Just ran a full comparison of three training methods on SDXL: full DreamBooth, LoRA, and LoRA extraction. Here's what I found.

The Forge / 工坊 Models / LoRA / 模型与 LoRA
1

Just ran all three methods with the latest Kohya GUI, base model is official SDXL 1.0 base. Full DreamBooth with text encoder training, used a rare token + class token (like ohwx man), paired with a set of manually curated regularization images.

Training data is 15 of my own photos, deliberately kept at mid-tier quality—wanted to show that even this level of dataset can work. Did 150 repeats, 1 epoch, which translates to 4500 steps based on Kohya’s repeat logic.

For LoRA training, I followed the hyperparams from that public tutorial, feels like there’s room for tuning. For LoRA extraction, I compared FP16/FP32/BF16 precision, extracted from the DreamBooth model at 128 rank, 128 alpha—rank and alpha could probably be optimized further too.

Overall, full DreamBooth gives the best results but costs the most, LoRA is the best bang for your buck, and extraction depends on whether you already have a full model lying around.

2

Honestly, even with a mid-tier dataset you can still get decent results, which is pretty solid.

3

128 rank seems a bit high, isn’t it? Probably easy to overfit.

4

Training the text encoder together definitely gives better face reconstruction.

6

Extracting LoRA saves VRAM, but only if you already have the full base model.

7

Collecting regularization images yourself is way too much of a hassle.

8

BF16 extraction, how’s it working? Waiting for the details.

9

128 works fine for training art styles, but for characters I usually just go with 32.

1 Like