Yes it is, you can read it on Hugginsface.

Silmas Not really, I just checked again and the HF link in the description has absolutely no mention of HD nor v50. Where exactly are you looking at?

Nexdoor bottom of description page says "P.S Chroma1-HD is Chroma-v.50"

Nexdoor Here, look at the message by lodestones, https://huggingface.co/lodestones/Chroma/discussions/99

Nexdoor https://huggingface.co/lodestones/Chroma1-HD/tree/main 

Kaalciv Silmas Sorry folks, I think you misunderstood what I said. This information is currently not in the links provided in the description here on Civitai. Both links you shared are not listed here either. My point is that having this clearly stated here would prevent lots of people asking whether HD is v50, as it's already happening on Discord, Reddit, here, etc.

Cheers

I just did it using Ostris AI Toolkit on Google Colab and it works very well... considering the quality of the base model, Chroma may be the next big thing :) (Qwen Image is also very good and Loras can be trained the same way as well)

Kohya sd-scripts works with it

Cryphe Thank you. Can you tell which base model to choose FLUX.1-dev or FLUX.1-schnell?

aramintastudio can you share the notebook please.

mag225658920 Chroma. there's a Chroma model type you can choose.

mag225658920 you need to checkout the sd3 branch in order to have the Chroma model type. On the main branch it's still not supported.

aramintastudio I managed to train it locally on Ostris AI Toolkit, on RTX 3090 it took 3 hours.

aramintastudio are you gonna reply or not?

flux union pro 2.0

Cryphe flux union pro 2.0不能使用吧？我尝试了很多次都是全噪点图，请问你是如何设置的？

I tried using this controlnet model and it seems it doesn't work for chroma.

Yes, it does! Just update your Forge to the latest version.

Akalabeth I feel sad for you using SD WEBUI

2P2 I use anything that brings great results — ComfyUI, Forge, whatever works. I don’t limit myself when it comes to tools or techniques.

20 steps

from 4 steps if you use the Flash version

this version works well over 20 steps but flash version which runs on cfg 1.0 and doesnt support negative prompt has good results at the usual sub 10 step count

It is an Apache 2.0 licensed and uncensored "V1 base model" incorporating some innovative solutions by an extremely small team with very limited resources compared to the labs and larger groups training and retraining stuff. So I guess the hype is more about its output capability relative to those facts.

It appears it became a bit overtrained on hi resolution towards the the end (49-50). It is currently being investigated with some retraining going on if I understand it correct.

StreamofStars Thank you for the additional info. It was no attack on Chroma, i more like try to understand how to use this :)

The hype is the LLM-lite clip plus the uncensored training set. You can write in real sentences and it's quite likely to give what you want.

It doesn't have the intense skin detail of Qwen, or the clarity of WAN2IMG, but it has a breadth of understanding none of those have. I use Chroma as my base image pretty much always now. SDXL is still the king, based on time, but Chroma's ability to actually understand prompts means it takes about the same time anyway.

Also people often miss the thing about Chroma being the absolute best quality when it comes to artistic, non-realistic scenarios.

They are obsolete, a one should use Flash instead

2P2 Flash model is worse, unfortunately. Flash LoRAs have baked-in styles which may go against the prompt.

The first link is 404

gabrielx whoops, the developer of that LoRA is moving files to folders, it seems. It's in there: https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main/hyper-low-step

dobomex761604 thank you! <3

how does NAG work? I saw the official nag chroma workflow and it has two negative prompts, one for normal negative and one for nag negative. What are their differences?

minecraft8964 I don't remember what the difference it because on some examples negative prompt goes into BOTH. So that's what I do personally, didn't see issues because of that.

You do realize this is a chroma model right ? Which is great on it's own, why TF would anyone wanna finetune another model based off of qwen ? I recommend you look for a qwen fine tune.

kakkkarot I think you misunderstood him, he's asking if the author of this Flux1schnell finetune called "Chroma" would consider finetuning Qwen using the same dataset in the future, seeing how Qwen is a tremendously more powerful base model than F1S. It's a great idea but probably not viable.

InvictusAI I am sure Lodestone could do it if there is funding secured for it. Chroma with its 8.9B parameters has so far cost roughly $150K to train. Qwen is a heavier model so...

InvictusAI ahh that makes sense, I'm sure lodes could do it.

InvictusAI Yes, I mean if the author has any plans to fine-tune another qwen base model in the future. I also really like the Chroma model. It can help me generate concepts that other models cannot. It's a really great model

qwen is 20billion parameter and flux schnell was 12bill so if chroma flux took almost 7months chroma qwen would take a year

jimecherry  yeah 20 Billion worth of overrated horse crap.

Hyped to see one of your finetune on top of it

you are finally here, cant wait to see what you will do

Certainly it is not, and some concepts are not there.

Silmas I mean, if the model doesn't do what you want it to do - it doesn't mean the model is bad. It only means that the model is not for you.

2 days later after my initial comment I still stand my ground - so far it's been pretty OP for my use cases. Even the most niche fetish gens are possible to gen with Chroma. Like, really niche ones. You do need to prompt it and Chroma does prefer longer prompts, but that's not that much different from having to adapt to pony or Illustrious prompting.

And I'm not even talking about different aesthetic kinds like dreamcore, weirdcore, liminal space, pinterest aesthetic with glitters and stuff, etc.

Thanks to T5, the prompt adherence is great, and if you also load fluffyrock clip-l alongside with it, it only gets better.

coyotte any plans for Chroma-based Lustify? :)

coyotte did you try last dev versions? https://huggingface.co/lodestones/chroma-debug-development-only/tree/main/HD

they are MUCH BETTER than uploaded here version

ArtisanDreamer I'm still on version 48 as v50 was really subpar in my tests. I need to give these a try, thanks!

As for the chroma version of Lustify - it took lodestone 7 months to finish training on 512 res on enterprise level gpus. Realistically speaking, I don't have nearly as much of compute power or financial capabilities to afford that, so for now that's out of questions. However I do want to contribute to Chroma ecosystem... Tryna find ways to do so.

coyotte The model is supposedly very receptive and should not need extensive and expensive runs to finetune. (*edit -even more so now if his VAE-free pixel space version pans out).

78 public posts = made about 100?

Not interested in your random human loras

If you want Civitai support there kind of needs to be a sufficient number of loras categorized with that model type. If you never post them it never gets categorized and then theres never a category

I heard that Patreon blocks everyone who posts AI realistic content and AI content with real people?

Stokky I have had no issues as long as NSFW is behind a paywall they seem to be fine i have had my account 2 years now

makiaeveli I have posted one, linking chroma as the resource but there not much else i can do, civitai won't know its for chroma as far as i can tell  

你是否考虑把chroma的训练过程制作个教程，我到现在还没有在网络上找到可行的或者比较好的完整教程，其它模型好多都有，chroma几乎没有。你是否会认真考虑这个？

Which setup / params are you using to traing your Loras?

Can you please tell which tool you are using for LoRa training? I have never trained one but chroma seems to be the perfect base to make some.

@winasrar Chroma has been a breeze to train with. Only tricky part is captioning your images with a natural language prompt, and even then, chroma supports tags, so you could just use tags if you really wanted to.
My current trainer has been ai-toolkit, with the default chroma setting (also makes training SDXL models easy). You don't even have to change any settings. My "step count" advice would be to do around 216 or 256 steps per image.
Even on a 4090, training time is 2 hours with 15 images for 3,000 steps, along with full 24GB VRAM usage. Good luck.

@Jack_Purley I'll second the recommendation for ai-toolkit. I'd never used it before and gave it a shot specifically because of its chroma support. Very user friendly gui, easy to set up, etc. I think it sat at around 21.5 GB VRAM use for me through most of it. Though to reiterate OP's comment, I really just wish civit would make a category for chroma at this point.

What is your patreon?

I've seen Kohya scripts can now train on Chroma

Yes.

In fact if Chroma wants, it'll bypass negatives related to photorealism. I received images even less realistic and less detailed with your suggestion.
https://civitai.com/posts/21111540

damn rip

typically anything remotely cartoon related will pull the model to cartoons, jsyk

also you're using tag prompts, which are gonna typically gonna be worse on a flux model

typically anything remotely cartoon related will pull the model to cartoons, jsyk

Sometimes does, sometimes not, dependend on a bunch of factors. Evenmore, I often recieved cartoonish/anime images for sci fi with nothing cartoon related in vers.38-46 (they fixed this further). It seems that the checkpoint heavily tuned towards anime/cartoons so just need someone to release finetunes towards realism.

> also you're using tag prompts, which are gonna typically gonna be worse on a flux model
TBH hadn't found sufficient difference for me.

Which "hacky nunchaku quant" are you using?
From this repo?
https://huggingface.co/rocca/chroma-nunchaku-test

EliteLensCraft yeah that one, there's a reddit thread with a workflow if you need it

okay watch this video: https://www.youtube.com/watch?v=egn5dKPdlCk&t=659s

biznis Thanks, I'll check it, want to compare it to my current setup with a 12 step low step lora from https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main

Sadly, the only way of getting good results for now is going for the slowest settings.

Euler is fine with beta or normal, but you need to go fp16, default weight type, not fp8, and high amount of steps, at least 30 to get good results consistenly. Also don't use the low steps loras.

Also Cfg has a big impact, you need around 3 or 4.

Euler/Beta or DPM2_a/Beta ; CFG = 4-6.

Just need to wait until it gets fine-tuned for aesthetics. Base models are rarely reliable for generation. But it has all the right ingredients.

@piconejo honestly I think it also needs multistep res Ala RES4LYF

Sounds reasonable, but also like very time consuming....Like 1-2 min to generate a pic?

@LovelaceA It can get up to 4 minutes

@piconejo That is sloooooooooooooooow....

This is just too slow!

In generating real-person images, I have made numerous attempts. Eventually, I found that approximately 26 steps yield the best results when using Euler Beta. If I set it to 30 steps, the images always end up being overly saturated or over-sharpened.

@lukeleo Thanks for the info. May I know would you recommend use the FP8 version or the FP16 one?

best results for me are on: res_2s with sigmoid_offset and 20 steps. As those require additional nodes like res4lyf package and the sigmoid package, it might not be for everyone, cfg is 4-5 for me atm mostly 5 but back on v48.

Can second the "do not use low step loras", for me those just destroyed what the model did best...

as I am on a 3060 12gb, I use the Q8 version and it is totally fine... fp8 is too big to fit in vram and takes ages as result, fp16 would be a dream but far from possible...

As for speed, on 1472x1024 and image takes about 6 min. I am not pressed for time, so that is totally fine by me but if people are complaining on 2 min, chroma might not be for you... in contrast to hidream, flux and qwen, chroma is the slowest but on my old system those models also take about 3-5 min per picture... and that is with sage attention on those that work with it (qwen does not like sage )

@LovelaceA I have only used fp16, but I believe fp8 may produce poor details. For me, the quality of the Chroma fp16 without adding any Lora is one of the most satisfactory models for me.

Give it a weak or so and it might clear up.

Chroma is very much the result of research. Constant experimentation and iterations is the main reason behind the various models you describe.

Chroma1-HD (v50) was originally meant as the finalized version (with the annealed version as an alternative variant of the final model) but it turned out a bit premature and did not stick the landing as well as the dev had hoped. So there has been some restraining on the works from v48 that are much more promising (like chroma-debug-development-only /HD and the new VAE-free Pixel Space (instead of regular Latent space) model that just started coming together.

fp8 learned svd - https://github.com/Clybius/Learned-Rounding
heun - The real name is Flash, an official distilled version. It's actual technical name was flash-heun-8-steps-cfg-1
merged - It's then a one breeds various models, probably not official. They are the versions which should be really different in your opinion!
I think you should use Annealed with a realistic lora

@2P2 are there realistic lora´s for Chroma? i found no one

@nanunana With the right prompting sometimes you wouldn't need a LORA for realism at this stage Chroma achieved.
But sometimes it does get the wrong idea or mixes styles, so a couple I use are:
https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/blob/main/Chroma-ProjReal_lora_rank_64-bf16.safetensors
https://huggingface.co/MisterMango23/chroma_realistic_lora_style
the second one can really force realism at 1.0 or above, but could get messy. So try lower weights at first. I don't know why it doesn't get much attention

@ailu91 Thanx for the links, but these Lora´s are bad.

After many experiments, I found that using simple descriptions like 'this is a photo' when describing the picture is easier to produce a realistic photo than using more detailed descriptions, especially words like 'photolistic', which can make it less realistic. I think for Chroma, the more these words are added, the easier it is to identify as a realistic painting.

mark

Hi. What's new about this version ? TY.

@Kawai_ It is an improvement addressing issues with the previous version that did not stick the landing as expected. The previous chroma1-HD was overtrained on a single resolution (1024) which caused some issues. This new one is trained on 512, 768, 1024 and 1152 res giving it capacity to "generalize" better output at various resolutions including above 2mpix.

How is it?

@biznis in my opinion and after some tests. I still think the version before v49 are still better for realistic images. better than the initial release of Chroma1 HD but hands, fingers and feet are still not good.

@Beezer79 I tried v48 and HD v2 on the same workflow and v48 was way better. I was really disappointed. Maybe HD needs different settings, but from what I tried v48 was consistently better. HDv2 smooths out the fine details, the image looks simplistic like a low-step model.

You can try 'huge, erect penis', for me, if I use 'huge penis', it is very difficult to get a flaccid penis

@lukeleo  flaccid soft plump virile dangling hanging cock, plump soft squishy bent penis, penis is soft and bend forwards. Perfect penis and scrotum anatomy.

if you say too many things surrounding "small" it'll turn more into a clitoris depending on the other words in the prompt. That's what the "virile" "plump" are for, although idk how important each word is anymore

What resolutions are you using. HD should have more info on higher res. But the model itself is meant to be a base for further fine tuning. Of course training could have continued in more epochs if Lodestone had the money for continued training to spare.

Try Chroma1-HD-dc. It uses v48-dc weights instead of standard v48 weights.

@StreamofStars you mean this version? https://huggingface.co/silveroxides/Chroma-Misc-Models/tree/main/Chroma1-HD Thanks for the hint. i give it a try.

@Beezer79 Yes. And yes, for some resolutions and prompts I prefer v48 over the New HD too. But my guess is that the HD versions will be easier to finetune on higher resolutions

This is becoming increasingly confusing, and v48 remains the best version. Why not simply state that this is the final version and provide individuals with the opportunity to continue training or prepare loras for it?

@nanunana v48 was only trained on 512x512. (you can bypass a lot of compute if you let a model learn new concepts in lower resolutions before finishing up with a few high res epochs. If you are referring to v48-dc it had some 1024 training I believe. In any case the HD models are those with the most "knowledge" and might be more sensitive to prompting . In anyway, these are all base models designed to be finetuned. You can finetune v48 too if you wish. All of it is open. But it is messy as it is a "non-corpo" dev that is experimenting and innovating as it is developed.

I have no problems with v49+ (same-ish result). And there are already some users who trained Chroma Flux loras. HD will not be final, just a checkpoint

@StreamofStars Thank you for pointing out the two HD-dc versions. I tried them, and, sadly, they suffer from the same loss of fine details. The images are more simplistic, show repetitive patterns, and slightly blurry. So, v48-dc is currently the best version. For example, when I generate a photo of landscape with flowering rolling hills, v48-dc draws individual blades of grass with shadows, variation in grass density, and stems going to flower petals. With all HD versions the grass turns into a regular green pattern, especially in the distance, and the flowers are just splotches of color flying in the air. It just looks more like a painting than a photo.

@huc2cvt678 If you're using comfyUI: Try setting min_padding = 0 and min_length = 256 or 512 in T5TokenizerOptions. See if that makes a difference. (I've noticed that quality can be degraded if negative and positive prompt have different length without padding, and the effect sounds similar to what you're describing.)

@huc2cvt678 Same her, fotorealistic don´t work with HD-dc

@rybkon417 I already had those tokenizer options set.

can't relate at all. HD looks better for me with euler simple

Are you sure your using the newest one? Seems like civitai is not updated. It looks much better at higher resolutions.

I can confirm your observations. As a rule of thumb, I always use the FP16 precision of any text encoder since I just load it into system RAM of which I have plenty. This frees up VRAM for the diffusion models. System RAM is relatively cheap these days (and much cheaper than VRAM).

i want to test this flan-alpaca, has anyone a link i only find very big versions for llm´s...

Flan appears to be the best option, I have been using it.
I know GNER, it's just not worth it because it's specialized for named entity recognition, which is not useful for text encoding.

t5xxl unchained doesn't work with Chroma, the clip is too different and is limited for now.

Also check t5 efficient xxl, but Flan beats it, try anyway.

There are other finetunes, but we need a NSFW one.

---

To use fp16 instead of fp8 or quants is just a recommendation. It's worth it to save some (V)RAM, and it is possible to get 5 digits on each thingy even with low quants of t5, I don't have to reroll for better anatomy

I'm using flan_t5_xxl_TE-only_Q8_0.The gguf that here https://huggingface.co/easygoing0114/flan-t5-xxl-fused/tree/main

Of all the encoders that I've known, this one is the most realistic

@nanunana Here's where you can find Alpaca:
https://huggingface.co/Clybius/t5xxl-encoders-comfyui/tree/main

@2P2 I mean sure you can get good anatomy with fp8 or a quantization. I am just saying you'll consistently get complete anatomy more often with fp16. I spent hours testing. I can take a prompt in fp8 with one encoder, then use it's fp16 version and watch the digits appear. It's not just digits, rooms will get more details, fuzziness gives way to clarity. Every one of those encoders I have both variants (besides alpaca), I can see the difference of fp16 over fp8.

--

In summary yes you can get good results with fp8, but with fp16 you encounter good results more often.

@mrsanders1313840 I'll check it out, thanks.

@mrsanders1313840 That should be the same encoder as the flan-t5-xxl I've been using, but it's definitely different.

https://huggingface.co/easygoing0114/flan-t5-xxl-fused/blob/main/flan_t5_xxl_TE-only_FP16.safetensors

Verses:
https://huggingface.co/silveroxides/flan-t5-xxl-encoder-only/blob/main/flan-t5-xxl-fp16.safetensors

Easygoing is missing the encoder.embed_tokens.weight

@2P2 Here are my thoughts on GNER.

For what ever reason lots of people who are messing with image gen are interested in it.

Silveroxides has a repo dedicated to it.

https://huggingface.co/silveroxides/GNER-T5-xxl-encoder-only

You can find it on this site as well.

In fp16 is produces great results.

It seems to adhere to negatives better.

It has better recursive/self referential adherence.

--

High-quality photography. Film photography, film, film grain. Realistic, clear details, 4k. A high definition, realistic photo.

Woman(A) 25 years old with blond hair.

Woman(B) 30 years old with black hair.

Man(C) 40 years old, and is over weight.

All Actors on couch.

Woman(A) is sitting on the left side of the couch sticking her tongue out.

Woman(B) is sitting on the right side of the couch winking.

Man(C) is sitting in the middle laughing.

--

GNER will follow that prompt better than any of the other text encoders.

@zoot_allure855 your T5 testing and sharing the info made quite a difference, I THANK YOU FOR YOUR ATTENTION TO THIS MATTER!

You have to guide the model a bit, mate. If you use words like "candid photo" "selfie" or photographic terms you will get realism, if you use "anime screenshot" then you will get anime, and so on

@Dantor makes no difference thats whats annoying basic prompt,

photo realistic, female expression of anger bearing teeth mouth open reflection of fire in eyes, hyper realistic, high detail, high contrast, sharp image.

very simple prompt

@StreamTabulous yeah if you use stuff like "hyper realistic" you will get digital paintings because that term is associated with it. When you describe a photo you wouldn't call it "hyper realistic", would you?

@Dantor oh thank you, sdxl was always photo term and chatgpt thanks ill move away from that. and yes i described my photo Restoration work as that as where full restoration they where not a photo anymore with adding skin texture fabric texture etc id say they are hyper realistic recreations

Read this thread to get it
https://www.reddit.com/r/StableDiffusion/comments/1j4biel/comment/mg81j11/?context=3

Eat this if you have to, lol

@2P2 and?

groomer needs mental checks

Crazy talk. With a few LoRAs, I'm getting better images than Flux Dev and Flux Krea. Check out the latest images on my profile. This model delivers and it's not even fine-tuned yet.

Si tacuisses, philosophus mansisses!

thanks for the link. after my first test. the v2 version works very well for me and is really fast on my test notebook with an old Quadro RTX 5000 16GB mobile GPU. 12 steps (8 steps are to low in my opinion) cfg 1, euler beta. no loras. just the default comfyui text to image workflow.

It brings a new issue > torch.OutOfMemoryError: Allocation on device
if your computer can't run fp8 unet

That is a basic merge for now

thanks so much for linking this, do you know what part of this is actually speeding up inference speed? like my s/it is way faster on this, about double speed. I'd like to try to apply it to the main model if i knew what it is. (i mean i would like to try to use the main model, but it's actual speed is too slow)

@TheNecr0mancer 

Looking at the safetensors metadata for this merge you'll see it is a merge of

Chroma1-Base.safetensors (48) into Chroma1-HD.safetensors (50.rev1) at fp8_e4m3fn.

They used t5xxl_flan_new_alt_fp8_e4m3fn_scaled.safetensors as the text encoder, and two low step loRA;

- Hyper-Chroma-low-step-LoRA.safetensors (.4)

- chroma-flash-lora_r32-fp32.safetensors (.6)

You should be able to find all of these in the following repository.

https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main

I had tested some more. with a cfg of 2 you gain a way more better image quality for a acceptable more of generation time.

16G can use？

@lilililili123 yes. the model uses around 14 GB of my 16GB vram.

@Beezer79 wow yeah cfg 2 is great, but you need to be careful of not using too many steps or it will overcook

@zoot_allure855 Thanks for the heads up! With Chroma1-HD Q8 GGUF + Hyper-Chroma-low-step-LoRA.safetensors on a RTX 3070 32gb, 1024x1024 12 steps generation time takes less than a minute!

@TheNecr0mancer yeah! i always stay at 12 steps. more brings no positive impact for me. and over 20 steps its looked really overcooked.

I wholeheartedly agree with you that a guide would be excellent. There's a lot of conflicting information both here and elsewhere about what works and doesn't about and with Chroma, and it's been a frustrating to navigate as I've gotten started with it.

I'm using the sliveroxide v47 Flash Q8_0 version of the model -- been running it for about a week now -- and when I first started messing with it, some of the misleading info I've run across have been things like "the Flash versions ignore negative prompts" (the don't according to my experience), or "add the fluffyrock clip-l for better adherence" (Chroma appears to use t5 exclusively). Best course of action for me has been reckless experimentation to help figure things out as I go along.

So far, v47 Flash Q8_0 with heun/simple and the Comfy t5xxl_fp16 text encoder is working far better for me than the vanilla version of FLUX or C4pacitor, but with occasional problems with bad hands, feet and body merges. Up it from 8 steps to 12 or 16 with the same seed and those problems are usually resolved with the next generation. It's also slow as molasses to generate, but I could easily blame that on low-spec hardware running stable-diffusion.cpp with the Vulkan backend.

Until or unless a guide is written, a general concept people can think of before diving in would be "Think of Chroma as the love child of FLUX.1-Schnell and IllustriousXL." Use prompts with complex and specific scene descriptions, concepts like mood and feel, *booru tags, or any combination of the same. Use negative prompts like you would with Illustrious, too. Then, just play with it. Better living through reckless experimentation, and all that.

@MrSnichovitch Try another encoder, its results are better in my opinion. This model is highly dependent on the encoder. flan_t5_xxl_TE-only_Q8_0.gguf that here https://huggingface.co/easygoing0114/flan-t5-xxl-fused/tree/main

@mrsanders1313840 Already have. Tried both flan and GNER t5's and they don't make any notable difference over the standard encoder to me. Highly likely that I need to work on revising my prompts (I generate using descriptions of elaborate scenes rather than defining character portraits) before I would see a difference, but I've got everything I need for testing when the mood strikes.

There's the official distilled version lodestones/Chroma1-Flash

These speed loras work well, tested

https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main/chroma-flash

https://huggingface.co/silveroxides/Chroma-LoRA-Experiments/tree/main/hyper-turbo-flash

@Llewellyn thanks i'm looking at it

@2P2 nice thanks, what are all these different versions? i've never seen files labeled like this.

@Llewellyn lol the didn't add a single detail on the settings you should use with it.

this was in another comment
https://huggingface.co/Phr00t/Chroma-Rapid-AIO

@TheNecr0mancer Chroma Flash recommend 8 steps with CFG 1.0. And it's not perfect, breaking on common small details like hands. Flux.S sometime did better and faster (4 steps). But still IMO this is fastest uncensored model with closer to Flux capabilities.

@Llewellyn oh I didn't notice that breaking, but if you check the post I made you can see the type of jssues I've been running into

It is very easy to train. It loves to learn.

Easy
sd1_inpaint_example.json
or use MaskDetailer from Impact Pack

Tenofas’s workflow has inpaint and other modules built in. You can see how he set it up here: https://civitai.com/models/1582668/chroma-modular-wf-with-detaildaemon-inpaint-upscaler-and-facedetailer — it includes DetailDaemon, an upscaler, and a FaceDetailer.

At the moment we're in very early stages, this page doesn't even have the current models. Many cannot get past the new dialect/syntax required by this model. And, it's a tough sell to learn it when you have so many fantastic SDXL models that can produce amazing results with just a few danbooru tags.

That being said I know there are very talented creators that are currently working with this model.

For example:

https://civitai.com/user/SG_161222

Has been messing with it since the first HD dropped.

--

Once the fine tunes start dropping things will change.

@zoot_allure855 I really hope so. Even in its current 1.0 version, this model already demonstrates incredible capabilities. It follows the idea beautifully and responds to every adjustment in the positive and negative prompts. With enough effort and some basic knowledge, you can already create just about anything you can imagine.

@zoot_allure855 Evgeny (SG_161222) is one of my favorite model creators! Is he working on a fine tune of Chroma?

BTW, what's the best resource to learn more about Chroma and how to use it to its full potential? Meaning best practices, settings, prompting guide, limitations, etc?

@zoot_allure855 Agree. People stick with the models they already know, and learning new prompting can be hard. SDXL is a good model with solid fine-tunes, but in my opinion a lot of SDXL images end up looking pretty similar. To get something truly unique you often need more than one SDXL fine-tune (if you’re chasing different styles and aesthetics) and a huge LoRA folder. My 1 TB SSD is already full, send help 😭.

Tastes differ. Use what works for you. For me, Chroma covers both stylized and realistic in one model, usually without LoRAs, and it understands “human language”. I’m waiting for fine-tunes, but I’m sticking with Chroma because it already does what I want and more.

No Chroma base model type, but fucking Google's banana got its own one very fast. literally

@mmdd2543 

I do not know for sure if SG161222 is working on a fine tune. On hugging face he liked release 48, and release 50.

https://huggingface.co/SG161222/activity/all

He published a repo for chroma but then deleted it couple of hours later the day chroma hd released.

He's currently working on a fine tune for ostris/Flex.1-alpha. He appears to be doing this on a single 4090, so we wont know for sure until he finishes with Spark.1V (ostris/Flex.1-alpha).

He kind of has a pattern where he likes models, then starts working on fine tunes for them; time will tell.

--

I am not sure if any Chroma documentation exists out of this location and loadstones repositories.

--

I'd also like to point out that loadstones did not stop with the HD release. In fact he appears have step up his work since the release, pushing 20-30 checkpoints a day.

@zoot_allure855 "I'd also like to point out that loadstones did not stop with the HD release."

Wow! Seems like he's busy. Looking forward to what's next!

@mmdd2543 What's next is probably his work on Radiance- It's Chroma but without going through a VAE, directly in pixel space.

This location has the first release 50, which loadstone redid as it was over trained. Hugging face is the newest one.

There also is:

https://huggingface.co/lodestones/Chroma1-Base/tree/main

Which is release 48 and what was used by loadstone to remake 50; I prefer it to both 50 releases.

@zoot_allure855 Interesting. Thanks for the heads-up! Maybe this is why some people are complaining of bad images? I have been using v50 from Hugging Face up until now. I will check v48 as well. Hopefully he updates the one on Civitai so that more users can experience the improved model.

@mmdd2543 If you are a noob and can't use decent models, you get bad images. It's a rule

@2P2 To be fair, it's not their fault. Both the Civitai and the Hugging Face models are named the same so you'd think that they are exactly the same, when in fact they aren't. Not many people have advanced knowledge to check hashes and such.

the rev01 model is the newest one and the proper version, the older 1hd is v50 which had some issues and had to be recreated from v48/base. the new rev01 is v48 with higher resolution images trained in, so does have less artefacts than v48 (the color issues on the borders)

Whoah, that’s nice, It's strange I didn’t see this post earlier

https://civitai.com/images/97380155