Version 2 coming soon...<hr />This flow is for making character dataset images, using Face Swap, for Lora creation. (A fancy way of saying 'lots of pictures'.) It will take your loaded image, generate a front and back view, generate multi angle shots from those, then go on to take your character and put them in different clothes, poses, and lighting conditions, face swapping along the way.<hr />This flow is dedicated to <a target="_blank" rel="ugc" href="https://civarchive.com/user/lonecatone23">Lonecat </a>as I used his flows for ideas and for all of my learning. Thanks bud, for all you contribute to the community.<hr />Version 1.6Just some minor prompt and detailer tweaks.Should be a wee bit faster now. (Even put in a render timer. Quite handy!)<hr />Versions 1.5 I started playing with AI image generation less than a year ago at the time of writing this. Like many people, I jumped straight into ComfyUI and immediately got lost in a sea of nodes, spaghetti workflows, and “wait... what does this node even do?” moments.A lot of my early learning came from studying workflows created by <a target="_blank" rel="ugc" href="https://civarchive.com/user/lonecatone23">Lonecat</a>. The outputs were incredible, but the workflows themselves were absolute chaos to a newcomer — which honestly made them a fantastic way to learn. Nothing teaches ComfyUI faster than trying to untangle someone else’s 400-node masterpiece.Over time, I became interested in training my own LoRA models, specifically to create a realistic representation of my <a target="_blank" rel="ugc" href="https://www.thegreenhead.com/imgs/handbook-for-recently-deceased-from-beetlejuice-1.jpg">recently deceased</a> father using decades of family photos and old recordings of conversations we had before he passed away. Yeah... maybe a little creepy, maybe a little emotional, but it became a huge motivator to really understand how all of this worked.The biggest challenge quickly became dataset creation. I could find workflows that got “close enough,” but I kept running into issues with consistency, angles, expressions, anatomy cleanup, identity preservation, or just workflows that were nearly impossible for newer users to follow or modify.So I started building my own.At first, it was just a handful of borrowed ideas mashed together while I learned. Then it slowly evolved into something much larger:<ul><li>* staged dataset generation</li><li>* angular generation</li><li>* optional NSFW refinement</li><li>* face swapping</li><li>* automated detailers</li><li>* reusable latent/model systems</li><li>* modular switches</li><li>* cleaner workflow organization</li><li>* educational layout design</li></ul>At some point, it stopped being “someone else’s workflow with edits” and became its own thing entirely.This workflow is designed to be both production-usable AND educational. I intentionally kept many sections visually separated and organized so newer users can follow the logic, disable stages, experiment safely, and learn how the pipeline works without needing a PhD in node spaghetti.The workflow is heavily modular:<ul><li>* Front/Back generation</li><li>* Angular generation</li><li>* Dataset stages</li><li>* Optional NSFW refinement</li><li>* Face swap systems</li><li>* Detailer systems</li><li>* Switch-controlled sections</li></ul>You can run sections individually as you progress through dataset creation instead of generating everything all at once.This workflow was built through months of experimentation, failure, rebuilding, optimization, and learning-by-doing. Hopefully it helps someone else who is starting that same journey.

F2\K Q XL Ernie Dataset Creation - SFW/NSFW

Version 2 of this workflow became less about “adding more stuff” and more about rebuilding the architecture after finally understanding how ComfyUI actually works under the hood. This flow is dedicated to Lonecat as I used his flows for ideas and for all of my learning. Thanks bud, for all you contribute to the community. Major V2 changes include:<ul><li>Shared SAM model loading instead of repeated SAM loaders</li><li>Shared face model loading for all face swap stages</li><li>Consolidated latent reuse systems using Set/Get nodes</li><li>Reduced unnecessary VAE encode/decode operations</li><li>Reworked detailer structure and optimization</li><li>Reduced over-processing in anatomy refinement stages</li><li>Improved face swap quality and reintegration</li><li>Cleaner modular execution paths</li><li>Better workflow grouping and organization</li><li>Simplified dependency chains</li><li>Improved stage-by-stage dataset generation flow</li><li>Expanded and reworked optional NSFW refinement system</li><li>Additional dataset generation stages</li><li>Updated prompts and model selections</li><li>Improved readability and educational structure</li></ul>One of the biggest goals for Version 2 was maintaining visual readability while still improving performance and modularity. Many sections remain intentionally separated and grouped so newer users can follow the workflow logic more easily instead of trying to decipher a giant wall of node spaghetti.The workflow is still designed around staged execution:<ul><li>build front/back references</li><li>generate angular references</li><li>generate dataset stages</li><li>optionally run refinement/detailer passes</li><li>progressively improve consistency and identity preservation</li></ul>The result is a workflow that is significantly cleaner, more modular, easier to troubleshoot, and considerably more efficient than the original versions while still remaining approachable for people learning ComfyUI.The original versions were functional, but very much built during the “learn by brute force and experimentation” phase. Over time, I started understanding:<ul><li>latent reuse</li><li>VAE encoding/decoding overhead</li><li>model loading efficiency</li><li>detector/detailer costs</li><li>modular execution</li><li>workflow dependency cleanup</li><li>face swap optimization</li><li>memory management</li><li>switch-controlled execution paths</li></ul>Version 2 is a major cleanup and optimization pass focused on keeping the workflow educational while dramatically improving how the pipeline was structured internally.

Files with SHA256: 117340d3f514d9ea6eff1a619741b0ae707c158324b257ec3ddf7e66f696eb93