Hi friends, IAMCCS here — and today I’m releasing two new workflows for Qwen Image Edit 2509, both deeply integrated with the Prompt Enhancer and the upgraded IAMCCS_nodes tools.
This post expands and updates my previous posts and pushes it into a fully working release with:
✓ A Face Swap Workflow (v1.0.1)
✓ A Qwen Edit Workflow with LoRA support for Nunchaku low VRAM users (v2.0)
Let’s break them down.
Some words about Qwen-VL
Yeah, the “VL” stands for Vision-Language, but what it really means is:
a mini-brain inside ComfyUI that can read your images, understand them, and talk back.
And when you combine that with the IAMCCS_QE_Prompt_Enhancer and a properly structured face-swap pipeline… you basically get a self-aware assistant that writes the perfect prompt for you, tuned exactly to the face you want swapped.
Let me show you the whole thing.
What Qwen-VL Really Is (and why you want it)
Qwen-VL is a multimodal model by Alibaba capable of:
✓ Understanding the content of an image (characters, faces, clothing, composition, mood)
✓ Describing the image in natural language (a structured prompt-like output)
✓ Extracting semantic attributes (pose, lighting, gender, age, environment, style…)
✓ Answering visual questions (“what is the person wearing?”, “where is the light coming from?”)
✓ Generating “prompt seeds” that can be fused with your custom Prompt Enhancer instructions
In other words: Qwen-VL is your prompt ghostwriter.
It sees the reference image, describes it, and hands the description to your AI pipeline as a ready-made semantic block. This alone is a game-changer for face swap consistency.
How to Download Qwen-VL
Here’s the model I recommend (VRAM-friendly, fast, perfect for our workflow):
Qwen3-VL-8B-Instruct (4-bit)
You can grab it directly from the ComfyUI-QwenVL repo or pull it through ComfyUI Manager:
After restarting ComfyUI, the node AILab_QwenVL will appear.
Choose Qwen3-VL-8B-Instruct + 4-bit mode if you’re under 12GB VRAM.
Here suggestions from the repo:
Why Qwen-VL Matters in a Face-Swap Workflow
Because no face-swap model can guess what you want unless you tell it every time.
Qwen-VL solves this by:
→ Analyzing the face reference
→ Producing a perfect textual description
→ Passing that description into the IAMCCS Prompt Enhancer
→ The enhancer fuses it with your preset instructions
The result?
A combined prompt that keeps identity consistent even in complex swaps.
You get:
-
fewer generations ruined by wrong hair
-
better skin tone matching
-
pose-aware consistency
-
and superior “semantic compatibility” between source and target images
⭐ Workflow 1 — Qwen-VL + Prompt Enhancer Face Swap (v1.0.1)
Load Body Image → Auto Crop the Face
We use AutoCropFaces to isolate the face region of the target body.
This gives us a clean face bounding box used for the swap.
Load Reference Face
This is the identity you want to impose on the target body.
Qwen-VL Reads the Body Image
The Qwen-VL node generates an intelligent description of:
This text is automatically systematized, not chaotic like standard captioning.
The Prompt Enhancer Takes Over
Your IAMCCS_QE_Prompt_Enhancer receives two strings:
A) the semantic description from Qwen-VL
B) your preset (Swap Face, Maintain Consistency, Keep Lighting, etc.)
The enhancer merges these two into a structured multi-line prompt with optional toggles:
This makes the semantic conditioning ultra-solid.
VAE Decode → Final Result
And then you get your cinematic face-swapped output.
Why This Workflow Hits Hard
Because you’re stacking three intelligence layers:
1. Qwen-VL — “I see what’s in the image.”
2. Prompt Enhancer — “I convert this to structured command logic.”
3. Qwen Image Edit (DiT) — “I execute with high fidelity.”
Together they behave like a director, a writer, and a camera operator.
You sketch nothing.
You type almost nothing.
And still get results that look intentional.
⭐ Workflow 2 — Qwen Image Edit 2509 + LoRA Support (Low VRAM Fix)
This is the workflow built around the BRAND NEW:
➜ IAMCCS Qwen Image LoRA Loader FIX 4 Nunchaku (included in IAMCCS_nodes v1.3.0)
🔧 What you can do with this workflow
Add LoRAs to Nunchaku Qwen Image Edit reliably (even on a 3060, 8–12GB cards, older CPUs) and Merge LoRA styles with multi-image Qwen operations (pose transfer, background merge, relight…)
Use the Camera Angle presets from Prompt Enhancer 1.0.1 and choose your prompt.
Check my previous post for deeper instructions about the Nunchaku pipeline:
Preset: camera angles
Prompt: High angle
Second example.
Here’s a quick sketch I drew (9 sec timelapse).
I ran it through the workflow using:
STYLE EFFECT PRESET – PHOTOREALISTIC PROMPT
This is the result.
Handsome guy, isnt’it?
🔧 Download My Custom Nodes
Check my previous post for info about the new versions:
https://www.patreon.com/posts/update-iamccs-v1-143940470?utm_medium=clipboard_copy&utm_source=copyLink&utm_campaign=postshare_creator&utm_content=join_link
IAMCCS_nodes (new version 1.3.0!)
https://github.com/IAMCCS/IAMCCS_nodes.git
IAMCCS Prompt Enhancer (NEW VERSION 1.0.1!)
https://github.com/IAMCCS/IAMCCS_QE_PROMPT_ENHANCER.git
IAMCCS Annotate
https://github.com/IAMCCS/IAMCCS_annotate.git
Use Annotate inside the workflow to sketch areas, notes, expected direction, etc.
Your future you will thank you.
Attached: the workflows and two Pexels images to warm you up!
If you use these workflows, show me your results — I’d love to see what you create.
More coming soon ❤️