Since support for MOSS-Image in OneTrainer—the tool I usually use—is lagging, I’m testing a full fine-tuning (FT) process using a different tool.
I’ve set this up as a "Finetune Experimental Test" using a separate model instance.
ver. 0.1
Honestly, it’s been quite a struggle; I tried modifying various other tools, but they all failed—resulting in nothing but bizarre, garbled images. This time, I used "Fizgig." It’s an interesting tool, though it breaks the training process into four separate stages, which makes getting results a rather slow affair. It keeps the model in VRAM using NF4 quantization for the actual training, then writes the data back to BF16 format in the main system memory. Since the optimizer state is maintained in BF16 precision, the overall precision isn't actually that low; the theory is that any shortfall in computational precision is compensated for by a higher number of training steps (Since settings vary depending on the environment, it is highly likely that processing is being performed using BF16 in this instance.).
The output here is from the very first epoch, so this is strictly an experimental, preliminary test.
ver.0.13 LoRA
Actually, I had already finished up to epoch 3 by yesterday, but in terms of balance, epoch 1 felt the most practical to use. As the epochs progress, elements that weren't originally in an anime style shift closer to an anime aesthetic, while things that were already in an anime style in epoch 1 remain largely unchanged. I don't plan on releasing the full checkpoints, but I have created a LoRA based on the difference between epoch 3 and the base model, which I’ve added as an optional file.
ver.0.2
This is a test of fine-tuning (FT) using a local checkpoint (ckpt) within Fizgig. Since this cannot be specified via the GUI, I modified the configuration file directly. This also serves as a functional check, as changes to configuration files in this tool are not always applied correctly.
I used MOF-FET v0.1 as a base, along with a small dataset (just under 10,000 images) featuring a different artistic style. The model leans toward a "semi-realistic" aesthetic, though it lacks any particularly distinct characteristics. Compositional quality and prompt adherence have both declined; given how well-crafted the base model is, it is likely better not to stray too far from it.
To be honest, the results aren't great, but I’ve decided to release it as a "test model." Another reason is that I needed a suitable model to use as an example when explaining how to use the merger tool—there aren't many models available to mix, after all.
Incidentally, when generating images with MOSS-Image without providing a prompt, the output often consists of documents, erotic manga, or pamphlets; perhaps "eromanga" is the prompt this model handles best.
ver.0.3
It’s a bit hard to define, but this version uses a "base model" in the literal sense. I had originally intended to use a less powerful, pre-trained model as the foundation for this test, but an error—likely triggered by an update to Fizgig—prevented that.
Compared to ver.0.1, this is a standard, "typical" anime-style character model. Since it seems that lighter fine-tuning (FT) works well for MOSS-Image, I used this as an opportunity to test a more streamlined, less labor-intensive approach; the total training time was only about 3 hours. About an hour of that was spent on cache generation (though I suspect the cache might not even be necessary for a single-epoch run). Even so, it performs reasonably well.