---

license: other license_name: qwen-research license_link: LICENSE library_name: diffusers pipeline_tag: any-to-any base_model: Qwen/Qwen-Image-2.1 base_model_relation: finetune tags:

• tailsteer
• image-editing
• text-to-image

---

Qwen-Image-2.1 · TailSteer

Keep creating. Keep the details.

The generation and editing strengths of Qwen-Image-2.1, with greater stability across repeated edits. TailSteer is a final training stage that largely retains the base model's learned capabilities while reducing the drift that builds up when each edit becomes the next input.

Try the model · Single-turn results · Add & Preserve · Training data

+1.22 dB PSNR · 11.5% lower LPIPS outside the edit region at turn 10 on Add & Preserve.

Keep editing the scene. Keep the people.

Twenty successive stage edits put faces, clothing and background details under repeated generation. TailSteer reduces the accumulated texture drift while continuing to carry out the edits. Each branch feeds its own previous output into the next turn. View the final comparison · Prompts and frames.

Use your existing pipeline

Same architecture. Same official model layout. Same loading API. These are complete model weights: replace your original model ID or local path with Lyricccco/Qwen-Image-2.1-TailSteer. No adapters, merging, extra inference modules or TailSteer-specific runtime are required.

Use the same official environment as the base model. Only the model ID changes.

import torch
from PIL import Image
from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained(
    "Lyricccco/Qwen-Image-2.1-TailSteer", torch_dtype=torch.bfloat16,
).to("cuda")

edited = pipe(
    image=Image.open("input.png").convert("RGB"),
    prompt="Change the jacket to blue.",
    num_inference_steps=40, true_cfg_scale=1.0, use_kv_cache=True,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
edited.save("output.png")

For the next edit, pass the edited image back into the same pipeline. Use TailSteer throughout the original denoising trajectory; no model switching or preservation mask is needed at inference.

Keep the foundation model's strengths

Same prompt and seed in each pair. Full prompts and images.

The same checkpoint supports text-to-image generation and single-turn editing. Across Qwen-Image-Bench, BizGenEval, GEdit and ImgEdit, scores stay close to the base model. The table includes RISE as well, where accuracy is lower in this setting.

Benchmark	Base model	After TailSteer
Qwen-IB EN ↑	53.39	53.06
Qwen-IB ZH ↑	53.34	53.20
BizGen ES (%) ↑	57.03	57.38
GEdit EN ↑	8.11	8.04
GEdit ZH ↑	7.80	7.72
RISE accuracy (%) ↑	20.56	18.06
ImgEdit ↑	4.40	4.42

The same multi-image interface

A person and a pose guide remain ordinary inputs to Qwen. This sequence shows successive reference-guided edits with the base model and TailSteer. View the final comparison.

Measure what stays unchanged

Add & Preserve makes drift visible: add a small object inside the red box at each turn, retain previous objects, and preserve everything outside. It tests both object completion and background preservation across 100 images and ten turns.

Model	COA (%) ↑	PSNR@10 ↑	SSIM@10 ↑	LPIPS@10 ↓
Base model	67.36	16.93	0.438	0.270
After TailSteer	66.94	18.15	0.513	0.239

COA (Cumulative Object Accuracy) averages requested-object presence across ten turns. PSNR, SSIM and LPIPS measure the unchanged region. These are standard image-feedback results; curves use all 100 images. Metric data.

Run demos and benchmarks

With the TailSteer codebase installed in the official backbone environment:

tailsteer infer --backbone qwen-image-2.1 \
  --checkpoint Lyricccco/Qwen-Image-2.1-TailSteer \
  --policies tailsteer --inference-suite cafe-20turn --gpus 1

Prepare the original base weights and omit --policies tailsteer for side-by-side comparisons. The same suite interface runs text-to-image, single-turn editing, multi-turn editing and Add & Preserve benchmarks after their datasets and judges are configured.

How TailSteer works

Preserve early dynamics; steer the trajectory tail. TailSteer learns from the editor's own denoising rollouts, using unchanged source content as a preservation reference. It improves long-horizon stability using single-edit trajectories, without requiring curated ground-truth editing pairs or training on multi-turn editing chains.

This is the paper's 2,000-update checkpoint, with training steering onset 30 / 40. The onset controls training supervision; inference uses the trained model at every step. The packaged model was validated against the original inference export.

From TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing.

The base model license and attribution apply: LICENSE · NOTICE. Improved using Qwen.