Input a single full-body image, auto-upscale it to high resolution(2MP), and generate a multi-view showcase grid focusing on face, legs, and feet details seamlessly, good for Character LoRA dataset preparation.
V1.2
1. Flow Control: To avoid frequent model switching, loading, and unloading, added a Preprocess pipeline to obtain masks ahead of time.
2. Batch Image Processing: Organized and structured nodes into Subgraphs, and modified Face/Legs/Feet into Batch Image processing based on model types (Qwen/K9B/Krea2) step-by-step, improving processing speed.
Models
Krea2 - Moody Krea 2 Mix V8
Qwen - Phr00t/Qwen-Image-Edit-Rapid-AIO
└ INT8 Convrot/INT6 W6A8 Conversion(19G/14.9G):
HuggingFace - waveshaker/comfy_models/qwen
INT6 is just a bit slower than INT8 (about 33s vs 30s), but the model size is smaller and the results are pretty much the same.
Flux2 Klein 9B - Moody Desire Mix v3.0
└ INT8 Convrot Conversion - moodyDesireMix_v30_fp16_ctq_int8_convrot.safetensors
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 sam3/
│ ├── 📄 sam3.ptSAM3(Official: FP32) - facebook/sam3
└ FP16/Mixed FP8/FP8 version:
HuggingFace - waveshaker/comfy_models/sam3
Civitai - SAM3 - FP16/Mixed FP8/FP8
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 checkpoints/SAM3.1 - Comfy-Org/sam3.1
└ INT8 Convrot Conversion - SAM 3.1 INT8 + INT4 Native ConvRot
YOLO Detection Models
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 ultralytics/LoRA
Krea2
HighQuality - High resolution
Qwen
Qwen-Image-Edit-2511-Unblur-Upscale - Qwen-Image-Edit-Unblur-Upscale_20.safetensors
Flux 2 Klein 9B
f2k_9B_lcs_consist_20260415 - Consistence Edit Lora
V1.1
Fully replaced SAM3 Segmentation (RMBG) with comfyui-easy-sam3, resolving the OOM issue caused by the RMBG SAM3 failing to clear VRAM.
Added a toggleable Krea 2 refine step into every original Upscale step; Krea 2 enhances skin realism.
Note: There is nothing wrong with the quality of the RMBG custom node itself, and SAM3 Segmentation (RMBG) does provide a model unloading option. However, when the workflow requires heavy and frequent use of it to generate masks, the model unloading speed of RMBG's SAM3 Segmentation is too slow, which increases the total workflow execution time.
V1.0
⚠️ VRAM Notice: SAM3 Segmentation (RMBG) in the RMBG custom node currently causes VRAM OOM issues.
Please use either comfyui-easy-sam3 (with sam3.pt) or ComfyUI native SAM3 Detect (with SAM3.1) instead. Alternatively, I'll release a fixed version soon.The new version will use Krea2 for more realistic skin texture.
Description
Fully replaced SAM3 Segmentation (RMBG) with comfyui-easy-sam3, resolving the OOM issue caused by the RMBG SAM3 failing to clear VRAM.
Added a toggleable Krea 2 refine step into every original Upscale step; Krea 2 enhances skin realism.
Models
Krea2 - Moody Krea 2 Mix V8
Qwen - Phr00t/Qwen-Image-Edit-Rapid-AIO
└ INT8 Conversion - Qwen-Rapid-AIO-NSFW-v23_diffusion_only_fp8-int8_convrot.safetensors
Flux2 Klein 9B - Moody Desire Mix v3.0
└ INT8 Conversion - moodyDesireMix_v30_fp16_ctq_int8_convrot.safetensors
SAM3.1 - Comfy-Org/sam3.1
Detection Models
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 ultralytics/📂 ComfyUI/
├── 📂 models/
│ ├── 📂 sam3/LoRA
Krea2
HighQuality - High resolution
Flux 2 Klein 9B
f2k_9B_lcs_consist_20260415 - Consistence Edit Lora
FAQ
Comments (8)
v1.1. Bro, Can you add more detail for the hands?:) I haven’t found any good or even decent ways to fix hands. I suggest you make a Subgraph`s. Not just for organization, but to eliminate the performance issues with the ComfyUI interface. https://drive.google.com/file/d/1nLOigKPBuaD0ZK9AvH-gsfdG7lZ9yvXr/view?usp=drive_link
Thanks for your feedback. Hand detail fixing is definitely a high-frequency pain point right now.
1.Currently, my idea for improvement is using a YOLOv8 hand-related Detailer, or using SAM3 for step-by-step mask cropping to get the hands, then Inpainting them using a method similar to the feet approach and pasting them back into the original image.
2.As for Subgraphs, the reason I don't use them is that frequent ComfyUI updates have broken many of my previous workflows. Additionally, keeping the nodes out of subgraph helps new users better follow and understand the underlying logic.
I will improve these two points when I have time.
Also, Google Drive link requires permission to access.
@rogerpie I’ve fixed it; I think it won’t be possible to achieve a high‑quality result without ControlNet and OpenPose. That is, it’s possible to determine that there aren’t five fingers but four, but the neural network can’t correctly render the hand and its pose - a skeleton is needed.
@beautifulmur889 Have you tried extracting my face upscaling method from the workflow? Get the hand mask, and then you can try inpainting it. Qwen1 performs well at restoring blurry images into artifact-free images, Klein 9B is suitable for upscaling, Krea 2 is good for enhancing skin details, and finally pasting it back to the original image.
@beautifulmur889 Although Qwen1 is relatively slow, it's the absolute best at restoring blurry or artifact-heavy images into clean, high-clarity ones—all while sticking as closely as possible to the original image.
I’ll give it a try. Right now, I’m looking for a way to fix my hands, and that’s how I found your project — it’s quite interesting.
Check this https://civitai.red/posts/31393729 post, modified workflow is included.
Before processing, the original image is upscaled using SeedVR 7B with a denoise value of 0.5 to minimize changes to the original composition. In this specific image, the hand area is contained within the leg area. The processing steps are as follows:
- Leg Area Refinement: First, I used 2 Samplers combined with Qwen Image Edit to fix and refine the details of the leg area.
- Hand Area Processing (Based on 1MP Resolution): Next, the hand area is cropped and scaled to 1MP as the base resolution. On this 1MP canvas, the workflow proceeds as: Crop Image → K9B → KR2 (to enhance hand skin details).
- Composite & Merging: Finally, the processed hand is pasted back onto the leg area, and then the leg area is pasted back into the full-body image.
Note: Since hand positions vary greatly in practice, you may need to save the hand processing for the very last step, especially if the hands are obscuring the face or other body parts.