File size: 4,334 Bytes
aabfb2e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 | NOTICE
======
Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026
Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.
This repository is a derivative work of Qwen-Image-2.1 (https://huggingface.co/Qwen/Qwen-Image-2.1,
revision 790c92633540aa0cb11d9abf19eb46d861714758). It merges two derivative works of it into the base pipeline:
Viggle/Qwen-Image-2.1-viggle-turbo (v0.2.1) revision bb26a0f38e5fe6c124aaccc9187a87eed5d9ed13
madebyollin/texture-fix-vae-for-qwen-image-2.1 revision e9f84623d22c47f8bc9fb799bc54201fa53cf80b
The full agreement is in `LICENSE`, a copy of it is given to every recipient of these files.
Modified files, as required by section 2.b of the agreement:
transformer/*.safetensors the Qwen-Image-2.1 DiT with the Viggle turbo v0.2.1 LoRA (rank 256,
alpha 256, F32 factors from `peft_v0.2.1/`) merged into the 227
projection weights it targets; the sum is computed in fp32 and stored
in fp16. Every other tensor is the upstream bf16 value stored in fp16.
vae/ the Texture-Fix VAE by madebyollin (decoder fine-tune of the
Qwen-Image-2.1 VAE, encoder unchanged), fp32, byte-identical to its source
scheduler/ the Viggle turbo scheduler config (`shift_terminal` null instead of 0.02)
Unchanged from Qwen/Qwen-Image-2.1: text_encoder/, processor/, transformer/config.json,
model_index.json, LICENSE. Files added by this repository: README.md, media/.
Built with Qwen.
Upstream notices
================
--- Viggle/Qwen-Image-2.1-viggle-turbo ---
Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.
This repository is a derivative work of Qwen/Qwen-Image-2.1, produced by Viggle.
Built with Qwen.
It contains:
ADDED Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors - v0.2.1: a distilled LoRA adapter (rank 256, alpha 256) for the base transformer, sampled in 6 steps, stored bf16
ADDED peft_v0.2.1/ - the same v0.2.1 adapter in peft key format, F32 as trained
ADDED Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r128.safetensors - the v0.2.1 adapter truncated to rank 128 (alpha 128) by per-layer SVD, stored bf16
ADDED comfyui/ - ComfyUI custom nodes (viggle_turbo.py) and text-to-image / edit workflows, and the edit workflow's two example input photos (comfyui/input/); no model weights
ADDED Qwen-Image-2.1-viggle-turbo-v0.2-5step-lora-r256.safetensors - v0.2: the previous distilled LoRA adapter (rank 256, alpha 256), sampled in 5 or 6 steps, stored bf16
ADDED peft_v0.2/ - the same v0.2 adapter in peft key format, F32 as trained
ADDED Qwen-Image-2.1-viggle-turbo-v0.2-5step-lora-r128.safetensors - the v0.2 adapter truncated to rank 128 (alpha 128) by per-layer SVD, stored bf16
ADDED Qwen-Image-2.1-viggle-turbo-4step-lora-r64.safetensors - v0.1: a 4-step distilled LoRA adapter (rank 64, alpha 64), stored bf16
ADDED peft/ - the same v0.1 adapter in peft key format, F32 as trained
MODIFIED transformer/ - v0.1: the base transformer after a full-parameter 4-step distillation fine-tune, stored bf16
MODIFIED scheduler/scheduler_config.json - the base scheduler config with shift_terminal changed from 0.02 to null
No other file of the base model is copied or modified; processor/, text_encoder/ and vae/ are loaded from
Qwen/Qwen-Image-2.1 at runtime.
Source checkpoints: v0.2.1 LoRA - run v6_isg, step 700 (EMA student); v0.2 LoRA - run v6_isg, step 600 (EMA student); v0.1 transformer/ - run v4_anchor, step 400 (EMA student, full fine-tune); v0.1 LoRA - run v2_16gpu, step 400 (EMA student)
--- madebyollin/texture-fix-vae-for-qwen-image-2.1 ---
Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.
Modified files: diffusion_pytorch_model.safetensors and texture_fix_vae_for_qwen_image_2.1_bf16.safetensors contain the
Qwen-Image-2.1 VAE (https://huggingface.co/Qwen/Qwen-Image-2.1) with modified (finetuned) decoder weights.
The modifications were made by madebyollin in 2026. The encoder weights are unchanged.
|