wanx-troopers

Qwen

Qwen LoRAs

HF Space LoRA
HF:lilylilith/AnyPose 2511 AnyPose - Person in image 1 do the exactly same pose as image 2
HF:lightx2v/Qwen-Image-2512-Lightning ? Qwen-Image-Lightning v1.1 Lora
- Qwen-Image-Edit-2512 4-Steps Lora
- Wuli Turbo 4 Steps Lora
HF:prithivMLmods/Qwen-Image-Edit-2511-Unblur-Upscale Qwen-Image-Edit-2511-Unblur-Upscale
HF:vafipas663/Qwen-Edit-2509-Upscale-LoRA Qwen-Edit-2509-Upscale-LoRA

Note: 2512 is not an “edit” just “qwen image” but due to some sort of confusion we’re seeing the incorrect “Edit-2512” designation sometimes.

2026.01.13

Re QIE 2511:

if you want quality, skip the lora and go straight with full QE, res_2s + Beta57 (you may need the custom node RES4LYF to have these scheduler and sampler, use cfg 3.5, 30 steps, and you will probably be happy, all this with Qwen Image Edit 2511 obviously

2025.12.31

Qwen Image Edit 2512 released. bf16: HF:Comfy-Org/Qwen-Image_ComfyUI:split_files/diffusion_models.

2025.12.28

openpose with Qwen-Image-Edit-2509 … but also the anypose lora

Yeah the anypose lora is good and the most accurate, but with real people it changes their face and also body shape, which is the only problem.

I trained using ai-toolkit [a QIE-2511 LoRA] and didn’t like the results, and I tested a lot of things. Yesterday I was training with diffusion pipe, but there’s still some incorrect code regarding resolution matching, so I paused and will try to change the code later. Now I’m training using musubi-tuner and it seems to be training very well, and even on my 5090 without much difficulty.

2025.12.27 QIE Q4 + ZIT Low VRAM

niji-qui-zit by MadevilBeats aka “NiJi Dragon”; the shared a workflow 1st edits with a highly quantized version of QIE, like Q4 and then Z-Image-Turbo is used to clean up the artifacts

Q4 of QIE is a bit blotchy looking, kinda blurry, plasticy, etc. but it can still edit at roughly the same level as Q8 so I’m cleaning it up with ZIT. Mainly I just wanted to see if they could fit together on one GPU and how fast it would be in one workflow: yes and it’s less than 30 seconds typically [RTX5090]; ZIT upscaleAny

2025.12.26

Face/Head Swap by Alisson Pereira: civitai.com/models/2027766, patreon.com/posts/bfs-lora-qwen-140789769, youtube.com/watch?v=3CoK4pDl-TE.

If you get too much contrast / plastic skin using LightX2V Lightning, try steps = N/2 (e.g., 8 -> 4). Without Lightning -> lower steps (20 -> 16) and CFG ~ 1.2 - 1.5. If poses differ a lot -> increase LoRA strength (1.2 - 1.3).

Samplers: er_sde+beta57, kl_optimal, ddim_uniform, res_2s+beta57; Precision: best = fp16 | recommended = fp8/q8

Corza responded with face detailer qwen2509_seedvr2

2025.12.25

Phr00t’s mix of 2509 and 2511: HF:Phr00t/Qwen-Image-Edit-Rapid-AIO:v17

its a small mix handful of “realism” loras and “insubject” on the 2509 side of things, but just accelerators for 2511 side

euler_a/beta seems to work best; 4 steps minimum, generally 6 steps is worth it; grid artifacts varied across versions but havent been noticing them in v17

chat.qwen.ai for testing.

2025.12.24

Simple template workflow added to ComfyUI

cfg 6 is great on 2511 so far

EmptyQwenImageLayeredLatentImage turns QIE into T2I - start from no input image. Alternative advice: “you can just take QI (not QIE) ComfyUI template and just load QIE2511 model”

Alternative text encoders:

there’s no issue when regenerating the whole image; it’s editing in place, the elements that doesn’t change from the original image are degraded [contrast, saturation etc]

Suggestion for better LLM serving as a CLIP: from HF:silveroxides/Qwen2.5-VL-7B-MixedPrecision-ComfyUI

abliterated-mixed-fp8

2025.12.23

Highly anticipated Qwen Image Edit 2511 is out

Official workflow embedded in this image:

qie2511

qie2511

They seemed to have added a lot more color. The greens are popping and the color does not seem so washed out.

HF:dx8152/Qwen-Edit-2509-Multiple-angles camera angles LoRA works as before

Seems to improve realism of different angle LoRA HF:lividtm/Qwen-Edit-Loras:external by GH:filliptm. Note the author apparently has also got a character sheet LoRA - somewhere :)

Helpful node:

edit-model-reference-method

Ingi’s suggestion for node/sampler

TextEncodeQwenImageEditPlusAdvance-Irzjason

QwenEncodeQwenImageEdit is a simpler alternative from stock Comfy. reddit-published workflow.

Another artist:

1st k sampler with ddim beta57 and the second dpmppsde beta cfg 1; paradoxically … bypassing the lightning lora and 12 steps for each ksampler

LoRA “apply diff between img1 and img2 to img3”: link

2025.12.13

GH:gajjar4/ComfyUI-Qwen-Image-i2L so called “I2L”, “image to LoRA”, a super-lightweight approach to extracting a Qwen (Image Edit?) LoRA from an image; resulting LoRA is pretty weak

Qwen Edit Multiple angles LoRA good to create one image with views of the same character from different sides

Qwen Image Edit

Supposedly has affinity to Wan models since both come from Alibaba.

HF Space Model
Comfy-Org/Qwen-Image-Edit_ComfyUI split_files/diffusion_models/qwen_image_edit_2509_bf16
Comfy-Org/Qwen-Image_ComfyUI split_files/vae/qwen_image_vae
Comfy-Org/Qwen-Image_ComfyUI split_files/text_encoders/qwen_2.5_vl_7b
lightx2v/Qwen-Image-Lightning Qwen-Image-Edit-2509/Qwen-Image-Edit-2509-Lightning-8steps-V1.0-bf16
lightx2v/Qwen-Image-Lightning Qwen-Image-Edit-2509/Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16

Interesting: an attempt to rig a ControlNet on top of QIE 2509.

Notable Loras And Wf-s For Qwen Image Edit

Generates 1st frame for next scene in same location with same character. Sample workflow on HF next to LoRA.

HF Space LoRA
lovis93/next-scene-qwen-image-lora-2509 next-scene_lora_v1-3000.safetensors

Text editing similar to Nano Banana (dataset visualizer, dataset)

HF Space LoRA
eigen-ai-labs/eigen-banana-qwen-image-edit eigen-banana-qwen-image-edit-2509-fp16-lora.safetensors

https://civitai.com/articles/20190/headface-swap-workflow-qwen-image-edit-2509In

https://www.patreon.com/posts/head-swap-qwen-140536096

Qwen Piflow

PiFlow

Face Consistency With Qwen?

WithAnyone

PuLID ?

LoRA-s

This site is open source. Improve this page.