Open: How much is it going to cost to take a picture
Let's say you're gonna shoot a set of scripts。
the french retrospect package was selected, with three hours of pose in the studio, a morning off makeup, and the final selection — 20 fine-tuning and $2,800。
And then you might spit, "It's not like you're done with your face, but it's not good for you, and it's better for you than I am."
It's not bad, but it's not worth it。
Where is it? Not the photo itself, the whole set..Sites, lights, makeup, clothing, aesthetics and later stages of photographersI don't know. Most of the money you spend buys this combination of punches, instead of pressing that 0.01 seconds。
But if I tell you, all you need is a face selfie, plusComfyUIAnd the Z-Image-Turbo model of Ali's Open Source allows you to put your face on any kind of picture — Sabbonk, old-fashioned clothes, French retrogression, solar fresh, whatever you want。
It's not like thatAI face-changing“It is the extent to which it is more natural to integrate and to send out friends directly。
That's $2,800. It's free。
No bullshit, straight to dry。
I. Why Z-Image-Turbo
One might ask, "How can you choose Z-Image-Turbo
Three reasons, hard:
It's a "man like a man."
Z-Image-Turbo (known as ZIT) is the 6B parameter used to model the November 2025 open source of the Ali Tunyi LaboratoryIt's a real figureI don't know. The official phrase is "very good at producing images of high-security people, natural skin texture and photo-effects". Use it as a bottom map, right。
It's two years old
ZIT has an All-in-One (AIO) version of the text encoder and VAE packaged in a single document. Quantified version of FP8 only 8G VISIBILITY You can run, RTX 3060, 4060 and take off. By contrast, Flux stood still at the start of the 12G, and the threshold was short。
3 ️⃣ Apache 2.0, completely free of charge
Don't underestimate this. Many model licences contain restrictions, and Apache 2.0 means that you can use them, and business projects are not afraid. Is it really a little red book deal? No problem。
IN ADDITION, ZIT SUPPORTS A TWO-LANGUAGE ENGLISH-CHINESE TRANSLATION, AN EIGHT-STEP HYPER-SPEED RUN (RTX 4090, 2 SECONDS) THAT IS COMBINED TO BE ONE OF THE MOST COST-EFFECTIVE PRODUCTION MODELS。
II. HOW DOES THE AI FACE CHANGE WORK
Let's get this straight, or you'll get a fog in the back。
AI's face is in ComfyUI, and the core has two ideas:
I: Reactor route
YOU START WITH A HIGH-QUALITY TEMPLATE WITH ZIT -- LIKE A WOMAN IN A FLAG ROBE STANDING IN THE SUZHOU GARDEN. THEN WE'LL USE ONE Reactor The tool to replace the face of the person on the template with your face。
The advantage of that is to be simple and rude and to change it in a few seconds. The disadvantage is that “face-to-face” is sometimes more visible and that there may be traces on the edge。
Ideas two: Train your face to produce it
You train one with your own pictures LoRALET ZIT REMEMBER YOUR FACE. AND EVERY TIME YOU PRODUCE A PICTURE, THE MODEL DRAWS YOUR FACE, AND YOU DON'T NEED A LATER REPLACEMENT。
The advantage of this idea is that it is extremely natural to integrate -- because your face is "long" in the picture, not "posted" on it. The disadvantage is that it takes time to train (approximately 1-2 hours) and to prepare training materials。
What about the two routes? Let me get this straight:
| comparison dimension | ZIT+Reactor | ZIT + LoRA (direct generation) |
|---|---|---|
| Rationale | Sir, make a template and replace the face | Train your face, draw it when it's generated |
| Operation difficulty | It's so simple. Just a node | I need to train Lora |
| Figure Speed | Quick, 8-step + seconds to change face | Like normal life maps, 8 paces 2-5 seconds |
| Face resemblance | ♪ Medium, like but not exactly like ♪ | It's you |
| Shadows merge | It's normal. Occasionally, there's an edge | ♪ Perfectly integrated, no seams ♪ |
| Preliminary preparation | Download the model and you can use it | We need to prepare 10-20 selfies for 1-2 hours |
| Suitable for the scene | Quick-out. Send a circle of friends | I'm looking for quality |
| Visible needs | 8G+ (AIO VERSION) | THE TRAINING REQUIRES 12 G+ |
Think of three seconds to play first - choose Reactor routeI don't know. It's like you're really doing it LoRA route.
It's both courses。 Let's start with a simple Reactor and then go to LoRA's evolution。
III. Nanny course (up): ZIT + Reality quick changeFace
STEP 1: DOWNLOAD THE ZIT MODEL
ZIT HAS TWO KINDS OF INSTALLATION, I RECOMMEND AIO VERSIONSave it。
AIO VERSION (RECOMMENDED NEWCOMER):
Download z-image-turbo-fp8-aio.safetensors (approximately 10GB) for CommyUI/models/checkpoints/
- Downloading address: HuggingFace search SeeSeeSee21/Z-Image-Turbo-AIO, or mirror site hf-mirror.com
The version packed the text encoder (Qwen3-4B) and VAE in full, one file was completed and no other components were downloaded separately。
Partition version (suitable for a veteran):
| Documentation | Which directory | corresponds English -ity, -ism, -ization |
|---|---|---|
| x_image_turbo_bf16.safetensors | Photo by Flickr user ComfyUI/models/diffusion_models/ | Core Generation Model |
| qwen_3_4b.safetensors | Photo by Flickr user ComfyUI/models/text_encoders/ | Text encoder |
| ae.safetensors | Photo by CafyUI/Models/vae/ | Image decoded, shared with Flux |
- THE ADVANTAGE OF THE SPLIT VERSION IS THAT IT CAN BE FLEXIBLELY MATCHED, FOR EXAMPLE, THE TEXT ENCODER CAN BE USED TO SAVE FURTHER VISIBILITY IN A QUANTITATIVE VERSION OF THE GGUF。
Step 2: Installation of face change points
Open the CommyUI, make sure you're loaded I'm sorry(Unloaded first, this is ComfyUI's "appliance store"。
Search and install in Manager:
| Node Pack Name | corresponds English -ity, -ism, -ization |
|---|---|
| CCfyUI-Reactor-Node | The face change core |
| ComfyUI-InsightFace(usually automatically loaded with Reactor) | Face check and characterization |
Loading up to restart ComfyUI。
Step three: Download the face change model
Reactor needs a face change model file:
inswapper_128.onnx
Download channel:
- GitHub Search inswapper_128Confirm .onnx specification
- Or go to Hugging Face InsightFaceGot it inswapper_128.onnx
Download after to:
I'm sorry, but I'm sorry
- No insightface folder? Just build one of your own, ComfyUI will automatically recognize。
Step 4: Working-streaming
The logic of this stream is:Mr. ZIT made a template.
Plugin ZIT to generate templates - Reactor Face Swap - Final Fig
↑
Your self-censorship — Load Image — — — — — — Peter —
Specific operations:
Start with a ZIT Basic Map Workstream (loading the AIO model with a CHeckpoint loader)
ZIT PARAMETER SETTINGS:
| parameter | Recommended value | clarification |
|---|---|---|
| Steps | 8 | ZIT Turbo distillation model, eight steps enough |
| CFG | 0 (turbo mode) | CLOSE CFG GUIDE |
| Sampler | res_multiistep | ZIT SPECIAL SAMPLER |
| Scheduler | simple | Official recommendations |
| Resolution | 1024x1024 or 1920x1088 | The vertical map suits the image |
Generates a satisfactory template and then goes back to Reactor's face change point
increase Load Image Node, upload your face
Connect ZIT-generated templates to Reactor input_image"The image of the face changed"
Connecting a selfie to Reactor you know, source_image(Provision of face)
Put Reactor's output_image It's connected Save Image
Wait, it's easy here. Speak clearly:
- input_image: THE ZIT-GENERATED MAP YOU'RE GONNA HAVE TO CHANGE YOUR FACE
- you know, source_image: The image of your face
Don't turn it upside down, but instead, it's scary to put the face of the guy on your own。
After we're connected, dot Queue PromptA few seconds to figure out。
- 💡 draw attention to sth.Reactor automatically detects the largest human face replacements in the two maps. When there are many people in the template, it changes the biggest face. You want someone else? The Face_index parameter was found in the Reactor node, the first face and the second, and so on。
iv. How does it work? Five practical techniques
It's not enough to get a job and make a good movie
1. Source map (self-generated) mass is ground-based
Whether Reactor changed his face or Lora trained, the photos you provided directly determined the maximum effect. Several points:
- Front or microside (within 15 degrees)
- The light is even. The other half is in the shadows
- Don't exaggerate
- The resolution is at least 512 x 512 and the clearer the better
- Don't wear masks, sunglasses, and Liu Hai covering his eyebrows
2. The light orientation of the template map is consistent with the source map (Reactor route)
YOUR SELF-PORTRAIT IS LEFT-SIDE LIGHT, AND THE ZIT-GENERATED TEMPLATE IS RIGHT-SIDE LIGHT, WHICH DOESN'T MATCH THE LIGHT AFTER THE FACE CHANGE。
Solutions: Specify the light source orientation in the ZIT hint, such as "lit from the front" or "soft even lighting". Or a smarter way of doing it -- a positive flat on a selfie, the best fit。
3. Use Upscale for final refinement
After a face change, run over Upscale and zoom in 4x-UltraSharp or ESRGAN MODEL) REFERS RESOLUTION TO 2K OR EVEN 4K. SKIN-MASS, HAIR-TWIRL DETAILS WILL INCREASE SIGNIFICANTLY, AND THE WHOLE BODY WILL HAVE A STRONGER SENSE OF "WRITE."。
4. ZIT Prompt Enhancer does not ignore
ZIT has an enhanced hint -- the simple one you write, the model automatically expands on the details. You don't want to write a long, big speech? Turn this on, write anything, and ZIT finish it. It's also a very useful function under the Lora route, which will help you to describe the scene in a much richer way without affecting facial identity。
V. GUIDELINES FOR PLACE SEARCH
Pipe 1: Reactor changed his face on special leave
- Two reasons: first, the colour differences between the source and template maps are too wide, and second, the differences between the template resolution and the source maps are very large. The solution is to choose a selfie close to the skin colour, or to add a requirement like "metching skin tone" to the ZIT hint, so that the colour of the template is close to your selfie。
pit 2: mistake not found insideightface module
- InsightFace, on which Reactor relies, is not installed on some systems. Windows users ensure that there is a Visual C++ Redistributable (microsoft running library). If not, try the pip installation onnxruntime manually。
Pit 3: Visible stitches in the hair and neck of a face change (Reactor route)
- It's Reactor's disease -- it's just a face change, not a neck change. It's too different from your self-portrait, so your neck and hairlines will come out. The solution: a self-portrait with a new hair style close to the template, or directly on the LoRA route - LoRA is an integration process, and there are no seams。
Pipe 4: The ZIT-generated template is too small for Reactor to detect
- Reactor needs to detect face in the template to change. If ZIT produces a full vision, the small face may fail. Solutions: Add "close-up portrait" or "medium shot" to the ZIT hint to make the person a larger percentage of the picture。
Pit 5: Multiple-person photographs changed face only one person
- Reactor defaults on first seen face. When there are many people in the template, a face_index parameter is found in the Reactor node to enter the serial number of the person you want to change (0, 1, 2...). Or check the swap_all_faces option for all faces。
PIPE 6: THE ZIT SAMPLER WAS WRONG AND THE IMAGE CRASHED
- ZIT Turbo must use res_multiistep Sampler+ simple Scheduler, CFG set to zero. Do not use other samplers, with mischarted images that can be exposed, deformed or directly exposed to noise. If you're going to use the CFG as a guide (e.g. with negative hints), set the CFG to 1.2 and Steps to 12。
VI. Who does the package fit
Suitable:
- I'M CURIOUS ABOUT AI'S FACE. I WANT TO TRY IT MYSELF
- Think of a different style of personal writing but don't want to spend money on the studio
- It's from the media. It's all about people
- There's a comfyUI base
- Visible 8G or more (Reactor route) or 12G (LoRA training route)
Not suitable:
- I'm trying to impersonate someone else
- THERE ARE VERY HIGH COMMERCIAL REQUIREMENTS FOR THE QUALITY OF THE DRAWINGS
- Zero Base Pure White -- it's recommended to start with the ComfyUI Basic Graphics, and the face change is a step-by-step game
Final Thoughts
2,800 bucks for one set or zero for an infinite set
I don't have to answer that。
Of course, "0" is not very strict -- you need time to learn the ComfyUI basics, to download model files, to train for one to two hours on the LoRA route. But for these time costsUnlimited writing is free- Today's ancient wind, tomorrow's Saber, day after tomorrow, day after day。
And the Z-Image-Turbo threshold is really low. 8G visibility, an AIO file, 8-step out of 2 seconds. Apache 2.0 is free for business. Ali is really in the atmosphere this time。
ComfyUi changed his face just to start. After this set of jobs, you can go back to the nodes -- the ControlNet for the change of clothes, the layers for the change of background, the LoRA for the change of style... and play more than you can play。
Technical threshold? It'll be done in a weekend. Device threshold? The 8G's visible can run Reactor and 12G can train Lora。
The only thing that is needed is to try。