Hey everyone, First of all, thank you to everyone for supporting this project, and to the Krea team who made the best uncompromised base model. Unlike the previous run, where I had to un-distill the model, the Krea team kindly provided us with their real base model. The training is much cheaper than the previous Chroma run. Though cheap is relative, it’s still costing me about ~10K RTX pro 6000 hours (3+ months on 4xGPU machine) running locally, this includes a bunch of failed experiments and bugs during training. Also thank you to ComfyOrg for the 2 RTX pro 6000 donations ♥️
The Base Krea is also not collapsed which allows me to do OPD (RL) directly ontop of it. As a result, you will get a good generation out of the box.
You can download the un-quantized version here: https://huggingface.co/lodestones/Kroma
So what’s next?
There are several stale projects like my pixel space run chroma radiance and zeta chroma, but with the limited compute, I have decided to not continue them for now since compute is at a premium atm. Right now the focus is to create a better dataset (where about 90% of the quality comes from) while continuing to add more knowledge to kroma in the background. I do want to tackle video generation models like h3 in the near future once the dataset curation is done, but fine tuning the video model has astronomical compute cost so I'm looking for funding for more hardware to train a model. I firmly believe that hardware funding donations help a ton because we can indefinitely do more experiments in a sustainable way without having to worry about rental cost and psychological burden of “burned money” on failed experiments. I have a target of getting 4 more RTX pro 6000 but that’s like 60 grand in total which is an insane amount of money! So I don’t expect to reach this goal immediately, it's more about chipping in one card at a time at the pace of the accumulated donations. Every contribution helps, no matter the amount, this project simply isn’t possible without your support!
Support this project!
https://ko-fi.com/lodestonerock/
BTC address: bc1qahn97gm03csxeqs7f4avdwecahdj4mcp9dytnj
ETH address: 0x679C0C419E949d8f3515a255cE675A1c4D92A3d7
my discord: discord.gg/SQVcWVbqKx
Description
FAQ
Comments (42)
You rock!
I don't know yet for what I could use this model. For realistic images, quality is one order of magnitude worse than in the original krea 2
I am gonna use it for rare races, like Felkins, African wild dogs, etc. I treat it like a better furry fine tune of Anima with superb prompt adherence. Once the pose and anatomy fits, one can turn it realistic with other workflows.
PEBKAC
@kneiff can you elaborate a little more what what workflows/models can turn it realistic?
great model, thank you for the work!! it fixed the artifact issues I've been having with krea 2 turbo
What is Kroma?
Awesome work as always! For anyone wondering, LoRAs trained on Krea2 Raw seem to also be compatible with this version of Kroma (from short testing)
Yeah noticed that as well, although some of my LORAs reacted slightly differently to the baseline of this model so I had to tweak some of the scales. 4MP 12+ step euler simple with a single stage instead of the reference two stage upscale. Really happy with how it turned out, thank you for all the hard work Lodestones!
Thanks Lode! Please can we get it in w4a8 & NVFP4 formats 🙏
Thank you very much for creating this! Is the W8A8 version the INT8ConvRot one?
legend
Wooohoo! Chroma / Kroma is back, baby!
Loving it so far!
Thank you for the model. It's fantastic.
If it's convenient, could you also upload the bf16 v0.3.1 base ckpt here? Civitai now can train lora against custom krea2 ckpt. It would be really great if we could train LoRa against the kroma base online. 🙏
You want to train against the teacher model instead. I'll poke Lode about uploading the other model types here.
@WithoutOrdinary please do! we could post the model ourselves but it would be better if lode did it.
I think I saw a project called Zeta Chroma. If I remember correctly, it was on HF. I used it at that time and the effect wasn't very good. However, seeing that it was constantly updated, it shows that the author didn't give up. It's worth supporting
Flux Chroma was the true success of Lodestone. The fact they were able to find a different solution for Krea2 is amazing
Very high-quality model with insane prompt adherence. No other checkpoints in here come close to this. But it still needs a realism LoRA on top to balance out the skin textures.
for training lora do better train on this base ckp or raw Krea2 one?
Hugging face page says its better to use their teacher model, but mine Krea trained char loras work very good with it
@Form27b_Sentiment thanks
@Form27b_Sentiment there has already been analysis that shows this OPD model is weight to weight equal to the teacher model anyway. It's some form of lora merge, supposedly most of the weights in the OPD model are just straight up Krea turbo, with all the Krea weights front loaded like a merged lora. I don't think it would matter right now, but yes if the OPD model actually does slowly get merges from more training, then the teacher would likely be your best bet to ensure your model is still congruent with original Krea2 weights
https://github.com/aoleg/opd-script
really all this is to say is if you use Krea2 turbo and a Kroma 0.3 lora, it should be equivalent
Fantastic speed/quality. Its a mix of krea and chroma(HD?) with magnificent prompt adherence!! Also, i tested 10 most downloaded loras for Krea on civitai and they are all working. I underline, the speed of generation is fantastic even on "potatoes"
Hey @Lodestone
Do you think release a FP8 or a BF16 of it ?
thx ;)
What is the point of FP8 when there is INT8 convrot (which offers better quality than FP8 and is faster on ALL RTX-series graphics cards)?
@socksosock Mostly correct, but it may depend on some user cases.
If a user is right on the edge of VRAM capacity and their system RAM spills over into the pagefile, the loading phase for INT8 ConvRot can cause minor bottlenecks, but you may get away with FP8.
INT8 ConvRot is highly optimized for specific UIs like ComfyUI and Forge. If you try to use complex external custom nodes, third-party extensions, or less common Python pipelines, they will occasionally throw errors or crash when handed an INT8 ConvRot model, whereas FP8 is universally accepted.
On high end card like A100, INT8 ConvRot will be up to 50% slower, due to missing kernel fusion optimizations. Most consumers do not use these enterprise classification cards, though.
@socksosock Not just RTX GPUs, on AMD as well. int8 just supports more GPUs overall.
@socksosock There's also the matter of prompt adherence. In my head-to-head testing of Krea2 and H3's FP8 and INT8 Convrot variants, FP8 consistently understood weirder concepts better than INT8. (I also thought FP8 looked better, generally speaking, but quality is subjective so we'll gloss over that). Probably not noticeable for most people, granted, but it's a factor worth considering nonetheless.
I'm just sad @Lodestone did not answer any of us ... and so we have no news/info about it :(
@LoopingAi There are bf16 checkpoints available on Hugging Face (https://huggingface.co/lodestones/Kroma), and if you have the bf16 versions, creating fp8 versions yourself will be easy.
Feels like this chepoint brings out the best of my Loras. This is the only non official checkpoint I would recommend.
same for my Game Chracter LoRA
Quite versatile, right out of the box without any Lora's. I really do like it!
Incredibly good, just straight update to base krea.
I have been using a slower sample method of qin_zhang_2s+ddim_uniform at 8 steps, and then res2s+beta57 at 4 steps. Essentially a Krea2 sampling pattern. It's slower, but it pretty much guarantees the lowest chance for plastic flux face. The first pass I can play with the noise injection to keep variety and realism. The second pass makes sure to reduce some of the natural depth of field and motion blur of Kroma. essentially I get half the output, but they're generally twice as good. Euler can still give you some of these output quality, but consistently requires much more prompting. If you're using an LLM with a prompt pattern you can like, then Euler at 8 will likely still be superior. But for my little brain, I find this sampling method can get the best out of even short and vague prompts
this model is the one. best overall image model i've ever used and it's NOT close. can't wait for the updates.
what are good settings to use with the identity edit lora and is it supposed to work well with it?