Chroma Merge
This model is a custom merge designed to streamline the power of Chroma for everyday use. By combining Chroma HD with Cenkrechro V25 and GonzaloMoChroma, I’ve aimed to create a "best of all worlds" experience that retains the iconic character and deep knowledge of the base Chroma model while significantly improving generation speed and ease of use.
What’s inside:
Base: Chroma HD
Additions: Cenkrechro V25 & GonzaloMoChroma
Optimizations: Integrated LoRAs for improved anatomy and lighting consistency.
Why this merge?
In my opinion, Chroma is the best model currently available.. My goal is to bring more creators into the Chroma community by providing a versatile, "plug-and-play" version of this incredible architecture.
Recommended Settings:
Sampler: er_sde
Steps: 10-14
CFG Scale: 4-5 (with the png workflow)
I suggest the GNER-T5 clip
I hope you enjoy the results! If you like the model, please share your generations in the gallery below—I’d love to see what you create.
Credits
https://civarchive.com/models/1330309/chroma
https://civarchive.com/models/2032955?modelVersionId=2300930
https://civarchive.com/models/1836040/cenkrechro
https://civarchive.com/models/2182526/gonzalomo-chroma
https://civarchive.com/models/1662740?modelVersionId=2299345
Description
This is the best I can do for an fp8 version of the model.
FAQ
Comments (30)
Thanks for adding the FP8 version, even though you can just use the full model at fp8 weight fine, nice to have the smaller size. Great model btw
it's not working for me I get a empty output
I downloaded this clip https://huggingface.co/silveroxides/GNER-T5-xxl-encoder-only/blob/main/gner-t5-xxl-enc-float8_e4m3fn_scaled_stochastic.safetensors
i also tried other t5 clips that i have used with other chroma models and it didn't work either
I followed the workflow from the older version you shared
Have you tried the stock chroma workflow, it should work with any t5 clip.
@panorios yep, same issue, preview all black with any t5 clip i choose. I am.tryong the fp8 btw. Could you share your workflow for those?
i see in the terminal a big warning:
unet unexpected: ['img_in.comfy_quant', ...
and a ton of elements in it. Updating comfyui didn't help
Have you tried with the chroma hd model? If you cant use it it could be a problem with your comfy install. If you download the png's from my first model you can use them in comfy as workflows.
@panorios chroma hd is too large for my gpu, but gonzalomo chroma or uncanny photorealism chroma both work
@xhorxhi Try with the gonzalomo workflow. It is working for me.
But is this the right T5 clip or can you put a link in here so i can download the one young use
@warsound75 use the t5 that is working with any other chroma model you use. Or try the fp8 version. What hardware you use?
@panorios I using the model now, and working perfecly, Best model ever i.m.o. You think it the model makes it. I'm using a RTX5070 TI. Takes about 6 to 7 second to make a image on 1216x832 and 11 sec to make one on 1840x1232. er_sde and beta57. Just amazing
I got really good results with this one and way less anatomy issues. I tried with deis_2m. Great work
It's really good, nice work. I suggest using an SDXL refiner at the end also
I have no idea how you did it but this checkpoint is witchcraft. Prompt adherence is amazing, lora usage is great, wooh. It just lack some advice with the best configuration to use.
Great job!
Yeah, this thing is nuts
Contrary to the recommended settings, for anime style images I would suggest running this at CFG 3.5 (dont go above 4.0) and below with the CFG Norm node in ComfyUI. It improves the detail on all image textures. Trust me you'll love it.
Also just to let the creator know, the workflow you mentioned in the description is not present in the showcase images. ComfyUI auto converts the images to .webp, which removes any attached metadata.
Hi, thank you for your feedback. I suggest right clicking on the image - open link in new tab-save as png and you have the workflow. In this workflow you can go as high as 5 CFG. On a standard workflow it depends on your scheduler and the style you want.
POV: You're casually surfing through Civit and accidentally discover THE FUCKING BEST CHECKPOINT THAT HAS EVER SEEN THE LIGHT OF DAY! WHAT A BANGER! Well hidden behind a really crappy name. :P xD
Allow me to ask: What are your sampler/scheduler recommendations? Below, you only write er_sde, but in your workflow (which doesn't seem to be your real workflow), you have Euler/Simple set. What settings did you use to create the sample images? Steps, sampler, scheduler, CFG...
Happy you liked it.
The fp8 version is using Beta and er_sde. (I used krita, so just the standard wf). you can use the gonzalomo wf too https://civitai.com/models/2182676
The fp16 images have the workflow for comfy that is using the custom noise. Use the png image for wf in comfy.
In general chroma is an amazing beast, there are no (best settings), you can use different settings for different styles or moods.
@panorios Like it a lot! Thank you for your reply. I don't really like Gonzalomo's workflows; for me, they somehow deliver crappy results and there's always something wrong. I can highly recommend Tenofas' workflows: https://github.com/Tenofas
I prefer to use Res_Multistep—I'm always torn when it comes to schedulers. But I mostly use beta or beta57.
This seems to be the best Chroma Model out there. Discovered by accident. Converted to Q8 - using with SDXL refiner (Gonzalomo workflow).
Thanks for this model.
Great, happy you liked it.
Hey would you share the Q8 version you generated and the workflow link? Thanks!
@mark_sausage439 Hi, my internet connection is useless for uploading. You can easily convert using this guide I made for qwen. It works for Zimage as well. Follow all the steps to the letter.
https://drive.google.com/file/d/1QdmtXijMROZC8nTpKPJGQ_Xlt8OIJK6U/view?usp=sharing
Here's the full workflow - https://drive.google.com/file/d/1722UbcbdTImAzgE8kY-1mrlXCKjw8GhZ/view?usp=sharing
Note- use the gonzalomo sdxl model in the workflow. That seems to be the best one. you can use sdxl loras as you like.
If you change refiner denoise to 0, the refiner will not change the face or background but the sam models will change breasts, nipples, face etc as you want. Values like 0.2, 0.3 will change the face. Experiment with it and find what you like.
@panorios Are you planning any new version of this?
@JobDoe364 Yes, hopefully in a month I will release V2
Thank you for making this model. But I have an issue though. The 'er_sde' give me the best result but the image is grayed out. I'm not using CFG Scale (just keep it at 1) and don't have any negative prompt involved. Would this be the problem? I have not tried your workflow since I hate install too many custom nodes. Thanks for your time.
I guess you could go at least at 2 CFG. There is no need for negative prompt if you have a good description. are you using any Loras?