Viggle Releases Qwen-Image-2.1-viggle-turbo v0.1 Preview

Viggle Releases Qwen-Image-2.1-viggle-turbo v0.1 Preview

At a Glance

Item Value
Repository Viggle/Qwen-Image-2.1-viggle-turbo
Published 2026-09-22
License other
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code at collection time. Dates are JST.

Overview

Viggle has released the v0.1 preview of “Viggle/Qwen-Image-2.1-viggle-turbo", a distilled model based on Qwen/Qwen-Image-2.1. This model is a “student" model trained using Distribution Matching Distillation (DMD), and supports both text-to-image generation and instruction-driven image editing using 1 to 3 reference images.

The standout feature is that while the base model requires 40 transformer passes, this model can complete generation in just 4 passes. It is also designed not to use Classifier-Free Guidance (CFG).

Specifications

The specifications listed in the documentation are as follows.

Model Architecture

The repository includes two types of models that can be chosen depending on the use case.

  • transformer/: A fully fine-tuned transformer. This is the currently recommended version as it directly replaces the base transformer and offers higher editing fidelity than LoRA.
  • Qwen-Image-2.1-viggle-turbo-4step-lora-r64.safetensors: A LoRA adapter (rank 64). Used by loading it on top of the base transformer at runtime. While the download size is smaller, it is reported to perform slightly worse compared to the fully fine-tuned version.

Inference Settings

  • Inference steps: 4 steps
  • CFG scale: 1.0 (CFG is not used)
  • Negative prompt: Not used

Output Resolution

The main resolution areas during training are as follows.

ratio 1024² area (T2I + editing) 1536² area (editing) 2048² area (T2I)
1:1 1024 × 1024 1536 × 1536 2048 × 2048
4:3 1184 × 896 1760 × 1344 2368 × 1760
3:4 896 × 1184 1344 × 1760 1760 × 2368
16:9 1376 × 768 2048 × 1152 2720 × 1536
9:16 768 × 1376 1152 × 2048 1536 × 2720
3:2 1248 × 832 1888 × 1248 2496 × 1664
2:3 832 × 1248 1248 × 1888 1664 × 2496

License

  • License: other

Performance

This model is currently a v0.1 preview version and is reported to be a work in progress. The publishers caution that the current weights should be treated as a preview rather than a complete replacement for the base model.

Quality Evaluation

According to the model card, 4-step text-to-image has reached a practical level. However, it is reported that in complex image editing tasks, the quality is visibly inferior compared to the 40-step base model in some scenarios.

Specifically, a drop in quality has been confirmed in the following cases:

  • Complex editing tasks such as composition using multiple reference images, face swapping, identity-preserving edits, or instructions containing multiple constraints.
  • In complex edits, duplicate people or ghost-like figures may appear.
  • Identity drift (a phenomenon where the subject’s features change) may occur when given instructions such as “keep everything the same".
  • Small or long rendered text may break more frequently than in the base model.

It is also specified that validation against the teacher model has not been performed for output at 2K resolution. Note that evaluations at this stage remain qualitative by the publishers, and no quantitative benchmark scores are provided.

Strengths and Use Cases

The model excels at text-to-image generation and image editing using 1 to 3 reference images. Particularly in instruction-based image editing, the order of the reference images controls which image corresponds to <image1> or <image2> in the prompt. Additionally, if the output aspect ratio is not specified, it follows the aspect ratio of the last reference image.

The primary areas used during training are 1024² and 2048² areas for text-to-image, and 1024² and 1536² areas for image editing. While it can operate at other sizes, much of the training data is based on these resolutions.

How It Differs from Similar Models

This model is a distilled model aiming to speed up the capabilities of the base model, Qwen-Image-2.1, through a significant reduction in steps (from 40 steps to 4 steps). While previously released quantized versions such as unsloth、画像生成モデル「Qwen-Image-2.1-FP8」公開 and unsloth、画像生成モデル「Qwen-Image-2.1」のGGUF版を公開 primarily target inference lightweighting and memory saving, Viggle’s model specializes in drastically shortening the generation process itself using DMD (Distribution Matching Distillation) technology.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 7.1B parameters (taken from the base model Qwen/Qwen-Image-2.1)

Your VRAM Quantization File size Est. memory needed
48GB (RTX 6000 Ada / A6000, etc.) BF16 30.8GB 37.0GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. This release is an adapter (LoRA etc.); the table shows what the base model Qwen/Qwen-Image-2.1 needs. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

How to Get It

The model is distributed in the following formats:

  • transformer/: Fully fine-tuned transformer (safetensors in bf16 format)
  • Qwen-Image-2.1-viggle-turbo-4step-lora-r64.safetensors: LoRA adapter (rank 64, safetensors in bf16 format)
  • peft/: PEFT format adapter (F32 format)

The repository is available on Hugging Face at Viggle/Qwen-Image-2.1-viggle-turbo.

Usage requires the diffusers library, but since QwenImage21Pipeline is not yet included in the official release version, installation using git+https://github.com/huggingface/diffusers.git@80c7ed262aeffbeb43ef13ae04baeb9b84515a69 specifying a specific commit hash is recommended. Additionally, the peft library is required when using LoRA.

Related Articles

Sources

Image, Video and Audiodiffusers,DMD,LoRA,Qwen-Image-2.1,Verified,Viggle,画像生成