<h1 id="pruna-qwen-image-2.1">Pruna-Qwen-Image-2.1</h1>Unofficial mirror. This listing mirrors the original project for reference and convenience. It is not created, maintained, or endorsed by Pruna AI. Please credit the original authors and preserve all license, attribution, and modification notices.Original project: <a target="_blank" rel="ugc" href="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1">PrunaAI/Pruna-Qwen-Image-2.1 on Hugging Face</a> Base model: <a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1">Qwen/Qwen-Image-2.1</a> Project/GitHub: <a target="_blank" rel="ugc" href="https://github.com/PrunaAI/pruna">Pruna AI / pruna</a><h2 id="overview">Overview</h2>Pruna-Qwen-Image-2.1 is a set of LoRA adapters for <a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1">Qwen-Image-2.1</a>. The adapters enable text-to-image generation and image editing in 5 or 8 steps while keeping the base pipeline, text encoder, and VAE unchanged. Training is based on DMD and improved using Qwen.<ul><li>5 or 8 inference steps</li><li>Up to 6.3x faster in the original benchmark</li><li>No CFG</li><li>Text-to-image and image editing</li><li>BF16, CUDA GPU required</li><li>LoRA strength: 1.0</li></ul><h2 id="variants">Variants</h2>Both adapters are v0.1. Use only one adapter at a time because each has its own sigma schedule.FileStepsTrade-off<code>p_qwen_image_2.1_8step_v0.1.safetensors</code>8Higher quality; recommended default.<code>p_qwen_image_2.1_5step_v0.1.safetensors</code>5Higher speed, with noticeably lower visual quality.<h2 id="visual-examples-from-the-original-model-card">Visual examples from the original model card</h2>These examples are hosted in the original Hugging Face repository and are included here with source attribution.<img src="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/resolve/main/assets/Pruna-Qwen-Image-2.1-t2i-1024.png" alt="Text-to-image 1024 example" /><img src="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/resolve/main/assets/Pruna-Qwen-Image-2.1-t2i-2048.png" alt="Text-to-image 2048 example" /><img src="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/resolve/main/assets/Pruna-Qwen-Image-2.1-i2i-1024.png" alt="Image editing 1024 example" /><img src="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/resolve/main/assets/Pruna-Qwen-Image-2.1-i2i-2048.png" alt="Image editing 2048 example" /><img src="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/resolve/main/assets/edit_grid.jpg" alt="Image editing comparison grid" /><img src="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/resolve/main/assets/latency_t2i_1k_2k.png" alt="Original H100 latency chart" /><h2 id="prompts-editing-and-resolution">Prompts, editing, and resolution</h2>The adapters were trained at 1K resolution with simple and upsampled prompts, text-to-image generation, and single- and multi-image editing with up to 3 reference images. Prompt upsampling is optional; detailed prompts usually work better.<ul><li>Start at 1024 x 1024.</li><li>Use at most 3 reference images for editing.</li><li>Higher resolutions, including 2K, and more reference images may work, but are outside the training coverage and quality may vary.</li></ul><h2 id="benchmarking">Benchmarking</h2>Original benchmark: official Qwen model-card example prompt, BF16, batch size 1, one NVIDIA H100 80GB, median of 3 requests after one warmup. It includes prompt encoding, denoising, and decoding, and excludes PNG saving, model loading, and warmup. These timings do not imply equal image quality.PipelineBase, 40 steps8-step LoRA5-step LoRAText-to-image, 102431.44 s7.60 s4.98 sImage-to-image, 10247.05 s2.01 s1.43 sPublished benchmark configuration: base with KV cache on; Pruna adapters with KV cache off; LoRAs unmerged; no CFG, compilation, or CPU offload.<h2 id="quickstart-(diffusers)">Quickstart (Diffusers)</h2><pre><code>pip install 'torch&gt;=2.4.0' 'transformers&gt;=5.17' accelerate peft pillow
pip install git+https://github.com/huggingface/diffusers@6256aa7666cedd47443adc8f82da9a10e110b09c

import torch
from PIL import Image
from diffusers import FlowMatchEulerDiscreteScheduler, QwenImage21Pipeline

STEPS = 8 # 8 for higher quality, 5 for higher speed
SIGMAS = {
 5: [1.0, 0.94, 6 / 7, 2 / 3, 0.4],
 8: [1.0, 14 / 15, 6 / 7, 10 / 13, 2 / 3, 6 / 11, 0.4, 2 / 9],
}[STEPS]

pipe = QwenImage21Pipeline.from_pretrained(
 "Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights(
 "PrunaAI/Pruna-Qwen-Image-2.1",
 weight_name=f"p_qwen_image_2.1_{STEPS}step_v0.1.safetensors",
)
pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(
 pipe.scheduler.config,
 use_dynamic_shifting=False,
 shift=1.0,
 shift_terminal=None,
)

# Text-to-image
image = pipe(
 prompt=(
 'A glowing neon shop sign that reads "QWEN IMAGE 2.1", mounted on a brick wall '
 'in a narrow city alley at night. Heavy rain, wet pavement reflecting pink and '
 'blue light, shallow depth of field, cinematic photograph.'
 ),
 width=1024,
 height=1024,
 generator=torch.Generator("cuda").manual_seed(42),
 num_inference_steps=STEPS,
 sigmas=SIGMAS,
 true_cfg_scale=1.0,
 use_kv_cache=True,
).images[0]

# Image editing
image = pipe(
 prompt="Change the background to a sunset beach",
 image=Image.open("input.png").convert("RGB"),
 generator=torch.Generator("cuda").manual_seed(42),
 num_inference_steps=STEPS,
 sigmas=SIGMAS,
 true_cfg_scale=1.0,
 use_kv_cache=True,
).images[0]</code></pre><h2 id="recommended-settings">Recommended settings</h2><ul><li>Use the sigma schedule matching the adapter; keep <code>shift=1.0</code> with dynamic shifting off.</li><li>8-step: <code>1 -&gt; 14/15 -&gt; 6/7 -&gt; 10/13 -&gt; 2/3 -&gt; 6/11 -&gt; 0.4 -&gt; 2/9 -&gt; 0</code>.</li><li>5-step: <code>1 -&gt; 0.94 -&gt; 6/7 -&gt; 2/3 -&gt; 0.4 -&gt; 0</code>.</li><li>No CFG: keep <code>true_cfg_scale=1.0</code> and do not pass a negative prompt.</li><li>Keep LoRA strength at <code>1.0</code>.</li><li>Use detailed prompts describing subject, setting, lighting, and style.</li></ul><h2 id="limitations">Limitations</h2><ul><li>This is a first version and quality is below the base model.</li><li>The 5-step adapter is faster but visibly lower quality than the 8-step adapter.</li><li>Short or vague text-to-image prompts give weaker results.</li><li>Other step counts, schedules, or CFG values are not supported.</li><li>This is not a standalone model; it requires the Qwen/Qwen-Image-2.1 base weights.</li><li>This is not a replacement for the base model when full quality is required.</li><li>This is not a finished release; weights may change in future versions.</li></ul><h2 id="license-and-attribution">License and attribution</h2>The original repository declares qwen-research / the Qwen Research License Agreement. This adapter is a derivative of Qwen-Image-2.1. Review the license before using or redistributing it, preserve the original <a target="_blank" rel="ugc" href="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/blob/main/LICENSE">LICENSE</a> and <a target="_blank" rel="ugc" href="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1/blob/main/NOTICE">NOTICE</a>, and retain attribution and modification notices. Qwen is Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.Mirror notice: This is an unofficial mirror. It is not created or endorsed by the original authors. No additional commercial rights are granted by this listing; do not sell this model or merges made from it unless the original license and all applicable permissions explicitly allow it.Source: <a target="_blank" rel="ugc" href="https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1">https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1</a>

Turbo - Few-step LoRA adapters - Pruna Qwen Image 2.1

p_qwen_image_2.1_5step_v0.1.safetensors

pruna 5 steps

pruna 8 steps

PrunaAI turbo LoRAs can be used for editing/text-to-image Current Viggle turbo LoRAs can be used for text-to-image Current Best settings for Viggle/PrunaAI 4-step: er_sde/beta res_2s_ode/bong_tangent, 5-8 steps, cfg 1-2 (+ lenovo/zjourney refiner lora to remove some of the overbaked, high-contrast, yellow tint) All examples are generated with 1.0 strength I did not train, distill, or create the original Turbo LoRA weights. This model page provides only modified compatibility versions intended to let the remaining compatible LoRA adapters load with ComfyUI's built-in LoRA loader Full credit for the original Qwen-Image-2.1-viggle-turbo-4step-lora-r64.safetensors belongs to the <a target="_blank" rel="ugc" href="https://huggingface.co/Viggle">ViggleAI</a> Team. Full credit for the original p_qwen_image_2.1_5/8step_v0.1.safetensors belongs to the <a target="_blank" rel="ugc" href="https://huggingface.co/PrunaAI">PrunaAI</a> Team. Viggle statement for v0.1: "Status: v0.1 preview, work in progress — this release still falls short of the base model. On complicated image editing (multi-reference composition, face swaps, identity-preserving edits, instructions with several constraints) it is clearly worse than the 40-step base model. Text-to-image at 4 steps is usable. We are still working on it and will update this repository as the distillation improves; treat the current weights as a preview, not a replacement for the base model." PrunAI statement for v0.1: "Pruna-Qwen-Image-2.1 is a set of LoRA adapters that let Qwen-Image-2.1 generate and edit images in only 5 or 8 steps. The adapters load on top of the base model, so the pipeline, text encoder, and VAE stay unchanged. Training is based on DMD. Improved using Qwen.v0.1: first release, work in progress. Pruna-Qwen-Image-2.1 does not yet match the visual quality of the base model. We are still improving the distillation and will update this repository as new versions become available."

Files with SHA256: 021a6228a0fcd217275190b89072416a2f28548e8e18fe049531dc9b77ef89bd