? or that training on it is worse? cause int8 will only go faster if you have right cards, dependencies, etc.
the reason you should train on the higher weights is so it will have a higher chance of working on lower weights too. Minimax was a clear aberration from that at the beginning, where people were recommending to train on lower weights. It likely won't be much difference at this point, though, just use whatever the training program you're using suggests.
If VRAM or storage is more of a concern, I'm lowering the weight of my clip before the weight of my model. Im just saying whenever training anything you should generally use the largest weights you can. Older model versions would show larger difference in these weights, like you'd generally want the bf16 of everything to make sure it'd be accurate on models at least before Krea2. but these more recent qwen encoders are so good that the lower weights generally don't lose as much quality as they used to. You seem to be arguing that the int8 manipulation is even worse than a gguf manipulation, but frankly these qwen weights are so baked I can't tell the difference since Krea2. I don't use Qwen or editing though.
I don't know what you want me to prove lol. I generate using the int8s on all modern models, but I only ever trained krea2 with the higher weights. So no, I have no data to back it up cause I didn't test, i just used the best practice.
@makiaeveli I'm not debating speed. I'm interested in quality.
"the lower weights generally don't lose as much quality as they used to.."
This is where actual perplexity data is helpful, so people can make an informed decision. I am not an expert in quantization but I have read that "edge cases/outliers" are what are targeted first. What if those are what is needed to obtain a quality result? I can see that if one is generating standard 1girl one might be fine with heavy quantization. Good luck with male genitalia, particularly in video situations involving a lot of different angles and movements.
"You seem to be arguing that the int8 manipulation is even worse than a gguf manipulation"
I have tested Wan 2.2 with Q8_0 and Int8-convrot and the latter was worse than the former. That's in terms of quality. The speed isn't as important to me because I have plenty of experience obtaining dozens of videos with things like turbo enabled and having no use for them, versus far fewer videos that took a lot longer to generate because the quality was higher. A higher pile of inferior results isn't faster, it's the illusion of speed. I'd rather generate one great video in a day than twenty substandard ones.
Before the INT8-convrot fad there were people who were absolutely sure that a version of FP8 was so great that it was pointless to use 16-bit precision for generations. As soon as INT8-convrot became the latest thing, though, suddenly those people no longer said FP8 is so great. (In fact, one of these people actually started bloating his/her model with more and more 16-bit data to try to compensate for the quality loss of FP8, the same FP8 that that person had lectured me previously about being all one needs.)
There certainly doesn't seem to be a shortage of 'experts' with little tolerance for opinions that are outside their range of familiarity. It's becoming really tiresome, to the point where I'm tempted to just not engage with any more provocations. I just got lectured and laughed at by someone who posted a penis that looked like the result of radiation at a plastic factory. This is because I dared to say that steps are needed for quality male genitalia. But, hey... I'm not the 'expert.'
@ss9999 lmao
you can download them both and test them. I literally just told you the cut off for when int8 convrot became good enough to not tell the difference. huffing "perplexity" results its meaningless. You're meant to use your eyes. If you saw some fried garbage, it's likely because they were using the SFT model, a shitty turbo lora, shitty configuration, or all the above.
I don't really know why you're stupid enough to think that I'm just "some expert" here lying to your face when you're literally the one abusing terms like perplexity and GGUF8 when talking about DiT models. I have nothing to give you because you're clearly incapable of being helped
like "little tolerance for opinions that are outside their range of familiarity." is one of the most hilarious fucking statements I've ever read. You're the fucking moron on this thread bro. lmfao wan 2.2 lmfao. yeah, and that moron that shared you some fried image was also likely a moron.
if you don't care about speed that's cool. but any fucking sane human being reading this thread should know that modern int8 convrots on nvidia cards with the right dependencies will give the same quality as bf16 at twice the speed.
you're literally masturbating to some analogy about your own experience and saying you'd rather disregard me. then fucking do it? and keep talking nonsense about perplexity?
another one for the block
Getting only very blurry and deformed results using a very simple workflow, a lot worse than the regular Moss model
I'm not having that result. It has improved realism for me in some things but eyes do seem too blurry.
it's the base model. if you're using the SFT model normally, that would make sense
This is a very early fine tune, some results may be subpar. Especially if sampling is don't poorly. We have no control over how Civitai's onsite works, if at all.
The team is unlikely to see any comments here. I only check Civitai once in a while.
It works at 25 steps CFG3, unfortunately a lot of the images are mangled. Increasing the steps to 50 doesn't fix it either.
if you add some negatives and use a more intense sampler (like res2m) you will start getting resolved photos. My day one photos were like anima-level. But with increased steps and better sampler, almost zero outputs are mangled. This is also a base model, so you really should try adding a turbo lora (like lodestones) at 0.5 strength to make sure it resolves. those three things, improving the sampler, increasing the negative tags, and adding a turbo lora, will make all your outputs work
@makiaeveli Cool, Thanks.
So add the lodestones 8 step lora at strength 0.5
Change to res_2m simple ?
and what would the negative prompt be?
@ferrrett33 one of my images here https://civitai.com/images/146069786 is a decent example of the changes i mentioned. i still use 20 steps, and use res2m+bongtangent with the clownsampler to get the "bongmath" and the eta/denoise controls, but its not really needed for this model so much. you could copy the nodes to get the pos/negative prompts. there are uploaded images in that gallery with the negatives though if you click to the first image (this one https://civitai.red/images/146069771). I'm sure you could try bumping the steps to 16, 12, 8 and see what happens, but you'll likely need to find a new "nice" turbo lora value
the best moss model.
I wonder why so many people say this is blur
just refine your prompt and workflow plz.
