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 model is just the UNET so you will need to supply the Flux VAE as well as the clip.
I recommend using the GonzaLomo Chroma Refiner Workflow for good results.
Recommended settings:
Steps: 8-12
Sampler: euler, dpmpp_2m
Scheduler: simple, beta
CFG: 1.0
FAQ
Comments (53)
GBRX this is an interesting model, but it appears to be a flux model.
This model is missing the distilled_guidance_layer(s). Comfy checks for those layers to determine if the model is chroma. If comfy does not find those layers then the model is treated as flux.
Because of the channel differences between flux and chroma when images are generated using this model in comfy they have contrast issues, giving them a surreal look.
Zoot Allure, I get it!! Takes me back to French class :-D... So @GBRX will Flux Lora's work with this Checkpoint also? I was going to give it a spin and see what's what...
@zoot_allure855 Thanks for the insight. I'm still figuring out the best way to work with Chroma. It has the potential to make some amazing images.
But Chroma is Flux Schnel based.
@Learning2025 It is architecturally based around flux but there are critical differences.
You can read about them here:
https://huggingface.co/lodestones/Chroma1-HD
It is different enough that comfy treats it like a different model.
The source code for the chroma/flux model detection is here (I highlight the line that it starts on):
https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/model_detection.py#L251
And here is it's unique model implementation in comfy:
https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/ldm/chroma/model.py
@GBRX Your workflow definitely produces nice images! Please keep playing with chroma, looking forward to what else you create.
@zoot_allure855 You were loading in a Flux lora, which caused the layer errors. Certain blocks from Flux loras will not be loaded with Chroma, but most will. Almost all of my Flux loras work fine with Chroma, as long as I keep the strength low.
This model does not show any layer/block errors when used by itself.
@blastermaster123 That's the info I'm looking for... I've created a ton of Flux Loras so it would be cool if they can work here too!
@blastermaster123 No I wasn't. I was looking at the logs in comfy, I saw the model loads as flux in the logs; I thought that was strange because the model is labeled as chroma here. Chroma models loaded in comfy will literally say chroma in the logs. I then loaded the model in python to inspect the layers and saw the distilled_guidance_layer(s) are missing.
I confirm the issue, Forge recognizes the checkpoint as Flux Schnell type, not Chroma. On high CFG it gives Chroma-alike images, on low ones - Schnell-alike.
@zoot_allure855 Good to know, and I apologize for the incorrect assumption. I will say though, the genitals definitely look Chroma generated. Even with the best genital loras loaded with Flux, it still struggles. This model produces great looking genitals similar to Chroma.
@zoot_allure855 Thanks everyone for the excellent feedback! I'm hoping to make the next version need little to no SDXL refinement so I've obviously still got some learning to do.
@GBRX We're right there with you my friend!! Thanks for the hard work. How can we support/subscribe to help your efforts?
@GenoMachino Thanks! Just continuing to use my models is support enough. I've also got a Patreon link in my profile, if you're so inclined.
@GBRX I am in fact so inclined!! :-D
Okay, I've Gen'ed a few images... HOLY COW!! Wow... Super fun. Great work! Thank you! I'll play with a few Flux Lora's to test that out also...
does this model support flux loras or chroma ones?
I tested some of my favorite Flux loras and they work adequately. I was suprised they worked alot better than another fork of Flux - Krea. Krea did not like my loras at all, while this Chroma tune works with them almost flawlessly.
@mark_sausage439 It should support both pretty well.
Fantastic model. I've been pretty happy with Chroma 1-HD, but this is my go to for now.
@GBRX I converted your model to chroma.
You can download it here.
https://huggingface.co/zootkitty/gonzalomoChroma_converted/resolve/main/gonzalomoChroma_v12FP8UNET_converted.safetensors
You still need your workflow to make those pretty images but it does produce different results compared to your normal model.
Probably would have been better to convert the base model instead of the FP8 one.
I know I did not get permission to host your converted model on HF, so once you download it let me know and I will delete the repository from HF.
Based
Why the conversion? If I may ask..
Awesome, thanks @zoot_allure855! Would you mind briefly describing the conversion process you used? (I have a technical background so please be as technical as you need to be)
My main concern is making sure that the model can produce good quality images but if Comfy doesn't treat it as a Chroma model then that's obviously a huge problem so I want to make sure I do this correctly going forward.
@Learning2025 Well primarily because I thought it would be interesting to do. I had no clue what the outcome would be and in the end I liked the gens so I thought I share it.
@GBRX This morning I looked into what exactly the distilled_guidance_layer(s) do. I compared the tensors between between the standard HD model, uncannyPhotorealism, and SPARK preview. The tensors were identical, unless it was FP8 and even then in FP8 they were just adding scale weights. So, I took loadstones architecture comments and did exactly what he said he did; quote: "I replaced a 3.3B parameter timestep-encoding layer with a more efficient 250M parameter FFN, as the original was vastly oversized for its task."
TLDR: I removed flux's 3.3B parameter timestep-encoding layers and added 250M parameter FFN layers.
In practice it is actually a simple script to do this, I uploaded to the repository in HF:
https://huggingface.co/zootkitty/gonzalomoChroma_converted/blob/main/convert_model.py
@zoot_allure855 Thanks again! I see what the script is doing.
I did some A/B testing with your converted model and the image quality is definitely better. I prefer the composition of the original though so I have some tinkering to do for v2.0.
At least now I can put out a proper Chroma version.
@GBRX I'm glad I could help in some small way. Really looking forward to see what you do with your next iteration of this model.
@zoot_allure855 Thanks, I downloaded your mod, and used the workflow of @GBRX, the outputs don't look as good as the images on this post. Especially female vulvas, they are deformed. Do you mind sharing your WF?
@Learning2025 I am using GBRX workflow as it is. 10 steps, euler, beta. I made both his KSamplers random seeds.
But since GBRX is using no negative and cfg of 1 you'll need to be quite expli^cit with how'd you like your lady parts, you may even need multiple sentences.
For example a innie p would require.
She has a tight p...., with no labi^a. Her p.... is puffy with no clit....
Obviously use the correct words where those . and ^ are.
@Learning2025 And a final piece of advice, chroma was created using e621-captions and a image set that has first class support for f^urry content. So you need to describe what you're looking for in such a way that uses e621-captions and offsets the f^urry content which tends towards meaty lady parts.
Is there a gguf version?
Yes, I do plan on quantizing this as well. What level of quanitization would you like to see? I was planning to do Q6_K at least but may do Q4_K_M and Q8_0 as well if there's interest. I find Q6_K to be the sweet spot between size and quality.
@GBRX I always go for Q8 if available, but honestly, with the amazing work you are doing, I fully trust your judgment :D
@mark_sausage439 Thanks! I'll just do Q4, Q6 and Q8. It's pretty quick and easy to quantize. It'll also give me the ability to see which versions are most popular.
@GBRX Personally I prefer NF4 if it's possible.
@mphobbit I'll take a look at doing that as well.
goddamn you cooked
I know this is a stupid question, but I'm just learning this stuff.
I can't find the diffusion model or CLIP here or on Hugging Face.
Where do I find those?
Also... any particular LORA you've had good success with?
The CLIP is just the regular t5xxl_fp16 text encoder. You can use lower variants too, like FP8 or one of the quantized versions.
The diffusion model is the Chroma model on this page. It's just the UNET so you hook it up like any other Flux or Chroma UNET-only model.
@GBRX Thanks for taking the time to reply! I got it to work! Amazing!
Does the converted version work with Flux loras or with Chroma loras? Thanks!
Either version will work with both Flux and Chroma loras.
Great model and workflow. It unlocked many mixing possibilities, superb quality! I have made some merges so that it can work with negative prompt, not as fast, but decent. If I may suggest, you can use the er_sde sampler for sharper details. Oh and one more thing, your model is listed in the character section, I think more people can find it in the base model category. Excellent work. Thank you.
Thanks for the suggestions!
An absolutely wonderful checkpoint. Incredibly good understanding. I'm using Euler/beta with 10 steps and 1.0 cfg. It's fun to create. Thank you for it :)
Thanks! I'd be interested in hearing about the color issues you're seeing though. I do want to make the next version as good as possible.
@GBRX In my opinion, the skin tone is too brown/yellow, for example in these two pictures.
https://civitai.com/images/112613221
https://civitai.com/images/112613175
i testet many hours now and in my opinion, euler a/simple with 10-12 Steps and 1.0 cfg do a better job. But unfortunately, it did not solve the oversaturation problem.
@Gorean Ah, I didn't think to try Euler A yet. You might be right. It has a little better contrast and maybe composition than Euler.
God damn, dude! Amazing model. Good, fast and accurate. After all this time from sdxl, this is my new toy with Zurbo