image21-turbo-4step-lora — project status

No trained LoRA is published here yet. The project is paused at quality-gate G3-FAIL.

This repo was created for a rank-64 LoRA that would make Qwen/Qwen-Image-2.1-Turbo produce 4-step images matching its own 8-step output. The plumbing gates all passed (CPU tiny-model checks, push preflight, bit-exact sampler parity: max|Δ| = 0.0 between a hand-rolled 8-step Euler loop and pipe(...)), but the distillation quality gate failed at smoke scale — the trained student stayed slightly further from the teacher than naive Turbo@4, at every prompt count, step count, learning rate and loss weighting tested. Full evidence, timings, root causes found and fixed, and next steps are documented in ysharma/image21-turbo-trajectories.

Change notice (for the eventual deliverable)

This derivative of Qwen-Image-2.1-Turbo is subject to the Qwen Research Licence (non-commercial). It keeps the licence text, adds this change notice, and is not named with "Qwen" as the main repo name. Built with Qwen.

Intended usage (once the gate passes)

import torch
from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1-Turbo", torch_dtype=torch.bfloat16).to("cuda")
pipe.load_lora_weights("ysharma/image21-turbo-4step-lora")
image = pipe(
    prompt='A neon shop sign that reads "QWEN 2.1", rainy night',
    sigmas=[1.0, 0.95418, 0.89508, 0.704534],   # 4 steps, replacing Turbo's 8
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]

(This snippet has not been executed against a trained adapter — the adapter does not exist yet.)