base_vs_distill.md

Base (100 NFE) vs Distilled (8 NFE)

← Back to README.

Run Base and Distilled Model

# Taking T2I for example
# Run Base
python examples/t2i/inference.py \
    --model_path SenseNova/SenseNova-U1-8B-MoT \
    --jsonl examples/t2i/data/samples.jsonl \
    --output_dir outputs/ \
    --cfg_scale 4.0 --cfg_norm none --timestep_shift 3.0 --num_steps 50 \
    --profile
# Run Distilled
python examples/t2i/inference.py \
    --model_path SenseNova-U1-8B-MoT-8step-preview \
    --jsonl examples/t2i/data/samples.jsonl \
    --output_dir outputs/ \
    --cfg_scale 1.0 --cfg_norm none --timestep_shift 3.0 --num_steps 8 \
    --profile

---

Text-to-Image

SenseNova-U1-8B-MoT (100 NFE)	SenseNova-U1-8B-MoT-8step-preview (8 NFE)

Image-Editing

Reference Image	SenseNova-U1-8B-MoT (100 NFE)	SenseNova-U1-8B-MoT-8step-preview (8 NFE)

Existing Issues

Two issues have been identified in the SenseNova-U1-8B-MoT-8step-preview (8 NFE), and we are actively working to resolve them.

• Grid artifacts may occur in certain instances.

• In image-editing tasks, the generated results may exhibit altered color tones.