Based on my Centerfold Flux 5, trying to merge in Chroma and Krea.
Diffuser version 2.3 https://huggingface.co/Tiwaz/CenKreChro_V2_3
For version 2.3 I will add the quantizations under the same "tab".
Big thanks to Spooknik for making the SVDQ versions https://huggingface.co/spooknik/CenKreChro-SVDQ
Q4 KM works fine in Comfy UI but not in Forge for example.
If you need the diffuser version: https://huggingface.co/Tiwaz/CenKreChro (not 100% tested yet).
You can try using it as low as 12 steps even if I prefer 30.
Description
FAQ
Comments (20)
Hi! Do you know how long it take to convert an FP16 Flux Dev checkpoint to NF4 SVD quant? I have a custom model that I would like to convert to NF4, but I think I read some troubling comments saying that you need several H100 videocards to convert a Flux Dev model in a reasonable amount of time. Is this true? I have an RTX 5090 32GB and 96 GB system RAM. Do you think this would be enough and how long would it take with this setup do you think? Thanks in advance and for the wonderful models!
Hello, I made the SVDQuants for this model. You need at least 48gb of VRAM and even then you need to when with low dataset sizes which makes the model calibration less accurate. I use an RTX Pro 6000 (96gb VRAM). It takes around 18-20 hours per quant.
I have a WiP guide here - github.com/spooknik/deepcompressor-guide
error loading this model in ComfyUI:
CheckpointLoaderNF4
ERROR: Could not detect model type of: K:\ComfyUI_windows_portable\ComfyUI\models\checkpoints\flux\cenkrechro_v23SVDQFP4.safetensors
yes, you need the nunchaku node in comfyui.
@TiwazM I have this node, but for some reason it doesn't see the model in the folder:
https://drive.google.com/file/d/1lO60DzhD2aKF7JR22PuJWdQBqks9hTmc/view?usp=sharing
https://drive.google.com/file/d/1SvbG6w5el4XxDSJMvYE3Y1uGbqrntaww/view?usp=sharing
@modzhahead158 aah I see Q4KM is the gguf node. only the v1 and v2.3 svdq fp4 work with nunchaku if you have a 50 series card, if you have another you need to get the int from https://huggingface.co/spooknik/CenKreChro-SVDQ
@modzhahead158 search for gguf and it is the one by City96
@TiwazM thankyou, I downloaded svdq-int4_r32-CenKreChro-V2.3.safetensors from huggingface and now everything works fine on my rtx4070 12gb:
https://drive.google.com/file/d/14ZEDa8qdYXK4FUnRarjbHwCjzsM657GF/view?usp=drive_link
Excellent model, able to create high quality images as displayed. Thank you bro.
Fantastic model! I much prefer version v1, it produces more realistic people and gorgeous effects, also more flexible. The newer ones are a step back for me.
they indeed are a step back.
truly a hyperrealistic model. mainly cenkrochro variants, cenkrochroflash... but there is a problem in regards to subjects skin generation, subjects often tend to have these black spots, like blackheads covering the exposed parts of their skin/body, any idea how to get rid of these? sometimes it is soo bad the subjects look like someone took a s**t in front of turned on ventilator and subject was standing behind the ventilator. the environment and subject details otherwise are top notch, so this irks me a bit lol. negative prompting seems to do nothing to solve this. maybe some magical scheduler/sampler combo? i tried many and cant get rid of the blackheads entirely.
do we have gguf versions or e5m2 of the latest 2.3 ? I tried converting like I did 1.0 but it is giving errors with the new model.
If you can use them the nunchaku versions are better quality and smaller file size.
Let me try tomorrow to create both.
As I'm in an amd card, I would also like to have both.
I was able to GGUF v2.3, it is available here - https://huggingface.co/BigDannyPt/CenKreChro-GGUF
@BigDannyPt from full fp16 model, yes ?
@xpnrt yes, GGUF always need FP16, even if that means to multiply the weights through a script ( I'm still not sure how this impacts quality but I also done this with the following one if you want to test, but is wan 2.2 https://huggingface.co/BigDannyPt/WAN-2.2-SmoothMix-FP16 )
Since FP16 is available, I started from it
@BigDannyPt I know it requires fp16 as base but for some models people do it from fp8 since fp16 isn't always available, wan is more forgiving at lower quants. so from fp8 q4 or q5 should be usable.
@xpnrt yeah, I have some other models that I convert like that, unfortunately, not always the OC does the merge in FP16, that's why I have a script to try to multiply the weight on FP8, but I don't know how that impact the quality.
But this one, I can guarantee that It was from the FP16