A Qwen Image 2.1 studio workflow. One model load, then four jobs you turn on one at a time: text to image, single-image edit, multi-image edit, and a pose editor. Each job can upscale with SeedVR2 and save a PNG with Civitai metadata.
## What you can do
- Text to image — prompt in, image out, then an optional SeedVR2 upscale.
- Single image edit — change a photo from a prompt. A mask is optional. Use one when the area is small or when several similar objects are in the frame. Put the photo on image 1 and the mask on image 2, and say <image1> and <image2> in the prompt (no space). Tell it to change only the masked region, or the edit spreads over the whole picture.
- Multi-image edit — several reference images in one instruction.
- Pose editor — VNCCS Pose Studio sets the pose, your photo sets the person. Image 1 is the pose render. Image 2 is the character photo. The positive prompt is Draw character from image2. Pose Studio’s view size and the sampler size stay the same (1024×1024). Keep the mannequin’s muscle, weight, and height near the middle.
Use the group bypasser and leave only one job enabled.
## Settings
- Steps 8, CFG 1, denoise 1, with the 8-step LoRA on. That LoRA is built for CFG 1. Higher CFG fights the pose and breaks hands, arms, and legs.
- Generate at 2MP or higher. Use the Resolution Master presets. The Krea 2 Turbo set is a good default. The name is only a label.
- Pose LoRA VNCCS_QI2_PoseStudioV1.1 stays at strength 1, in front of the 8-step LoRA.
- SeedVR2 tile size stays at 1024. Tile count 2 is a normal 2× upscale. Higher counts run one full pass per tile. On a 16 GB GPU, use the 3B SeedVR2 model and leave Cache Model on. Keep Manual Load bypassed unless you mean to use it.
## Saving
Each job has a normal save and a Civitai save. The Civitai save writes the prompt, seed, steps, CFG, sampler, scheduler, and model name into the PNG. Sampler settings for each job live on that job’s sampler settings node.
## You will need
- UNET: qwen_image_2.1_int8_convrot
- VAE: qwen_image_2.1_vae_bf16
- Text encoders used in the graph: qwen3vl_8b_nvfp4_heretic, qwen3.5_9b_qwen_image_2.1_pe_t2i, qwen3.5_9b_qwen_image_2.1_pe_i2i
- LoRAs: p_qwen_image_2.1_8step_v0.1, VNCCS_QI2_PoseStudioV1.1
- A SeedVR2 upscale model if you use that stage
Tested on an RTX 5070 Ti 16 GB.
Description
Changelogs:
Removed a few leftover nodes.
Fixed some node connections.
Fixed a few small typos.
Rearranged some groups to make the layout cleaner.
Separated the Upscale phase from the Text-to-Image group.
Reworked the SeedVR2 Manual Load section.
Added an easy way to download the models i used.
Added LoRA support for the Single Image Edit and Multi-Image Edit groups.
Added the "Speedup" nodes to the Multi-Image Edit group.
FAQ
Comments (6)
this workfolw is so good. thanks for good workflow, and what is diffrent betweean 0.1 vs 0.2 ?
I’ve made some cleanup changes and organized things more clearly. You can find a detailed report in the “About this version” section.
I just discovered that Qwen Image 2.1 also works surprisingly well as an upscaler / photo restorer in this workflow.
Use the Single Image Edit group. Load the photo there and try a prompt like this:
---
Restore this old photograph. Remove dust, scratches, stains, fading, and grain. Improve clarity and natural detail. Keep faces, clothing, pose, lighting, and composition unchanged. Do not beautify or modernize.
---
To change output size, open the Prompt Generator subgraph. On Text Encode Qwen Image. Resolution is 0 by default.
- Under ~500×500: I set 1024
- Larger scans: I set 2048
Notes:
- Width and height should be divisible by 32. (can work with weird resolution, but it's not optimal.)
- Tiny details can still break.
- Pose and face can drift a little. If that happens, turn PE off and/or disable the 8-step LoRA.
Not a dedicated restore pipeline, but it works surprisingly well.
Final score: 8/10 ~ Fun to use!
P.S - If the image resolution is too low even for SpeeVR2, try this version instead.
(Or try both at once…)
I'm working on a complete rework of this workflow to make it way simple to use and understand, it should be ready by next week.
(i don't have much time on my hands...)
IMPORTANT CHANGE, MUST READ!
Version 1.0 is built and tested only for the new TURBO version of Qwen Image 2.1 (link inside the workflow). I cannot confirm that the older Qwen Image 2.1 model still works with this file. If you still use the old model, keep the previous workflow.
What is different for you:
- You pick a mode and press Queue. Text to image, one photo, several photos, or pose. The workflow runs only that mode. You do not bypass groups anymore.
- The canvas is much smaller. The old file had four copies of almost the same setup. This one has a single path: Start, Models, LoRAs, Resolution, Prompt, Queue, SeedVR2, Save.
- A lot of extra custom nodes are gone. The speedup nodes, the group bypasser, the "send this everywhere" nodes, and the extra save nodes are not in this version. What is still needed is listed in the workflow notes.
- The 8-step LoRA is no longer used. Turbo already includes that speed.
- There is one prompt box. In pose mode the pose prompt is filled in for you, and the pose LoRA turns on by itself.
- SeedVR2 is a simple menu: Off, On, or upscale a picture you already have. You do not bypass the generation group for that.
- A light film grain was added. It runs at the end, including when SeedVR2 is off.
- Saving is one node. Civitai info can be written if you add the small optional helper from the notes.
The prompt enhancer is still optional. If you do not use it, you can leave those two extra text models off. Everything else stays on.
(Workflow available in a few hours)