FLUX De-Distilled Fine-Tuning & LoRA Tests: Bleeding and CFG Results
FLUX de-distilled models train best at guidance scale 3.5 instead of 1, yet these DreamBooth and LoRA experiments show they still do not fix bleeding.

Highlights
Anti-bleeding tests train two people together on FLUX Dev, Verus Vision 1.0b, PixelWave v03 and FLUX Dev De-Distilled with 28-image datasets. Also tests the CFG impact for stylized images such as anime or paintings on regular FLUX Dev, where CFG 3.5 is worth testing.
What it tests
Experiments with de-distilled FLUX models for full fine-tuning / DreamBooth and LoRA training, based on our FLUX fine-tuning workflow (tutorial, configs post).
Key findings
- With de-distilled models, the only setting to change in our FLUX configs is Guidance Scale: 3.5 instead of 1. 1, 2.5 and 3.5 were tested, and 3.5 was best.
- For inference on de-distilled models, CFG 3.5 gave the best quality compared with 2.5 and 3.0.
- Community fine-tuned models that are not de-distilled, like PixelWave, train the same way as the base FLUX Dev model.
- FLUX Dev De-Distilled has less bleeding than FLUX Dev when two people are trained together, but it does not fix it.
- Higher CFG on regular FLUX Dev for stylized images such as anime or paintings is hit or miss.
Models tested
FLUX Dev, Verus Vision 1.0b, PixelWave v03 and FLUX Dev De-Distilled, trained with 28-image datasets and compared on full grids (dataset post).
Video tutorial
Train FLUX on Your Own Photos With Just a 6 GB GPU: Full Fine-Tuning Course (Kohya, 2024) · chapters and summary
Related

Qwen Image 2512 Training Comparison: The Best Model to Train
Qwen Image 2512 is the new king for training: the video compares trained Base, 2512, Edit 2509 and Edit 2511 models and shows a big leap in realism.

FLUX 2 vs FLUX SRPO & New Kohya SS GUI Premium FLUX Training App
FLUX 2 Pro compared with a trained FLUX SRPO model, plus the new Kohya trainer V35 that trains FLUX on GPUs with as little as 6 GB VRAM.

FLUX LoRA vs Fine Tuning / DreamBooth Comparison: Batch Size 1 vs 7
FLUX LoRA and Fine Tuning / DreamBooth compared with 15 and 256 training images at batch size 1 and 7, for both realism and stylization.