Hi im No_Thing, this is my first Workflow ever :)
Philosophy Engineered for clarity and quality. Designed with strict UI logic and a clean left-to-right architecture, eliminating scattered nodes and unnecessary visual clutter. A streamlined tool built for immediate output.
I made this workflow as a project to learn more about Comfy UI, im used to node based workflows with Blender and Davinci resolve so i know how the noodle wiggles. Any feedback is welcome, especially how to make the workflow better, we are a community, help me to help you to help me :)
Krea 2 Ultra-Detail & Dual-Sampler Pipeline
Turn your wildest SFW/NSFW dreams into absolute high-res perfection. This ultra-streamlined ComfyUI workflow pairs dual-stage Krea 2 sampling with 2K DiT upscaling for unmatched skin texture, razor-sharp detail, and zero censor filters.
* FLUX UNET: [snofsSexNudesAndOtherFunStuff v1.4 Distilled](https://civarchive.com/models/2416142/snofs-sex-nudes-and-other-fun-stuff-merged?modelVersionId=2985440)
* CLIP: [Huihui Qwen3 8B Abliterated v2 FP8](https://huggingface.co/YifeiDevs/Huihui-Qwen3-8B-abliterated-v2-FP8-Comfy)
* VAE: [Flux2 VAE](https://huggingface.co/Comfy-Org/flux2-dev)
* Upscale LoRA: [Flux2 Klein Image Restore V1](https://civarchive.com/models/2474084)
* Diffusion Model: [Lustify NSFW v1.0 Krea2](https://civarchive.com/models/573152/lustify-nsfw-checkpoint)
* CLIP / Text Encoder: [Qwen3 VL 4B Heretic](https://huggingface.co/Kutches/Kr3a)
* VAE: [Qwen Image VAE](https://civarchive.com/models/1912333/qwen-image-vae)
*VAE [Kijai/MiniMax-H3-TAE] Preview Window, works with minimax h3 https://huggingface.co/Kijai/MiniMax-H3-TAE/blob/main/vae_approx/taeh3.safetensors
ComfyUI-Crystools-MonitorOnly:
github.com/BobRandomNumber/ComfyUI-Crystools-MonitorOnlyComfyUI-Crystools:
github.com/crystian/ComfyUI-CrystoolsComfyUI-Pixaroma:
github.com/pixaroma/ComfyUI-Pixaromargthree-comfy:
github.com/rgthree/rgthree-comfyComfyUI-KJNodes:
github.com/kijai/ComfyUI-KJNodesComfyUI_essentials:
github.com/cubiq/ComfyUI_essentialsComfyUI-SeedVR2_VideoUpscaler:
github.com/numz/ComfyUI-SeedVR2_VideoUpscalerComfyUI-Krea2T-Enhancer:
github.com/capitan01R/ComfyUI-Krea2T-EnhancerComfyUI-Image-Saver:
github.com/alexopus/ComfyUI-Image-Savercomfyui-vrgamedevgirl:
github.com/vrgamegirl19/comfyui-vrgamedevgirl
Description
Version 1.0 Has toggle for upscaling.
FAQ
Comments (8)
Hey, I tried your workflow but the Flux Upscale output looks nothing like the input — it’s like it’s generating a brand new image instead of actually upscaling it. Also, the main part uses Krea2 and then runs it through Flux2 for refinement, which I don’t really get. And I keep running into errors too. Any chance you could help me out? Appreciate it!
BTW thx for ur work~~~~~~~~~~~
♥
Krea2 expects conditioning with 12x2560=30720 features (a 12-layer Qwen3-VL stack) but got 12288. Load the text encoder with CLIPLoader type 'krea2'.
you most likely are missing the lora for flux, image restore lora, its linked in the note within the workflow,
the problem with the clip is most likely just wrong clip if it does not match the conditioning, 99% of the time its the wrong model you have loaded,
flux refinement increases sharpness of fine detail, detail like individual hairs being sharp and in focus, even skin pores have more detail and it all comes to having the detail and data for video models to work with, im saying this out of experience of using video models, not that i actually know any technical things, its just "trust me bro", i hope this helps
The skin texture and details are amazing, genius!
thanks for the tip :)
Thank you for sharing this workflow, I do however find it quite difficult to use upscale. I've notice a lot of your images doesn't use upscale provided in this workflow, It does Improve sharpness overall of the image and gives a boost of quality but also quite a downside as it gives this weird lines of texture.
Seedvr creates weird lines with the sharpest model, i guess it can be bypassed by changing to a softer model and not using downscaling between flux and seedvr, or then just removing seedvr completely.
What i would do if id have the tools, is create a tile controlnet upscale with sliced tiles, 1500 pixels per slice then upscale them individually with a small florence model to describe the sliced image, but unfortunately 9b klein nor krea 2 has Tile CN, the tiles never align and if they do its a miracle, its something ive been trying to figure out but dont know how, its unfortunate.