Workflow Qwen Image Edit Rapid AIO Workflow/Prompt discussion and guidance.

tramend7

tramend7

Lurker
Joined
Jun 2, 2026
Posts
8
Reaction score
0
Hi all, I've been experimenting with Phr00t's Qwen Image Edit Rapid AIO locally, and was looking to start a discussion on best workflows and prompts. I find it to be vary variable in terms of output and I haven't quite been able to nail the parameters to consistently maintain accuracy to the base image, and avoid mis-generations. The discussions on Hugging Face are relatively limited so I figured I'd try starting a thread here, in case any others are using this model and wanted to contribute.
A couple things I have found did help was using an image resizer to resize the second image when using 2 or more images. I've also found going below around 0.85 denoise consistently generates either pure static or very badly warped images. Going above 6 steps or increasing CFG by large increments also produces poor generations.
 
Hey, i've used Phr00t's AIO checkpoints extensively, and I've also been looking for prompt guidance. I'd use a system prompt (act as a img2img editing prompt NSFW generator using this input image) to feed into a relatively uncensored model like Grok, or an abliterated model that u run locally, and then using that prompt for the generation. What GPU are you using to run the model? I would recommend using this LORA, https://civitai.com/models/1939453/consistence-edit-lora?modelVersionId=2585269, at various strengths. It is great for improving adherence to base image. I'd stick with CFG 1.0, denoise at 1.0, and 4 steps, using euler_ancestral and beta scheduler. Are you using phR00t's workflow in comfyui?
 
I’m currently using my laptop lol so definitely limited GPU wise. I have a 3060ti but need to rebuild pc since cpu died. Thinking of getting 64gb of ram when I rebuild (probably the worst year in history to consider this tho…)
This is my usual prompt template:
“Use Image 1 as the base image to edit.
Preserve the subject’s identity and facial appearance exactly. Keep the same face, facial structure, facial features, eye shape, eyebrows, nose shape, lips, jawline, skin tone, hairstyle, hair color, hair length, hair texture, and facial expression. The face should remain clearly the same person and should not be redesigned or altered.
Preserve the subject’s body exactly. Keep the same body shape, body proportions, anatomy, pose, posture, limb placement, hand placement, leg placement, torso position, shoulder position, and overall silhouette. The body should remain unchanged unless the requested edit directly requires a small necessary adjustment.
Preserve the original composition and environment exactly. Keep the same camera angle, framing, perspective, lighting direction, background, room or environment, objects, scene layout, and overall image structure. The setting should remain unchanged unless the requested edit directly requires a small necessary adjustment.
Apply the following edit only:
[EDIT]
Make the edit look natural and realistic. Preserve realistic skin texture, realistic image texture, realistic fabric texture where relevant, realistic lighting, realistic shadows, and fine detail. The final image should look like the same original image, with only the requested change applied.
If any instruction conflicts, preserve the original identity, pose, and environment while applying only the requested edit. Change only what is necessary for the edit and leave everything else unchanged.”
But I tend to do more simplistic prompts. I have a fairly high error rate, but I have found that certain input images produce errors quite often when I re-run whereas others don’t. Also my hardware sucks which is probably another cause.
Agreed on 4 step, 1CFG, and 1 Denoise.
Workflow is as phroot recommends, I just added a node to resize image2/3 to the same size as image 1.
 
the best model to ever exist when it comes to sex pose, none other comes close
spilling some sauce here, find a way u can add multiangle node in ur workflow, magic will happen
 
Will give that a go. I'm guessing you're not the type to just share your workflow?
 
if u serious, ask for claude 1st, workflow is just beginning, there s also learning curve with prompting
 
Any tips on face swaps, and keeping originals facial expressions on generated export? Or body swaps?
Swapping in Qwen Edit tips appreciated.
 
A tip for prompting i2i for all kinds of qwen: be specific.
It s the biggest learning curve but with chatbot like grok it s much easier as u can reverse the process by asking it to write prompt for sex pose reference image u upload:

example: titfuck for pov

Dark-skinned hands of pov reach from low sides to her breasts, his hands prominently push her huge breasts together against his penis shaft . His penis shaft is fully hidden into her cleavage. The girl's breasts are pressed tightly around the penis shaft. his penis shaft is sandwiched between her soft breasts. The huge breasts are squeezed tight, completely surrounding the shaft on both sides. her breasts are mashed together so tightly that the breasts skin is flattened around the shaft, creating a visible deformation in the breast shape, The shaft is completely hidden from view, with only the head of his penis visible at her upper chest. The direction of his penis head is clearly upward and forward.

tongue kiss

he passionately tongue kiss her, his lips completely fused to girl's mouth, His mouth passionately swallowing her lips. his lips presses fully over her mouth,

Basically u have to rewrite the prompt sentence with different wording so the pose will be accurately shown in output
 
I tried this, with instructions for retaining face and body etc, but I found that with longer prompts it sometimes just ignored the whole prompt, have you had this? Also have you experimented with different tones when prompting? I tend towards more instructional and short sentences, I'd describe your prompt above as "narrative", as you more or less just describe the scene. Unsure what is better.
 
i never had problem with narrative prompt, i find it easier since u can copy paste from grok chat, here try out this basic aio workflow:

 
What's the samplingauraflow and cfgnorm doing? How did you calibrate those? They do seem to have made it more consistent.

Also general question - I find if my desired output size deviates too much from the image input then the output can get really variable. Either total nonsense output or it ignores the prompt entirely, I guess that's why you have the resizers?
 
i no longer use this wf, but ask claude for tweaking

rule of all successful i2i sex pose generation is u need a clear full body shot of the female subject
 
I find that using SAM3 masking is better in keeping character consistency when removing clothes. Without it, some parts of the body or the posture always changes.
 
been tweaking with krea2 for a while now but qwen still remain undefeated to me when it comes to consistency and spatial interaction
 
Do anyone knows if it's possible to create choking images with it? I would send an example but I don't know if it's against the rules
 
You must log in or register to reply here.
Back
Top