Rebels MrFlow — Staged Upscale Workflows for Krea-2 & ZIT
______________________________________________________________________________________________________
Two ready-to-run ComfyUI workflows (Krea-2 and Z-Image Turbo) built around MrFlow, a training-free staged sampling method: generate at low resolution, SR-upscale in pixel space, re-encode, then do a short 1-step refine at full res. Big speed/VRAM savings over sampling straight at target resolution — built and tested on an RTX 3070 (8GB VRAM).
Method credit: MrFlow by Xingyu-Zheng et al. (arXiv:2607.01642) — https://github.com/Xingyu-Zheng/MrFlow
Ported to Krea-2 and Z-Image Turbo by RealRebelAI.
1. INSTALL THE CUSTOM NODES (REQUIRED)
Grab the node pack from my GitHub:
https://github.com/RealRebelAI/Rebels_MrFlow
Clone it straight into your custom_nodes folder:
git clone https://github.com/RealRebelAI/Rebels_MrFlow
to "ComfyUI/custom_nodes/"
Or download the repo as a ZIP from the green "Code" button on GitHub and unzip the ComfyUI-Rebels-MrFlow folder into ComfyUI/custom_nodes/.
Restart ComfyUI. No pip installs, no extra dependencies.
This adds 5 nodes: ZIT Mr. Flow Preset, Krea-2 Mr. Flow Preset, Mr. Flow Upscale + Encode, ZIT Mr. Flow Refine, and Krea-2 Mr. Flow Refine.
The two workflow JSONs (ZIT_MrFlow_workflow.json and Krea2_MrFlow_workflow.json) are also in that same repo, and attached to this post — drag either straight into ComfyUI.
2. GRAB THE MODEL FILES YOU NEED
If you're running Krea-2:
Full GGUF weights + required text encoder/VAE + patched loader instructions are on my model page:
https://civarchive.com/models/2726875/rebels-krea-2-ggufs
Quick list if you don't have them yet:
- Custom node (required for GGUF loading): RealRebelAI/ComfyUI-GGUF_KREA-2 — https://github.com/RealRebelAI/ComfyUI-GGUF_KREA-2 (standard GGUF loaders don't recognize Krea-2's architecture tag yet, you need this fork)
- GGUF weights: https://huggingface.co/realrebelai/KREA-2_GGUFs → ComfyUI/models/unet/
- Text Encoder: Qwen3-VL-4B-FP8-Scaled.safetensors → https://huggingface.co/Comfy-Org/Qwen3-VL/tree/main/text_encoders → models/text_encoders/ (load with CLIPLoader, type must be set to "krea2")
- VAE: qwen_image_vae.safetensors → https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae → models/vae/
If you're running Z-Image Turbo (ZIT):
Full details on my model page:
https://civarchive.com/models/2169770/rebels-z-image-turbo
Quick list if you don't have them yet:
- GGUF weights: https://huggingface.co/unsloth/Z-Image-Turbo-GGUF/tree/main → models/unet/
- Text Encoder: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/text_encoders → models/text_encoders/ (load with CLIPLoader, type must be set to "lumina2")
- VAE: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/vae → models/vae/
Only grab what you're missing — if you're already running either model, you likely have all of this installed already.
3. INSTALL THE REALESRGAN X2 UPSCALE MODEL (REQUIRED)
The staged pipeline needs a real SR model for the pixel-space upscale step — this is the whole point of MrFlow, don't skip it or swap in a plain latent upscale.
Easiest way — via ComfyUI Manager:
Open the Manager button → Model Manager (or "Install Models") → search "RealESRGAN" → install the x2 variant. It drops into models/upscale_models/ automatically.
Manual alternative:
grab RealESRGAN_x2.pth from the official releases page (https://huggingface.co/rklaumbach/RealESRGAN_x2/blob/main/RealESRGAN_x2.pth)
or
grab 4x_foolhardy_remacri.pth from the official release page (https://huggingface.co/FacehugmanIII/4x_foolhardy_Remacri/blob/main/4x_foolhardy_Remacri.pth)
and place it in ComfyUI/models/upscale_models/ yourself.
THE WORKFLOW
Both graphs follow the same shape: KSampler at low resolution → decode → RealESRGAN x2 upscale → re-encode → short 1-step refine at full res → save. The Preset node drives resolution/steps/cfg/denoise for the whole chain — change your target size in one place and it propagates. GGUF loader swaps in for the diffusion model in both, keeping VRAM low without sacrificing refine quality.
Workflow JSONs are attached below — drag straight into ComfyUI.
Description
zit in pixel space
FAQ
Comments (4)
This looks really interesting also what is the difference between this and the great https://huggingface.co/Comfy-Org/PixelDiT
this gens low res like pixel diffusion but UPSCALES in pixel space instead of genning in it. saves a ton of time but at a greater quality loss admittedly. this seemed like more of a concept than a standard release if you ask me. cool concept but not really useful. i just posted it becaause i put the work in to test it type shiii :)
@realrebelai Do you think it's honestly worth it or do you think it's better just to do a two-pass workflow and start with a lower resolution and then upscale it to a higher one as that is what I usually do with anima with the turbo lora for example I start at 512 by 512 at 12 steps on the first k Sampler and then I upscale latent by to 1024 by 1024 in a second ksampler at a denoise of 0.5 or 0.55 at another 12 steps or 8 or even 6. It adds so much detail and basically a hi Res fix. I have used Comfy-Org/PixelDiT with Krea 2 and it's great takes longer tho I also like the RTX super resolution node.
@AnimaXx id say for speed yeah its worth it, but for quality no. pixeldit would do it better or even PID at this point if you could get the math right