Here's a workflow using Chroma to generate base image and SDXL to refine it.
It's the result of my long hours of practice on ComfyUI, applying tiled diffusion upscaling and detailers for various body parts. I tried to keep it as simple as possible, but it's a complex workflow that isn't necessarily easy for a new user to grasp.
Chroma is very ressource intensive so you can use quantized versions here which will be lighter for your graphics card model.
Here's a list of the different models needed to use it:
My SDXL merge model (you can use another one but this one is mine so it's better^^)
Detection models (Folder also present in the archive provided with the workflow, put them in /models/ultralytics/bbox)
Xinsir Tiled Controlnet (rename it as "Xinsir_controlnet-tile-sdxl-1.0.safetensors")
Lora for Nipple detailer (this one is mine too ^^)
All nodes used should be able to be downloaded via the Manager but I have not tried with a fresh install, I will provide details if necessary.
You can change the aspect ratio for images of various formats, and a landscape true/false button is used to calculate tile sizes for upscaling.
I recommend a base image resolution between 1.2 and 1.4 megapixels, depending on your preferences, the output image will be 2x larger.
You can also mute/bypass the Grain filter at the end of the chain if you wish.
Leave a review or tip me if you wish, it will be greatly appreciated, I give you here all my experience and my know-how on ComfyUi in order to obtain the best possible images with Chroma and SDXL.
I hope you have fun with it :)
Description
In this version the image goes through upscaling and detailers twice, after some tests, I think that this version is less subject to anatomical deformations than v1.0 which only had one step of upscaling and detailers.
FAQ
Comments (21)
Due to recent ComfyUi updates the Resharpen node no longer works, so I removed this node from the workflow as it was not essential, if you want to update ComfyUi remember to remove Resharpen from your list of custom nodes. Have fun ;)
i need image to image chroma is it possible ? make it now
Just add an image loader node and a vae encode to replace the latent input of the Chroma sampler, you can use a scale to megapixel node between this two last nodes to control the input resolution, feel free to experiment :)
graam558 I had luck doing this. Make sure you change the Sampler's denoise so it is lower than 1.0 (if you leave Denoise at 1, the new image will look completely different from your initial image.) Try 0.6 Denoise to start.
amazing work as usual man! A few newbie questions.
SOLVEDI ran the workflow and only generated the first and second images, are the upscales turned off by default ? Also the load loras as shown in the picture (https://imgur.com/a/i161J2J) are red, except the nipples one, and I put all the LORAS in the workflow in the correct folder Models/Chroma. Is there a selector besides the 4 ones on the "chroma" group?
Edit: found the solution, right click the Load Lora node and VIEW lora info, you should see the correct path the lora should me. Or right Click and Expand you can set a new path for the lora.
First image comes up, but then fails here:
Tiled Upscale 1
SamplerCustom
mat1 and mat2 shapes cannot be multiplied (32400x16 and 64x3072)
Something is wrong with the models, did you follow the instructions and download the correct models as indicated in the description?
@graam558 if the instructions said use ChromaHD and the WF said use chroma-unlocked-v48..., which is correct?
both, they are just different versions of Chroma
@graam558 Here are the models I downloaded and used:
Chroma model - Chroma1-HD.safetensors
Clip - converted-flan-t5-xxl-Q8_0.gguf
Flux VAE - flux_vae.safetensors
Chroma's Lora - the 4 Chroma LoRAs from HF
My SDXL merge model - graamXLJuicyMerge_v10.safetensors
Detection models - from your WF zip
DMD2 Loram - dmds_sdxl_4step_lora_fp16.safetensors
Xinsir Tiled Controlnet - diffusion_pytorch_model.safetensors from HF renamed to Xinsir_controlnet-tile-sdxl-1.0.safetensors
Lora for Face detailer - real-humans-PublicPrompts.safetensors
Lora for Eyes detailer - Super_Eye_Detailer_By_Stable_Yogi_SDPD0.safetensors
Lora for Hands detailer - 鲜创一派@Lady Hand_SDXL.safetensors renamed to Lady Hand_SDXL.safetensors
Lora for Nipple detailer - Detailed_nipples_xl.safetensors
I should note that have the same error with your SDXL HD Photos WF.
Update: I was previously testing a wan22 wf that turned on sage attention. I thought I read somewhere that you have to manually turn it off. So I did and then the Tiled Upscale worked. Weird. Now the wf fails at Face Detailer. I'm trying to debug that now, if you have any advice.
@ogbtjyqat450 Forgive me but I don't have much time to answer you, I have a lot of work at the moment and without more information I won't know how to help you, I hope you will succeed in running this workflow which is the result of long hours/days of experimentation and brings the best results that I have seen on SDXL
@graam558 No problem. I'm breaking it down into pieces zeroing in on what is going wrong.
noob questions: what's missing from chroma for you that makes it not worth using on its own?
I usually do multiple refinement passes which would take forever with Chroma, and tiled diffusion doesn't work with it.
@graam558 thanks for that- so i started using chroma and find it super hard to even get decent realism even with loras! tbh im not sure i should even adopt it to begin with. idk how they did it but it seems like hunyuan and wan do better and faster work with less garbled results! should i stick it out because im missing something?
For me, Chroma is much more creative and varied, with enormous possibilities, unlike Wan, which remains quite limited. Having tried all the models, Chroma remains my favorite. Qwen is really promising but still lacks Lora.
Here you'll find two Loras that I'm currently using to improve realism, in addition to those I linked in the WF description.
https://civitai.com/models/1908534?modelVersionId=2271596
https://civitai.com/models/2014953/reakaakas-enhancer-chroma-hd?modelVersionId=2280444
you can try this one too https://civitai.com/models/970862?modelVersionId=1944723 and most of the Flux loras works with Chroma :)
@graam558 ooh that's a big deal! thanks for telling me the flux trick, that makes a world of difference
@graam558thanks for these!