Hello!
Version 2511 v1:
This workflow works fine with 8GB VRAM, if you have a better card, you might want to disable the Lightning LoRA, add more steps, and cfg to every KSampler. Maybe 20 steps and cfg 2.5? You should try. Unlike the previous versions of this workflow, this has a custom node and uses a Q version.
Resources:
[Model] Qwen-image-edit-2511-Q2_K.gguf:
[CLIP] qwen_2.5_vl_7b_fp8_scaled
[VAE] qwen_image_vae
[Nodes] Comfyui-QwenEditUtils
About the prompts:
These prompts are just examples, so you might want to change them, because, depending on the original image they may work or may not work. Also, I totally recommend translating the prompts to Simplified Chinese, because in that language they will have more adherence. For example: First prompt is orbit the camera 45 degrees. Second prompt is right-side. Third prompt is zoom out. You can use English too, but, you'll need better luck.
_____________________________
Previous version text:
Hello!
People asked me to update this workflow, so, here it is! This version uses:
And 8 steps LoRA:
Also, I added a secondary image, if you want to add something to your new images, if you don't want, you can disable the load node. Remember always to change the prompts to do what you need.
The file you'll download has the JSON file as well as a PNG, so you can use either in ComfyUI
Cya!
_____________________________
Previous version (1.0):
Hello!
I'm kinda new to ComfyUI, so I made this spaghetti nightmare, this will create 6 different images from one image using Qwen Image Edit with 4 steps each one (I'm low VRAM!).
Files used:
Model: qwen_image_edit_fp8_e4m3fn
Lora: Qwen-Image-Lightning-4steps-V1.0
Clip: qwen_2.5_vl_7b_fp8_scaled
Vae: qwen_image_vae
Have fun!
Description
FAQ
Comments (38)
I just can't get qwen to generate images this high quality. Did you generate the original image in flux?
Yup, I made a lot of random profile pics for testing, they are mostly Flux Dev 1 with, if I'm not wrong, either Amateur Photo 6 (LORA), or Flux Skin Texture 2 (LORA), the variations are Qwen Image Edit, I don't use Qwen Image because it's so slow here. I rather use a Chroma + Pony I made a week ago, it's fast and kinda nice.
Also, this is a relatively clean workflow, compared to a lot of what you see on here. A nice one task workflow is best, AIO workflows are overrated.
The error "TextEncodeQwenImageEdit mat1 and mat2 shapes cannot be multiplied (5376x1280 and 3840x1280)" occurs in "Negative 6". Do you know anything about this?
Nope, but I see someone had that error, here, check if it helps: https://github.com/comfyanonymous/ComfyUI/issues/9423
@gabrielx Got it, thanks! I’ll check the link.
Qwen CLIP GGUF error, get the fp8 clip
You need to use a clip that has been "scaled." I ran into that too, and that's what fixed it.
Damn, I'm new to comfyui and it's very rare to get a workflow running the first time for me. But you did it! Thanks.
You're welcome! I'm actually new to ComfyUI too, just ditched ForgeUI, because it's been a while since it's been updated, sadly.
Fantastic resource, thanks so much for sharing 💛💛
Is there an easy way to force Qwen to change the character positions (not just the camera angle)? Sometimes it stubbornly refuses to modify the composition even if I prompt it heavily.
I've been testing on that by changing the noise and cfg in the ksamplers, but Qwen Image Edit kinda love sticking to the original most of the time.
This is one the best workflows running out of the box. No fancy nodes, still working like a charm. Thanks for sharing! Runs super fast on rtx4070
Best workflow! Simple and powerful!
Thanks! :D
You're new to comfyUI and you create a workflow this good? In that case, welcome to comfyUI! ffs. Also, I'm going to try this on my commercial gpu (I rent on demand gpus) and see how quick the gen time is without the lightning lora.
Thanks! Tell me how it goes! ☺️
@gabrielx Well, I completely reworked the workflow. lol.
@gabrielx Here's a link to the updated workflow, the biggest change is the power lora loader, and the ability to bypass generation stations, also I think I may have added the unified seed generator, I think I also unified the steps on the individual generation stations but you may have actually already did that. Can't remember. Also set everything up so you can see all six photos as you're working on your prompts. Well, its not letting me put the link in the comment, YAY CIVIT. Its on my profile page.
It is great workflow and insanely smart idea! May I suggest some upgrades for future version: 1. Probably it will be better to change LORA to qwen image edit lora 8 steps (created specifically for qwen image edit model) 2. changing text nodes to image edit model specific text nodes (with image input). Thanks a lot, keep up the good work!
Hi, I actually learned some days ago a couple of things I can do better, I'll try for version 2!
Nice work mate. I only replaced your qwen-image lightning lora with qwen-image-edit-lightning 8 steps. Worked just fine. Waiting for V2 with updates :)
I have also adjusted:
1. Text nodes to TextEncodeQwenImageEdit (negative+positive) [thanks to @sianosianz]
2. Replaced Scale Image to Total Pixels with Flux Resolution Calc + EmptySD3LatentImage for more granular control of the output [saw that in @Danrisi ultra real workflow]
now it's perfect :>
You just saved me (checks workflow connections)...
An eternity of frustration! Thank you! Where is that tip button!?
Slight correction and continuing accolades to you!! Prompt 6 was returning depth map occasionally:
This will fix it for anyone who needs a fix
(Low-angle view):(1.2)
I often find that when working with characters, it seems consistent and doesn't have any problems. Have you ever tried it with spaces like rooms or landscapes? I find it not very good in landscapes, right? :<<<
Ya, it's not good in rooms and stuff, for example, if your character is sitting in a room, and you want a profile of that character sitting, what it will do is moving the character's face to a profile view, but the rest will keep almost the same.
@gabrielx ohhh nice , Thank u <3
You are the 9th incarnation of the Buddha. Thank you for this.
Looks like a really great workflow! But he is not generating any new images. Output is just the same than the input picture. Sometimes he is generating pormpt3 "Obtain the back-side" bot not the other 5 view types. Any ideas?
Tried out a bit with the settings and changed the prompts etc. But now he is doing nothing anymore. even if i set it back to your stock settings. Something is going wrong here.
Logs says:
---------------------------------------------------------------------
got prompt
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Requested to load WanVAE
loaded completely 2270.0625 242.02829551696777 True
Using scaled fp8: fp8 matrix mult: False, scale input: False
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load QwenImageTEModel_
loaded completely 13091.646704483031 7909.737449645996 True
loaded completely 13091.646704483031 7909.737449645996 True
model weight dtype torch.float8_e4m3fn, manual cast: torch.bfloat16
model_type FLUX
Requested to load QwenImage
loaded partially 12984.746980926513 12982.860900878906 540
100%|██████████| 4/4 [00:23<00:00, 5.85s/it]
Requested to load WanVAE
---------------------------------------------------------------------
So why Clip text encoder is running on cpu instead of gpu?
And why he is detecting the model as "Flux" instead of "Qwen"?
I will run some tests, I think the lastest Comfy update did something.
I wonder if we could get an updated workflow for Qwen Image Edit Plus (2509). This workflow in particular has been incredibly useful to me, so thank you for your hard work 🙏