I’m trying to find an AI image generator that can take an existing photo and turn it into new versions or styles without losing key details. I’ve tested a few tools, but they either change the image too much or don’t follow the original closely enough. Can anyone recommend reliable options or share settings/tips to get accurate AI-generated images from a source picture?
Short answer, you want “image-to-image” with strong control, not pure text-to-image.
Here is what usually works best without losing key details:
-
Use tools with a “strength” or “denoise” slider
- Stable Diffusion (Web UIs like Automatic1111, ComfyUI)
- Leonardo AI
- NightCafe
- DALL·E 3 has weaker fine control, often overchanges details.
Set strength low, around 0.25–0.45, so it keeps structure and identity.
-
Use ControlNet for Stable Diffusion
If you are ok installing stuff, this is the most reliable option.- Install Automatic1111 + ControlNet extension.
- Upload your photo.
- Choose ControlNet models:
• openpose for pose
• depth or lineart for structure - Prompt only the style: “oil painting, studio lighting, high detail, same person, same pose”.
- Leave negative prompt with things you do not want: “different face, deformed hands, extra fingers”.
This locks composition and pose while letting style change.
-
Use “reference image” or “image guidance” features
Midjourney:- Upload your photo.
- Use something like:
/imagine prompt: portrait in anime style --iw 1.5
Higher --iw keeps more from your original. Lower value changes more.
Still tends to alter faces, so expect retries.
-
Use face-specific tools if the subject is a person
- FaceSwap + Stable Diffusion: keep exact identity, change clothes or background.
- Some sites have “consistency mode” or “same character” toggles.
Those help when you want stylization without a new person.
-
Split your workflow
If tools overcook the image in one step, do it in two.- Step 1: Run a mild image-to-image with low strength, focus on color and texture.
- Step 2: Upscale and clean with a face restoration model or an upscaler.
This keeps facial structure while improving style.
-
Concrete settings that tend to work
For Stable Diffusion Automatic1111 image-to-image:- Denoising strength: 0.3–0.45
- CFG scale: 6–8
- 20–30 steps, sampler DPM++ 2M or similar
- Prompt: “cinematic portrait, soft lighting, detailed skin, same face, same pose”
- Negative: “different person, different pose, extra limbs, blurry, low detail”
-
Things to avoid if you want consistency
- High strength/denoise, like 0.7+. That rewrites the image.
- Overly complex prompts that describe new poses, clothes, or angles.
- Face-heavy style prompts like “pixar style” without adding “same face”.
If you share what kind of photo you have and what style you want, people can suggest exact prompts and settings for a specific tool.
I’ll push in a slightly different direction than @cacadordeestrelas here and focus less on the “how to drive Stable Diffusion” angle and more on workflow choices that keep your original photo recognizable.
A few things that actually make a bigger difference than the specific tool:
1. Decide what must stay identical vs what can change
Most tools freak out when your prompt secretly asks for two conflicting things.
Before you start, literally write this down:
-
Must stay the same:
- face / identity
- pose
- camera angle
- composition (e.g., “full body, centered, background trees”)
-
Can change:
- color grading
- background type
- clothing style
- art style (oil / anime / cyberpunk, etc.)
Then, in your prompts, do not describe things that belong in the “must stay the same” bucket. Just describe style and mood. If you say “dynamic low-angle shot” but your photo is a straight-on portrait, the model will “fix” your photo by changing it.
2. Use “style only” prompts
Instead of:
“a fantasy knight, dramatic angle, golden armor, stormy sky, portrait of a man”
Use:
“same person, same pose, painterly fantasy style, soft brushstrokes, rich colors, detailed lighting, fantasy color palette”
And leave the original content (pose, camera angle, identity) to the image, not the text. This lowers the chance of it inventing a new face.
If the tool supports it, some flavor of “preserve subject” or “keep composition” option helps a lot. Many web tools quietly have a toggle like:
- “Respect original image”
- “Keep subject”
- “Content weight / Image weight”
Crank that toward “respect original,” even if it limits how wild the result looks.
3. Go for style transfer tools instead of general generators
This is where I slightly disagree with relying too heavily on full-blown image-to-image Stable Diffusion if you mainly want “same photo, new vibe.”
If your goal is:
“Same shot, but in watercolor / anime / line art / comic shading”
Then you might actually get better consistency from tools that are closer to style transfer than generative remix. Look for features like:
- “Photo to cartoon”
- “Photo to anime”
- “Photo to sketch / painting”
- “Filter” systems that accept a reference style image
These tend to lock your composition and just re-render it, instead of hallucinating differences. Less exciting, more reliable.
A surprisingly good workflow:
- Use a style transfer / filter tool first to keep the structure intact.
- Then run that stylized result through a lighter image-to-image pass if you want more detail.
You end up nudging an already stable stylization instead of fighting the model to keep the face the same.
4. Use reference style images instead of long prompts
If your tool allows giving:
- 1 image as “base photo”
- 1 or more images as “style reference”
Do that instead of writing a flowery paragraph. Models are better at copying style from another picture than from overly poetic prompts.
Example setup:
- Base: your portrait
- Style ref: a painting / anime frame / graded movie still you like
- Prompt: 1 short phrase like “same person, same pose, in the style of reference image”
The style image does the heavy lifting. Simpler text prompt = fewer weird hallucinations.
5. Post-process to restore details instead of trying to get it perfect in one go
Common pattern:
- AI image: good style, but face lost some likeness
- You try again with stronger constraints and end up in a loop
Alternative:
- Accept a slightly-off stylized version.
- Then use:
- face restoration / face enhancement tools
- traditional editors (Photoshop, GIMP, Affinity) to paint back small details from the original (eyes, mouth shape, birthmarks) on top of the stylized version with low-opacity brushes.
You don’t actually have to demand the AI keep every detail if you’re willing to nudge it manually after.
6. Start with smaller changes and stack them
Instead of jumping straight to “turn my office selfie into an epic cinematic fantasy scene,” do something like:
Step 1:
- Same photo, just better lighting and color grading.
- Use AI as a smart filter (like Lightroom on steroids).
Step 2:
- Keep that result as base.
- Change only background style (e.g., realistic office to painted backdrop).
- Avoid touching the face again.
Step 3:
- Optional subtle stylization (slight painterly look, mild comic shading, etc.)
By stacking gentle changes, you’re less likely to cross the line where the AI decides your subject is a whole new person.
7. Watch resolution and crop
A weird one, but it matters:
- Tight crops of the face tend to make the model obsess over facial features and sometimes “beautify” them too much.
- Overly large images can cause the model to treat different areas inconsistently.
Try:
- Crop to just enough context: head and shoulders if it’s a portrait.
- Work in a moderate resolution first (like 512 or 768 on a side in SD-land, or roughly 1–1.5K in typical web tools).
- After you like the look, upscale using an upscaler that promises to keep identity.
You can even upscale using non-AI options to avoid further changes.
8. When a tool keeps overchanging, switch type of tool, not just settings
If you’ve already tried a couple of generators and all of them replace the face or pose, try a different category:
- From: text-to-image or heavy image-to-image generator
- To: filter / style-transfer / “creative photo editor” type tool
The former want to “create,” the latter want to “transform.” That mindset in the model makes a very practical difference.
If you want to describe one example photo (like “waist-up portrait, indoor, neutral background”) and one target vibe (“Ghibli style,” “80s film grain,” “digital painting”), I can suggest a concrete workflow that doesn’t require learning a full Stable Diffusion stack.