détails

对脸部等进行化妆，细节增强处理，虽然都是变漂亮，但和美颜磨皮消除细节的过程刚刚相反。相信大家也应该更喜欢这个。

对脸部等进行化妆，细节增强处理，虽然都是变漂亮，但和美颜磨皮消除细节的过程刚刚相反。相信大家也应该更喜欢这个。

Just making it more real photo but in potrait position

A different approach in the space of upscaling, Rather than upscaling the final image, Latent upscaling in the middle of the process can increase the details in an image significantly. I have implemented this idea in the comfy UI workflow and made an 8K image. This workflow is very good for beginners because i have added few useful notes in it. So feel free to use it and post your work and feedback.

Pony Diffusion model to create images with flexible prompts and numerous character possibilities, adding a 2.5D LoRA of details for more styling options in the final result. The workflow has Upscale resolution to 1024 x 1024 and metadata compatible with the Civitai website (upload) after saving the image. I added the node clip skip -2 (as recommended by the model), remembering that in ComfyUI the value -2 is equal to 2 (positive) in other generators (Civitai, Tensorart, etc). After several tests, I realized the need to have a specific node group for details on faces, especially when using the Ponny model, so this new version has ComfyUI Impact installed with DDetailer Feature to detect faces and refine them automatically. I've left a note with the link explaining the details of how the tool works. NOTE: An interesting detail is that you can make the lines of the nodes “cleaner”, in other words, straighter. To do this, just follow these steps: 1 - Click on the ComfyUI “settings” gear, the icon is above the QUEUE PROMPT button, next to the name QUEUE SIZE: 1. 2 - In settings, scroll down to almost the bottom of the page and check the following option: LINKS RENDER MODE 3 - From the various options, choose STRAIGHT 4 - Click Close to exit the settings and you will now see the lines of the nodes “straight”, making it easier to understand the flows. The following additional nodes will be required: - UltimateSDUpscale (upscale) - SaveImageWithMetaData (metadata) - DDetailer (ComfyUI Impact Pack)

The Anytest ControlNet seems to be some kind of combination of ControlNet models, but no one knows which ones or how it works exactly, but is definitely the best model.It is very good at keeping the pose, proportions and of the input image, while changing style and details. No preprocessing needed.Use a strength of 0.50-0.60 if you are using sketches or non anime pics, of 0.70-0.80 if using anime pics or keeping the original style.Img2img is possible, but may require some tweaking.

最新的FLUX大模型结合lora和Controlnet用于图片的不同风格转绘。 This workflow is all built on top of FLUX. Xlabs-flux-controlnetV3 is great. FLUX's lora is also great.The details are optimized through subsampling, and the final image is of very high quality. （这套工作流全是基于FLUX搭建的。Xlabs-flux-controlnetV3很棒。FLUX的lora也很棒。通过二次采样进行细节优化最终出图质量非常高。） TEST： Black Myth Wukong style repainted（黑神话悟空风格转绘） You need to upgrade to the latest Comfyui, and the nodes that appear can basically be installed directly in the manager（需要升级到最新的Comfyui，出现的节点基本都可以在manger里面直接安装。） flux： https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main Xlabs: https://github.com/XLabs-AI/x-flux controlnet: https://huggingface.co/XLabs-AI (v3) lora： https://civitai.com/models/681691 https://www.liblib.art/modelinfo/3945f0e21a4b4bb5965b4319d927dd4f?from=search https://www.liblib.art/modelinfo/f00b5b28b87b4935a99f651499495ffc?from=search Xlabs-lora： https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main

Flux Kontext Fast + Upscale Workflow Status: Work in Progress - Still refining output quality This workflow combines the power of FLUX.1 Kontext with the Alimama Turbo LoRA to achieve fast, high-quality image generation and upscaling in just 8 steps. Key Features Fast Generation: Utilizes the alimama-creative/FLUX.1-Turbo-Alpha LoRA for rapid image generation with only 8 inference steps Context-Aware Editing: Leverages FLUX.1 Kontext capabilities for intelligent image editing and manipulation Two-Stage Process: Initial generation followed by upscaling for enhanced detail and resolution Efficient Workflow: Optimized for speed while maintaining quality output Workflow Components Stage 1: Fast Generation Model: FLUX.1 Kontext with Turbo Alpha LoRA Steps: 8 (significantly reduced from standard 20-50 steps) Guidance Scale: 2.5 (optimized for Turbo LoRA) Sampler: Euler with simple scheduler Stage 2: Upscaling Method: Ultimate SD Upscale with custom sampling Upscale Model: 4x-ClearRealityV1 Tile Processing: Intelligent slicing for memory efficiency Denoising: 0.2 strength for detail enhancement Technical Setup Required Models FLUX.1 Kontext DEV (FP8 quantized) Alimama FLUX.1-Turbo-Alpha LoRA 4x upscale model Standard FLUX text encoders (CLIP-L + T5) Memory Optimization FP8 model quantization for reduced VRAM usage Tile-based upscaling to handle large images Optional caching nodes for faster iterations Prompt Guidelines The workflow works best with detailed, descriptive prompts. The Turbo LoRA has been trained on high-aesthetic images (6.3+ rating) with resolution >800px, so it excels at: Detailed scene descriptions Lighting and atmosphere specifications Style and aesthetic directions Object replacement and background changes Current Limitations ⚠️ Work in Progress: Output quality is still being refined Some inconsistencies in fine details Occasional artifacts in upscaled regions This workflow represents an efficient approach to context-aware image editing with FLUX, trading some quality for significant speed improvements. Perfect for rapid prototyping and iterative design work.