Photonic Fusion Chroma INT8
I have taken the debloated version of Photonic Fusion Chroma and converted it into a permanent INT8 convrot format using ComfyUI.
Requirements & Compatibility:
This checkpoint requires the ComfyUI-INT8-Fast custom node extension or the updated, native ComfyUI INT8 loaders.
It is optimized specifically to speed up generation times on modern NVIDIA GPUs (especially RTX 30, 40, and 50-series cards) by utilizing native 8-bit tensor cores.
Generation Recommendations:
Steps: You can go as low as 15 steps, but I find 20 steps to be the absolute sweet spot for quality and speed.
Chroma checkpoints are incredibly detailed but notoriously heavy and slow to run. Baking the quantization directly into this checkpoint significantly speeds up inference times while dramatically reducing your VRAM overhead and system requirements! If you enjoy, please consider buying me a beverage - Coffee and Beer
Photonic Fusion Chroma - Debloated
Introducing the refined, BF16 version of the original model: Photonic Fusion Chroma - Debloated. I’ve taken the original 22GB beast and stripped away all the unnecessary "bloat" to bring you a proper 16+GB version that’s easier on your storage and VRAM, I’ve also tucked in a minor fix. Honestly, this checkpoint has been on a strict regimen of Ozempic to slim down to this fighting weight, but don't let the smaller size fool you; the Photonic structure and high-fidelity detail are punchier than ever. I recommend you delete V1 and save your disk space and some extra VRAM and use this version, this is the exact same model just slimmed down and better.
Huge thank you to VioletCrow for the Q4 GGUF version of my checkpoint, many in this community will be very grateful for your efforts to be able to use this model on far less VRAM now!!!
UPDATE 04/26/2026 and now the Q8 GGUF version as well!
Welcome to Photonic Fusion Chroma!
Man Chroma is a difficult beast to tame, after two failed attempts at a finetune the 3rd time was a charm. I thought I was going hang up my hat with my SDXL series but I am still around tinkering and I always wanted to add my own coat of paint to Chroma.
As always with my checkpoints it excels at male anatomy but the base Chroma model already did amazingly well at this I just trained in far more variety. I feel that anything base Chroma could do can still be done here but with more variety and with improved skin texture and more of that moodiness vibe I enjoy. And great news if you are not into gay males it still does women very well as well as all general types of art, fantasy art and animation styles. I hope you enjoy! I have no plans at the moment to create any smaller quantized or GGUF versions sorry.
I’ve spent weeks grinding on my own custom finetunes for Photonic Fusion, but it wasn't quite hitting the exact vibe I wanted on its own. To get that specific lighting/mood, I merged my weights with dawncreates Chroma. So yes, I merged my custom finetune with this other amazing model to get the balanced look I wanted and I hope you enjoy the final output!
ComfyUI I recommend Res_2s + Beta57 (but almost as good and even slightly faster euler + beta)
Steps 25-40 (I usually go 28-30)
CFG: 3-6 (I usually go with 4)
Resolution: up to you it does well with all common aspect ratios
Upscaling: lanczos 1.5x-2x, 4x-ClearRealityV1, 18-20 Steps, CFG 3-4, 0.27-0.33 Denoising. Optional: SeedVR 3rd final upscale.
Text Encoder - https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/t5xxl_fp8_e4m3fn_scaled.safetensors
If you find a better combination of settings ass always I ask that you please share with the community in the comments below!
For some of the sample images I have used some amazing prompts found right here on CivitAI, I wanted to test some amazing prompts I have seen for Z-Image, Flux and other Chroma models to see how well they performed on mine. So thank you to those that share their prompts!
Lastly, in this version the model does struggle some with sexual positions. If I make a second version I hope to address this.
Thank you Lodestone - https://civarchive.com/models/1330309/chroma
Thank you dawncreates - https://civarchive.com/models/2086389/uncanny-photorealism-chroma
Also I don't believe in locking my models that are available on TensorHub or CivitAI behind a paywall like some do but I gladly accept donated --> Coffee and Beer <-- your donations help me to keep motivated to continue making new checkpoints and LoRA's for all our enjoyment! To those that are generous enough to throw me a few dollars I thank you! <3
Description
The original model with a minor fix and stripped of all the bloat. The best full size version of Chroma checkpoint.
FAQ
Comments (23)
A Q8_0 version of the debloated chroma would be amazing 🙌
I know someone who may be able to help with that. I have a GGUF friend, I'll see if I can call on a favour but I make no promises hehe.
Man, it‘s the easiest thing in the world to make a GGUF quant, or other quants. How about you just teach yourself that instead of always bugging the creators with your special requests. Just an idea… ;)
@CupRunethOverWithLSD Unnecessary griping. If someone doesn't want to do a request they don't have to.
@ss9999 Exactly, and it wasn't even a request. OP can do whatever he wants, a quantized version is just accessible to way more people, with way more download and success. Anyway, that guy is high af on Lsd
How is the debloated version vs. the gguf version?
What is the "bloat" that you removed? Personally, I'd really like a Chroma checkpoint that has no anime, illustration, etc. Only realism. The non-realism is the bloat for me. However, I have read that having it in there is supposed to give the model better understanding of concepts so perhaps it's not bloat.
is it even possible to remove specific knowledge from a model like that
When I merged some of the other model mentioned in my about section with my trained model I accidentally backed in VAE and Text Encoder causing V1 to balloon in size plus had a flux clip which was completely unnecessary and unrelated to Chroma. So that is the bloat I refer to! :)
Also regarding the anime side of things I get it, it's not for everyone. I believe it would be impossible to cut that out as it's hard trained in the base model which all finetunes and merges stem from.
@StecFX That explains why the file was so big. Thanks for solving it!
@goeypants400 as fars i know quantisation does remove specific knowledge, i dont know if intently or not, yet to learn more on model training myself
@snap2887 No matter how many times someone puts the word "surgically" in their description of a quantization, throwing out data means data is lost. Normally that lost data has value because these surgeries are on finished products. Sometimes it's unavoidable to throw away data because the hardware one is using is too weak to handle the full model and sometimes it's arguably not going to lead to significant degradation (the main argument for ditching FP32 in favor of BF16 and FP16). Going below 16 bits isn't surgical in my view; it's drastic. However, there is disagreement about that. I remember running SD 1.3–1.5 in FP32 because FP16 was less reliable, more likely to be bland. That was despite endless claims that FP16 is indistinguishable from FP32. SD 1.x needed every shred of data because it had such horrible handling of anatomy. The larger the models become I think the more they can handle compression (versus a tiny model like SD 1.x) but the result will still be quality loss. 8 bits is severely compressed when the original model is FP32. But a model like Wan running in FP32 requires too much hardware, given the Nvidia monopoly and now the RAM pricing. I have had some decent results with Q8 using Flux 1D and Chroma (back when I was stuck using a 4070) so even 8 bits seems like it can be usable for some kinds of generations with some models. I try to stick with 16 bits, though, when possible and do text and VAE generally in FP32. Allegedly, FP8+ (which has some things still at 16 bits) is supposed to be better than Q8.
the debloated version is great. faster than the regular fp8 version in most instances. better consistent generation results so far as well. TY!!
LOVE to hear this thank you! :)
Please if you can make a Quant version , then i am sold :D, cant afford the 16gb harddiskspace for a single image model.
Thanks for this!
Results are quite good in general. Congrats!
I have some constructive feedback too, though. It surprised me that, even being a male-focused model, it has lost Chroma's ability to render foreskin. Your previous SDXL models handled it well, so this was unexpected. May be I'm not prompting it correctly or something, but so far I haven't been able to get any images with foreskin from this model. I can just use a lora for that, so not a big deal... But I'm curious: Is that concept present on the dataset?
Thank you and I value the constructive feedback. I overtrained on cut cock not intentionally I had plenty of foreskin in my dataset but clearly not enough. For some reason the training really biased towards the circumcised. But you are correct it's not you it just refuses to do foreskin so a LoRA for this version is definitely needed!
Thank you for this chroma checkpoint ! Results are very interesting in portrait photography with complex light and adherence is outstanding ! Nice with artistic results with fantastic @StuckyFan loras
But It's also a very dangerous and powerful model... Becareful in using it
Glad you like it! :)
There is a ComfyUI custom node called "calcuis/gguf." By using the "tensor cutter" node within this extension, it is possible to convert BF16 models to FP8 and save them. I have tested this method and confirmed that it works correctly. Naturally, this results in a degradation of quality, but if you are short on VRAM, it is worth the trade-off. If you require an FP8 version, please refer to the following link: