Chroma by Lodestone, quantized using GGUF by silveroxides and Clybius.
Original description:
Chroma is a 8.9B parameter model based on FLUX.1-schnell (technical report coming soon!). It’s fully Apache 2.0 licensed, ensuring that anyone can use, modify, and build on top of it—no corporate gatekeeping.
The model is still training right now, and I’d love to hear your thoughts! Your input and feedback are really appreciated.
What Chroma Aims to Do
Training on a 5M dataset, curated from 20M samples including anime, furry, artistic stuff, and photos.
Fully uncensored, reintroducing missing anatomical concepts.
Built as a reliable open-source option for those who need it.
See the Progress
Hugging Face Repo: https://huggingface.co/lodestones/Chroma
Hugging Face Debug Repo: https://huggingface.co/lodestones/chroma-debug-development-only
Live WandB Training Logs: https://wandb.ai/lodestone-rock/optimal%20transport%20unlocked
ComfyUI Inference node [WIP]: https://github.com/lodestone-rock/flux-mod
ComfyUI workflow: https://huggingface.co/lodestones/Chroma/resolve/main/simple_workflow.json
Training code!: https://github.com/lodestone-rock/flow
Support Open-Source AI
The current pretraining run has already used 5000+ H100 hours, and keeping this going long-term is expensive.
If you believe in accessible, community-driven AI, any support would be greatly appreciated.
👉 [https://ko-fi.com/lodestonerock/goal?g=1] — Every bit helps!
ETH: 0x679C0C419E949d8f3515a255cE675A1c4D92A3d7
their discord: discord.gg/SQVcWVbqKx
Description
FAQ
Comments (42)
can i use it with forge?
Sadly no
@levzzz so, only comfy? Maybe invoke will handle it?
@compgamer1337267 There is a patch for Forge currently, you need to move some files manually. Maybe they'll merge it later
https://github.com/croquelois/forgeChroma
Yes, just update your forge and d/l models
Can you help, I'm trying to use gguf with your workflow, but I get an error:
"Could not find ComfyUI-GGUF node: GGUF support requires ComfyUI-GGUF"
and if using Unet Loader (GGUF): "KSampler
'attention_mask_img_shape'"
Huh, weird. You do have to use the chroma model node due to the architectural differences, maybe you should try reinstalling comfy gguf
Don't use the classic gguf node. Use the chroma diffusion node instead. Bug fixed for me. Works fine.👍🏻
I'm having the same issue I have ComfyUI-GGUF installed but when I try to load a gguf model with the Chroma Unified Model Loader I get this error
2025-04-15T12:28:34.527841 - Traceback (most recent call last):
File "E:\stability\Data\Packages\ComfyUI\execution.py", line 345, in execute
output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
File "E:\stability\Data\Packages\ComfyUI\execution.py", line 220, in get_output_data
return_values = mapnode_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
File "E:\stability\Data\Packages\ComfyUI\execution.py", line 192, in mapnode_over_list
process_inputs(input_dict, i)
File "E:\stability\Data\Packages\ComfyUI\execution.py", line 181, in process_inputs
results.append(getattr(obj, func)(**inputs))
File "E:\stability\Data\Packages\ComfyUI\custom_nodes\ComfyUI_FluxMod\flux_mod\nodes.py", line 72, in load_unet
flux_mod = load_flux_mod(
File "E:\stability\Data\Packages\ComfyUI\custom_nodes\ComfyUI_FluxMod\flux_mod\loader.py", line 137, in load_flux_mod
ensure_gguf()
File "E:\stability\Data\Packages\ComfyUI\custom_nodes\ComfyUI_FluxMod\flux_mod\loader.py", line 27, in ensure_gguf
raise RuntimeError("Could not find ComfyUI-GGUF node: GGUF support requires ComfyUI-GGUF")
RuntimeError: Could not find ComfyUI-GGUF node: GGUF support requires ComfyUI-GGUF
@RBD675 Open E:\stability\Data\Packages\ComfyUI\custom_nodes\ComfyUI_FluxMod\flux_mod\loader.py
replace the line number 25 with
gguf = sys.modules.get("E:\stability\Data\Packages\ComfyUI\custom_nodes\ComfyUI-GGUF")
@RBD675 Undo the changes I suggested. It seems the new update fixed it.
11th of March 2025: The node used by your workflow are not available in comfyui manager yet. Please get them listed and checked.
Yes, you have to install it manually. The link is in the description.
Just open GitHub copy command and paste in terminal.
Even my grandmom can do it.
nice, but very slow, are there any existing loras
Not that I know of. Most training software doesn't support it :(
@levzzz @nanunana LoRAs: https://huggingface.co/silveroxides/Chroma-LoRA-Experiments
It is partially support some Flux lora.
@2P2 Thanx, i will try some
With <lora:Hyper-Chroma-low-step-LoRA:1> and 10 steps, it's ~ 4 sec/it on a 3060 12GB. You can increase steps for better detail.
This is the best Flux model! I would say it's the Pony of Flux!
Still waiting for ponyflow v7~ The main version is based on auraflow, but they said there might be a flux schnell version too.
Interesting project, sounds like just what we need at this point! I have briefly looked for a report or guide about the prompting for this model, haven't found anything in the HF repo.
One very important aspect to make sense of a model is to know about how the training data were captioned, so we can "speak the language" of the model.
Are you planning to release a prompt guide or dataset caption report?
Ah, well, I'm not the person who created the model, merely reuploading it here. I think the captioning is a mix of booru tags and JoyCaption, I've had success with both natural language and tags. It doesn't know characters very well as of right now, but might get them correct if you explicitly describe the clothes and appearance.
@levzzz thanks for your reply and sharing your experiences, this is helpful. I will have another look at the available resources and see if I can find out more
Pro: It actually can write words on images.
Con: It generates ugly horrible trash if use long detailed prompt with human character.
After multiple iterations I managed to get non horrible female which looks worse than SD1.5
Also it will work only with special FluxMod nodes, it will not work with usual nodes as normal Flux.
yeah, it's a little undertrained as of now, so artifacts are inevitable. think of illustrious 0.1
I told this in a review I left to the original Chroma.
When this model is uploaded to an online generation service, such as TensorArt, the service accepts it, but using it only you get useless black squares/rectangles independent of the prompt or the parameters that are entered.
It will be worth applauding if you manage to repair that crass mistake Lodestone made in creating a model that according to some is the Pony Diffusion V6 XL of Flux-based models.
oh well, seems like it's either a service issue (not loading the model with the right workflow) or an fp8 issue, because the fp8 version by clybius has the same issue
It's a different model architecture. The reason your website isn't working is because it wasn't set up to handle Chroma properly. It requires a custom comfy workflow. You can read about it on the huggingface page to see how it's different from flux. If you want to use the model on tensor, then you should try complaining to Tensor and/or comfy to support it. I find it funny that you whine to op about fixing some sort of mistake when they're just reuploading the quantized versions. They even give credit and link to where they're reuploading from.
Ok, @Mobbun, I will drop my hopes of getting a flux based model equivalent to Pony Diffusion V6 XL in another model.
Good model. But if you want the art style to be something better than scribbles. Using Pony/SDXL style tags like "masterpiece", "best quality", "digital illustration", "awesome quality" will make it look really good. Though the scribbles are kinda nice because not many models can do bad quality as well.
there are high quality examples too, make sure to check them out.
Nice work, this follows prompts pretty nicely. I personally go for photographic style mostly, I can see this defaults to anime/digital style and does a good job at that. If the photo output could be further polished (less artifacting, finer details) it would be amazing. But I understand if that is not your goal!
Try aesthetic 11 if you want a quality tag. I think there are other aesthetic # tags, but I don't know what they do.
Can this run in ForgeUI?
Sadly, no.
It can if you patch forge, I have it working on mine https://github.com/croquelois/forgeChroma
@jobobby04 Thank you, I'll have a look!
what about ReForge?
@jobobby04 what's the vram and ram usage with 1k pics? which quant do you use?
This model gets better with every update❤️❤️❤️