alvdansen/linnea-qwen-image-baseline

Linnea — Qwen-Image (Baseline)

The monotonic-training counterpart to `alvdansen/linnea-qwen-image`. Same 27-image dataset, same hyperparameters, same hardware — trained without the chained dataset rotation. Published alongside the chained version so the paired comparison in Forgetting on Purpose is reproducible. Read the paper · Source on GitHub.

For general use, the chained version is the better pick. This one is here for anyone who wants to run their own comparison — see Figure 2 of the paper for the cleanest difference between them.

Usage

Trigger word: linnea

Same inference recipe as the chained twin:

Sampler: euler
Scheduler: simple
CFG: 3.5
Steps: 45
LoRA strength: 0.8–1.0

Training Details

• —Base model: Qwen-Image (FP8 quantized, text encoder FP8)
• —Training steps: 35,750, monotonic full-dataset training from step zero
• —Rank/Alpha: 42/42
• —Learning rate: 5e-5
• —Optimizer: AdamW 8-bit
• —Caption dropout: 0.25
• —EMA: enabled (decay 0.99)
• —Noise scheduler: flowmatch
• —Precision: bf16 with qfloat8 quantization
• —Dataset: same 27-image combined dataset as the chained twin
• —Trainer: ai-toolkit by Ostris
• —Hardware: NVIDIA RTX 6000 Ada (A6000, 48 GB VRAM)
• —Wall-clock: ~24 hours

Identical to the chained twin in every parameter except the dataset schedule.