I occasionally get asked how I build LoRA training datasets, so I customized this tool based on a project by @tazmannner379 (thanks for sharing) .It generates captions from images or videos using a VL model and integrates with ComfyUI for visual feedback.<a target="_blank" rel="ugc" href="https://github.com/zxtopower-tech/LoRA-Caption-Assistant">https://github.com/zxtopower-tech/LoRA-Caption-Assistant</a>Uses paid APIs (Google or OpenAI-compatible). Local models require Ollama or LM Studio.<hr /><h2 id="2025-09-06">2025-09-06</h2>I'm finally releasing <code>Seoullina v2.1</code>.This was s'posed to be a small, quick fix, but it ended up getting delayed as a few things came up.TBH, I'm not entirely happy with how it turned out.The main issue's that the skin texture has reverted to something closer to v1.0. As some of you might've seen in the comments, my SSD completely failed. They couldn't recover the data, so they just sent me a new one. I tried my best to rebuild the dataset from scratch and replicate the v2.0 look, but I just couldn't quite get it right.Btw, the breast physics bug seems to be fixed.I know that's a whole saga, so for anyone who just wants the TL;DR, jump to the <code>About this version</code> section.<h2 id="2025-08-26">2025-08-26</h2>Alright, so after releasing<code>Seoullina v2.0</code>, I've spotted a few problems.The biggest issue is a bug that causes the breast physics to bounce around even when the character is completely still.I can't tell you how hard I laughed when I first saw it.😂😂My guess is that it happens on heavily quantized models, at low resolutions, or when cleavage is exposed, but I totally missed it. For this issue, please refer to the video that <code>kyo55966933</code>uploaded to the gallery (and thanks for the video!)I'm going to investigate the issue and prepare for version 2.1.<h2 id="2025-08-24">2025-08-24</h2>First, a huge thank you to everyone who has been sharing their work in the gallery. The results are absolutely fantastic!For <code>Seoullina v2.0</code>, I've made some important updates to the training process.Previously, my preferred method was to train only the low-noise model and use the Wan 2.1 Self-forcing LoRA at a high strength (e.g., 2). This produced great image quality, But I now believe the high strength was a major contributor to the deforming issue.However, I've found an approach that yields better results. By training both the High/Low models together and using the <code>Wan 2.2 Self-forcing LoRA</code>(<a target="_blank" rel="ugc" href="https://civarchive.com/models/1585622?modelVersionId=2080907">link1</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/1585622?modelVersionId=2081616">link2</a>), the deforming issue has been significantly improved.The training dataset has also been revised.It came to my attention that the less realistic skin texture from images generated by Flux was sometimes undesirable. To address this, I've updated the dataset, which has resulted in the character's appearance now being different from the v1.0.As the original goal is to create a fictional character, I believe this change is a positive step forward.<h3 id="recommendations">Recommendations</h3>The Wan 2.2 Self-forcing LoRAs (High/Low) are required!Sampler: eulerScheduler: betaShift: 8+Steps: 4+CFG: 1?<hr /><h2 id="2025-08-11">2025-08-11</h2>The Busansuji version has been released.In the <code>Seoullina v1.0</code>, there was a deforming issue.I wanted to fix it, but I couldn't find the exact cause, as it occurred even when training both the high and low models.For the <code>Busansuji v1.0</code>, I've selected a model that should minimize this problem.I'll continue to experiment with different approaches to find the best dataset configuration.<hr />For the Self-forcing LoRA, I suggest using the <code>wan 2.1 64-rank model</code>(<a target="_blank" rel="ugc" href="https://civarchive.com/models/1585622?modelVersionId=2014522">https://civarchive.com/models/1585622?modelVersionId=2014522</a>).And set the strength to 2 for the high model and 1 for the low model.Also, using the <code>euler</code> sampler is recommended for the best results.<hr />About This LoRAFirst off, I used the <code>Diffusion-pipe</code> for the training.I really want to say thank you to <code>tdrussell</code> for updating the tool so fast, right after Wan 2.2 came out.I made a total of 206 images with the Flux.D model and did my best to include many different conditions.For example, I used various lighting like day, night, indoor, and club scenes, and the shots were a mix of close-ups, medium, and full shots.But, I know it's lacking when it comes to camera angles.It's a shame that I couldn't put together data with more varied angles.However, I think the camera movement in the new wan2.2 model is better, so I believe this will overcome that weak spot.UsagePlease use the LoRA as both the High and Low model LoRA.

Wan2.2 - Korean Women

✅ Tested with Self-forcing LoRA for Wan 2.2Training Details:Model: Wan2.2-T2V-A14B - High/Low noiseLoRA Rank: 32Learning Rate: 4e-5Dataset: 76 samples with variable resolutions

Files with SHA256: 1853a113a3ca6bb938b54f26baaf1572528dccf95fc46b21a6ffca50ead8ea5d