I've always liked Flux even though it is heavily and unnecessarily censored. Chroma is basically more what Flux should have been from the beginning. It is based on Flux Schnell and has all of the benefits of Flux without most of the Flux censorship.
GonzaLomo Chroma is a blend of Chroma 1 HD with a number of LoRAs, including some of my own. The primary goal was to try to retain Chroma's creativity while making it more consistent and stable.
The images I post with this model are created using my workflow which generates the image using GonzaLomo Chroma and then does a light refinement using GonzaLomo v6.0 Photo XL. The amount of refinement is configurable by changing the denoise value on the refiner's sampler.
Description
This is a Q4 quantized version of the V3.0 UNET
FAQ
Comments (104)
Great... Now I have to start messing around with Chroma 馃檮
You won't regret it ;)
@GBRX聽why do you (and someone I assume borrowed from your workflow) have an SDXL refiner? What does Chroma lack that I need to address?
@lonecatone23聽As good as Chroma is at color and composition, SDXL is still better with some fine details, especially NSFW details.
Okay...why does Chroma say 20-30 steps and a cfg of 3, and yors cooks (quite well) at 10 steps with a cfg of 1. Am I missing womething? I've spent hours buiding a workflow and have no clue wtf is gong on
@lonecatone23聽I baked in the Chroma Flash lora
@GBRX聽and boobies apparantly. Lots and lots of boobies.
I'm so proud of you 馃ス
Do you have a normal version? I'm trying to figure it out.
@lonecatone23聽I don't have a normal version but the lora I used was chroma-flash-heun_r128-fp32.safetensors. So you can probably find that version and apply a negative value to it to create a 'normal' unet. I would start with -0.85 and adjust from there.
I cannot seem to be able to run the V3 Q4 .gguf model in ComfyUI, I always get "Could not detect model type" when using Unet Loader (GGUF) node, the same error appears when using ForgeUI Neo in the Flux pipeline.
Sorry, I don't know what the issue could be. I normally use the GGUF Loader node. I just tried it with the UNET Loader (GGUF) node and it also worked.
hmm so at the risk of staten the obvious in hopes of solving your problem folder placement is paramount - did you place the model in the "UNET" folder , NOT the diffusion model folder or checkpoint folder for the gguf version - ????? i've used both the v3 q4 and bf16 version for about 5 hours last night and they both ran flawlessly in ComfyUI - i have no experience with ForgeUI so I can't comment on that * but assuming you have it in the correct folder aka "UNET" - it could be a corrupt download aka your ISP had a minor hiccup when you downloaded and then resumed - this website has been acting funny the last week , so if all else fails re-download it .
** Also worth pointing out - ComfyUI has had alot of bug fixes in the last week they made a rather large leap going from version 13 - to 14.2 with many patches inbetween when all the AI audio stuff came out , i'm currently still usen comfyui v13 because i was spooked by how many bug fixes they were implementing upstream - so its possible if you have the latest and greatest version of the program that they're still ironing out things aswell * (it happens)
My bad, I had old version of ComfyUI, updating it fixed the issue and now it works. As a side-note - I am really, really thankful for the Q4 model, because I always wanted to try Chroma but on my 4GB Vram laptop I was stuck at SDXL at most. Amazing work.
@AlternativeAlter聽Awesome, glad it's working for you. I honestly don't see much difference in quality between the Q4 and FP8 versions.
I'm getting the "could not detect model type" for the fp16 gguf. I don't get why its a gguf but I can run all the other uploaded models here.
@Suhnny聽https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.15.0 , released this 21 minutes ago , ya they're flying through updates atm the original poster of this troubleshooting thread found success in updating, you might try that too since they literally 21 minutes ago released a (stable) v15 - from my understanding when they go a full version like 13, 14, 15 , as opposed to say 14.2 which is what it was at this morning - that means they feel confident enough in the release to deem it (stable) . so yeah suggest updating :) update** like a good little test subject I took one of my older machines that had an old version of comfyui on it and tested to see if the old comfy-gguf custom node would work with the new model - and i replicated your problem - the fix is in fact updating ^.- hope that helps anyone whose reading this.
@lin_mei872聽I'll give it a shot but last update broke my stuff and it took four hours for me to figure out how to roll back. So I'm gonna check out the github release for what I got now first and keep the current_requirements file to make it easier in case.
@Suhnny聽, after you update comfyui you're gonna wanna check for updates for the custom nodes specifically the comfyui-gguf custom node by city96 . yesterday after comfyui update released the v15 version - city almost immediately released a "nightly" version to his custom node , you're gonna need that for the model to work
@lin_mei872聽Update worked fantastically and now the 16bit model is running great. This is a more stable commit on 2/25/2026 it updated to. If anyone having trouble, this version is definitely one that works
Using simple workflow without any refiner for V3.0 Q4. gguf because I only had 8 GB vram, so far so good. What I didn't expect was that it can work with flux character lora, I have expecting it wasn't compatible based from what I read online about Chroma, but still trying it and after trial and error, it work in some extent. Character lora with 1.0 strenght will give blurry and too grainy output, but 0.4-0.6 seem to be a sweet spot, granted with that strenght, sometimes it wouldn't be 1:1 of the character, but it keep enough resemblance
@lin_mei872聽Chroma couldn't in FLUX loras properly till v45/46. Regard character loras: for example it can't into this lora while regular FLUX checkpoints does.
@lin_mei872聽you misunderstood me a bit.
As I wrote in another comment, I'm with Chroma since v40 and tried each version released. In v40-45 Chroma didn't support FLUX loras properly (I really spent time to generate pictures w/without lora with a fixed seed to compare). The situation changed in v46. But even now the checkpoint and its derivatives has difficulties with FLUX character lora (I gave an example earlier) so for characters I use regular FLUX checkpoints.
It's not a surprise, somewhere in early comments to the vanilla there was a warning that working of FLUX loras is not guaranteed due changes in the engine and one has to check exact lora's compatibility with Chroma.
@lin_mei872聽everythning is fine, English isn't my native language too :)
I've found the comment about FLUX loras, look in comment section: https://civitai.com/models/1330309/chroma?dialog=commentThread&dialog=commentThread&commentId=931628
Probably this user (ailu91) can give you more details about compatibility of FLUX loras on Chroma.
> I would say you'll find much more success with Chroma1-HD than you will the original line of Chroma in both capability and compatibility <
Typically I use the subj (GonzaLomo Chroma) and this finetune. Chroma HD I use on TA.
@lin_mei872聽A question: what controlnet usable for this one? flux controlnet?
Thank you Sir for the Model bf16 it is really good and it is able to create complicated pictures, Quick question since it is the same size as the regular BF16, why did you choose to create a GGUF one,
Thank you again great work
Ah, sorry, that should be labeled as FP16, not BF16. I just fixed it.
@GBRX聽Do you have a safetensors fp16 version, or only the gguf
@carterj099聽Looks like only gguf
Could you try uploading an undistilled model so I can train Lora directly on it? This is the only Chroma model I've been using lately. It would make sense to train Loras on it too.
Thanks, I'd have to investigate how to create the undistilled version so I'm not sure yet when I'd be able to do that. I would say that I create loras based on the official Chroma HD1 model and they work great with my model. In fact, most of the improvements I've made to my model have been from loras I created this way and then integrated into my model.
@lin_mei872聽Probably at least 16-24
you really should always use the base model to train, not a lora merged model.
@crombobular聽Yes, I understand the reasons for this. For large, popular base models, this is certainly true. But the launch of Chroma1 was botched and ill-timed, resulting in many unfinished, slow-running versions and the relatively fast Chroma1-Flash, which is completely useless. That's why Chroma1 doesn't have a base model. I have three official Chroma base models installed: HD/Base/v46, and I can't fully trust any of them. Chroma also has the unfortunate tendency to try to generate anime whenever possible. I only need a Chroma-level model for photorealism and complex hyperrealistic art, so half of my prompt size is an attempt to avoid generating anime and other art. So, which one should I train LoRa on? I think you should train it on the one you use most and that doesn't try to generate anime for you first in response to a simple request (there are plenty of excellent dedicated models based on Ill or NAI for anime). I was planning on making my own base model for my own needs, but I just didn't have the time to figure out how to get rid of the annoying anime. But with the release of GonzaLomo 3, I've lost the motivation to do so, so I think using this line as a base would be an interesting experiment. At the very least, it would be worth a try.
@Unicom聽"So, which one should I train LoRa on?" I already said. Chroma-Base if 512 res and HD if 1024 res. It really is not that complicated.
@Unicom聽I currently train LoRAs on Chroma1HD (5090 32gb), all of them have come out fantastic so far.
Any idea why using a Lora in the Chroma Base is not working?it doesnt follow it, which base do i need to train it? ive tried flux and chroma
I've created Chroma loras and they worked fine.
Is it recommended to use the flash heun lora with this, since it's already distilled?
Ah sorry, it already has the flash lora integrated into it. I probably should have mentioned that
PatchModelAddDownscale (Kohya Deep Shrink) node plays nice with this model. Base generation at 1920x1080 makes a huge difference for I2V keyframes. Obeying the magic multiple rules of course. In case anyone wonders why closeup I2V works great and everything else sucks. Zoom in to find devil. I don't know why everyone is crying about RAM though... guy makes a GGUF for you and you cry about that too... just go buy some RAM, it's so cheap you whiners! Next you'll be complaining about some other readily available, dirt cheap commodity, like HDDs...
When was the last time you bought RAM?
@shoringu466聽I think perhaps you failed to notice the sarcasm I applied with a trowel.
Fantastic work
does anyone have any tips to get consistent lighting across a batch of photos with this model? one gen will be bright and easy to look at the next gen will be dark and moody with heavy shadows, nothing i prompt seems to be able to make it stay consistent, also there seems to be a random camera flash effect that randomly shows up on some images without me prompting for it, and i cant prompt it out to keep it from happening,.
You're absolutely right but those issues are a side effect of the instability of the base Chroma model. I think I've gotten as much of that instability out as I could without hindering the creativity from the base model. It's a work in progress.
This is a labor of love. Chroma is such a sensitive model when prompting for realism. It's like panning for gold when using the base model.
Thanks very much for the feedback. The Chroma base model is awesome but it's very unstable so this model is my attempt to try to tame it a bit.
use "this is a real photograph" in your prompt.
why does every single face look the exact same, even from different ethnicity. has the most similar looking faces of any model ive used
Chroma is based on Flux Schnell so there's a default facial type. But you can get more diverse facial details through your prompt. That said, the goal with future versions is to provide more diversity.
@GBRX聽when i use just base chroma the faces look different, even when i use flux schnell its more different. something about the flash lora or about this model, but also in older version 1.2 the faces used to have more variety now its less
@GBRX聽it's definitely more homogenous than the raw chroma, though, as the previous reply mentioned, i'm not sure if it's because of the flash lora or your finetune. If you could manage to increase the variety a bit more, than this model would probably be among my main recommendations.
@m4rbleye聽Increasing diversity is always on the todo list. It's a very, very long todo list however.
It seems to me that version 3.0 has a much greater impact on faces than version 2.0. I use character LoRa's trained on chroma-base model, and with 2.0. i get perfect resemblance. Using version 3.0 the resemblance is gone. BTW great WORK!:)
i would like to keep character LoRa resemblence using Gonzalomo Chroma v30. Is it possible to train LoRa directly on this checpoint in AI-Toolkit? Or what model would you recommend to train character LoRa on for this checkpoint? Chroma1-HD?
I can no longer type in the prompt node. WTF happened? Is it from updating comfyui? I tried updating all and updating the custom nodes. See screenshot https://ibb.co/WN8hcxph
Sorry, I have no idea what that issue could be. It's a primitive node. Try deleting and recreating it
Since there is no input and the output is a string, it looks like you're connected to the wrong node, not Clip Text Encode. Can you create a new Clip Text Encode node?
This is works as described. No errors, bugs, artifacts, farts, glitches ,etc. recommended. Also, workflow suggested in description absolutely magnificent! Thx a lot, dear friend
I really appreciate you releasing an nf4 version.
This has the best output of any Flux-related model I've tried. It feels like it has some Illustrious/Pony energy.
still the only chroma model that works for me that doesnt lose detail, or that doesnt take 5 minutes to render.
Thanks for the great feedback!
I have tested nearly all of the Chroma models currently available on Civitai. In terms of both speed and quality, nothing outperforms GonzaLomo V3. If you are just getting started with Chroma, you can simply use this model鈥攖here is no need to look for a better alternative. If you prefer a 2D style or are looking for a different aesthetic or artistic direction, then it may be worth exploring other models at that point.
Hey, thanks so much for the kind words. Chroma is one of the most versatile models and I'm happy that people think I did it justice with my version. However, even though you gave me that glowing review I have to mildly disagree about its artistic capabilities. Chroma actually has a bias towards artistic depictions of a prompt and I had to do a good deal of training to try to bend that bias more towards photorealism. But that artistic capability is still in there! You just have to prompt for it. See this post for a few examples - https://civitai.red/posts/28289758.
@GBRX聽I don鈥檛 think this model is lacking when it comes to 2D or artistic styles. What I meant was that for people who are new to Chroma, it makes sense to start with this model first, and then explore other models only if they feel the need for further experimentation. Since I鈥檓 not a native English speaker, I rely on translation, so some nuances may have been unintentionally distorted. I appreciate your understanding.
@DoroArmy聽No problem, I certainly agree that people should try other models. I would never claim that mine are the best.
Not trying to start model wars, just out of curiosity, have you tried the recently released UnGloryHail model? How would you compare it to GonzaLomo Chroma?
@mmdd2543聽No I haven't tried it but thanks for the recommendation! btw, your link was wrong. This is the correct link to UnGloryHail, https://civitai.red/models/2580292/ungloryhail
@mmdd2543聽Since I have limited SSD space, I usually test models and only keep the ones I consider the best. At the moment, GonzaLomo is the only model I have stored. I did try the model you mentioned, but it didn鈥檛 leave much of an impression on me. The one I鈥檓 currently paying more attention to is the next improved version of Photonic Fusion.
I can't make the FP16/GGUF model work. I get this error:
"ERROR: Could not detect model type of: F:\AI Models\diffusion_models\Chroma\gonzalomoChroma_v30.gguf"I have updated the GGUF nodes from City69 to the latest version.
It will be difficult to answer because everyone's PC environment is different. It would be helpful if you could provide more specific information. I usually use kjnodes' GGUFLoaderKJ. The problem might be resolved by changing the custom node.
@DoroArmy Thanks for the quick reply! I can run any other GGUF model, but not this one. I tried with KJ node's GGUF loader, but it too gives an error:
ERROR: Could not detect model type of: F:\AI Models\diffusion_models\Chroma\gonzalomoChroma_v30.gguf
Do you think its possible that the model got corrupted during download? Let me try downloading again.
@DoroArmy聽Hm... I no longer see the FP16 GGUF model of v3.0 in the download section. Did the author remove it? Can you see it on your side?
@mmdd2543聽Yes, I removed it. Something happened to it when I uploaded it civit. It was supposed to be a BF16 unet, not a gguf and civit wasn't allowing me to update it. I'll have to reupload the file properly when I can.
@mmdd2543聽I reuploaded the proper BF16 version
@GBRX聽Thanks! It works, but strangely only in the latest versions of ComfyUI such as 0.20.1. If I run it in ComfyUI 0.12.3 from 3 months ago it would give me this error:
RuntimeError: ERROR: Could not detect model type of: F:\AI Models\diffusion_models\Chroma\gonzalomoChroma_v30.safetensors
Maybe I'll just have to move to the latest version to avoid the error.
@mmdd2543聽I also held off on updating ComfyUI due to issues with the subgraph feature, but I ended up having no choice but to update it. In this field, things become outdated after just six months.
@mmdd2543聽Yeah, I used to hold off on updating comfy unless I absolutely had to but a lot of new model architectures and other features require recent versions.
@mmdd2543聽Having the same issue. Tried provided refiner workflow and still get the same error.
It doesn't work:
AttributeError: 'NoneType' object has no attribute 'unet_key_prefix'
ValueError: Failed to recognize model type!
Any suggestions?
Which specific model version and workflow are you using?
16Gb, WebuiForge. But 8Gb works perfectly there.
@Krutofff752 Sorry, I don't use forge so it's hard for me to say what the issue could be.
Getting the same error on the bf16 model - AttributeError: 'NoneType' object has no attribute 'get_model_object'
@carterj099聽I just tried the BF16 version in ComfyUI and have no issue running it as the unet.
still the best. i just downloaded the v3 gguf (~17gb) back in April. I see there is a bf16. is there an improvement with this version?
BF16 is usually going to give better results than a Q4 gguf but many times the difference is barely noticeable.
With GonzaLomo Chroma bf16, you've done an incredibly precise and professional job. I feel like i am Photographer from a major international agency... Thank you so much!
Did you use the included workflow or a different one? I keep getting an error when trying to run bf16
@carterj099聽What's the error?
Awesome work, definitely my favorite Chroma model! I wonder though which Lora (Flash?) was baked in in order to achieve results in just 10 steps.
Thanks! I used chroma-flash-heun_r128-fp32
Awesome work, definitely my favorite Chroma model! I wonder though which Lora (Flash?) was baked in in order to achieve results in just 10 steps.
did someone try training lora using this checkpoint as a base model? i dont have resources for lora training on original chroma so if someone share his experience using gonzalomo as a base it would be nice
Personally, I always use the base Chroma model for lora training. I think it creates the most flexible lora that can be used with any model. That said, I haven't tried training against my model so I can't say that it doesn't work.
@GBRX聽and btw amazing model! thank you so much! im using v3.0
1.Can you give recommendations for t2i generations. I use forge-neo and it doesnt have preset for chroma so i use flux preset default settings. But i think there must be some other adjustments for better results. currently default settings i use are: Euler, Beta, steps 20, distilled cfg 3.5, cfg 1. Ive experimented a little but to much combinations exists to try. Can you share your recommendations? (i understand probably my question is irrelevant cause most people here use Comfy i see xD)
2.Following your advice ive trained lora in onetrainer - adafactor cosine rank 64 alpha 32 lr 0.0005. can you advice what training adjustments or generation adjustments should i use in order not to copy lora hero face for other participants?
@atrib946聽I put the recommended settings in the version description. Your steps seem too high. As far as OneTrainer, I normally just use the default settings for the most part and those work well.
Amazing model. truly fantastic. Can you, please, give some advie about prompting ? thanks you !
Chroma is a derivative of Flux so it works best with natural language sentences rather than sdxl-style tags. That said I've no problems getting good results from tags as well.
@GBRX聽yeah both are good, i wanna know what is better on you model. I'm getting used to now . TY !
One thing I can say with certainty: this is a magnificent model! The execution of the prompt is superb! Whatever you imagine will definitely be in the image. By the way, it perfectly reproduces the text in the photo.