mario-d-robot/qwen-image-edit-batch

Qwen-Image-Edit-2511-LoRAs-Fast — Batch Driver

Single-file headless batch driver for Qwen-Image-Edit-2511 inference with weighted multi-LoRA composition.

batch_run.py is a self-contained UV-script: drop the dep header, the FA3 processor, the QwenImageEditPlusPipeline fork, and the 4-stage batch driver into one ~1400-line file. No sibling imports. Runs anywhere hf jobs uv run can reach.

Quick start

hf jobs uv run \
    --flavor l40sx4 \
    --timeout 4h \
    -v hf://buckets/mario-d-robot/Qwen-Image-Edit-2511-LoRAs-Fast-storage:/data \
    -e HF_HOME=/data/cache \
    https://huggingface.co/datasets/mario-d-robot/qwen-image-edit-batch/resolve/main/batch_run.py

See BATCH.md for the full pre-flight + launch + collect workflow.

What's in here

File	Purpose
batch_run.py	The entire driver — UV-script deps, FA3 processor, Qwen-Image-Edit-2511 pipeline fork, 4-stage batch loop
models_config.json	Pipeline + Rapid-AIO transformer repos, pinned by revision
adapters_config.json	The 10 definitive LoRAs with incompatibility rules
job.json	Template showing the runtime interface (upload to bucket before launch)
BATCH.md	Detailed launch recipe, troubleshooting, examples

Brilliant patterns preserved

Lifted from luminous-ultra-studio (a 6-month headless pipeline project):

• Norm-preserving CFG in the denoise loop — comb_pred rescaled to ||cond|| so true_cfg_scale=4.0 doesn't saturate.
• Per-condition-image img_shapes for RoPE — without it, condition image tokens get treated as new-image tokens → subtle position artifacts.
• Mu-shift sigma schedule for FlowMatch at varying image_seq_len.
• Multi-image Picture N: token string so the prompt can reference "Picture 1", "Picture 2", ... by position.
• VAE tiling + slicing enabled by default — cheap insurance against decode OOM on large outputs.

License

Apache 2.0. Pipeline fork code derived from diffusers (Apache 2.0).