Need some information about both pretraning and swap training

[nettinParticipant]

Hello everyone, I’m new here.

First, let me talk about my experiences. On Windows, system applications reduce the available VRAM capacity, so I started training on native Linux. I’ve seen that it processes things at least 2x faster and I also have more free VRAM space.

1. When I set use_f16 = true in the model settings, the performance and VRAM usage become more optimal. The problem is that a pretraining from scratch collapses with f16. **Initially, the swap process also collapses with f16.** After a certain point, I can increase the batch size with use_f16=true, and the speed increases too, but it feels like the noise also increases. What are your thoughts and suggestions on this?

2. AI chats recommend src_random_flip = true and dst_random_flip = false for both pretraining and swap training, but I usually see the opposite in the forums. I would like to know your insights on this.

3. I would be grateful if you could inform me about the recommended starting settings for both training phases, and in which situations we should change which settings and continue training.

[nettinParticipant]

There are phrases in the tutorial, but that many steps means a lot of work. Is there a simpler method you can recommend?

Are the following statements true? (by deepseek aichat)

“Instead of strictly following that complicated plan, use a much simpler and more effective 3-Step Plan:

STEP 1: GENERAL TRAINING (Random Warp ON)

random_warp: True

batch_size: As high as possible.

All Other Settings: Default or Off (0.0).

When to stop? When the loss value stops dropping and becomes stable (plateaus), and in the preview, the face looks consistent but still a bit blurry. (~30,000 – 60,000 iterations)

STEP 2: REFINEMENT (Random Warp OFF)

random_warp: False (This is the most important change!)

batch_size: Same or increased.

Optional: If eyes/mouth are blurry, turn on eyes_mouth_prio: True and mouth_prio: True.

Optional: For color/lighting issues, start with true_face_power: 0.001 and increase it slowly.

When to stop? When the preview looks perfect and the loss value is dropping very, very slowly or is stable. (~15,000 – 30,000+ iterations)

STEP 3: ENHANCEMENT (GAN – Optional)

Start with gan_power: 0.01 and slowly increase to a maximum of 0.1.

gan_patch_size: 32-64

gan_dims: 16

When to stop? When skin texture and details are as sharp as you want. Don’t overdo it, or it will look fake! (~5,000 – 15,000 iterations)

✅ The Bottom Line:
The guide you have is technically correct but is too complicated and bulky in practice.

If I were you: I wouldn’t stress myself trying to follow a 10-step plan. I would just focus on perfecting STEP 1 and STEP 2. Only add the GAN (STEP 3) if you really need it for that final touch.

Simplicity is always the best strategy. Complex plans are methods that experienced users turn to for solving very specific problems. Your first goal should always be to build a simple model that works and gives good results.”

[Author]

Posts