Comparing Original Fidelity Across 7 Free AI Models When Using Hand-Drawn Reference Art
ねDo you ever use image reference features?
I use them from time to time, but I've noticed that depending on the model, it might generate something super close to the original image, or end up creating something completely different 😳
There are already some great posts discussing image reference features, like this one by Kobato🕊️_DOVE:
(Check it out here → )
And in the gallery, I've seen posts where creators like nazo-man seemingly fed the same reference image into multiple models:
(Check it out here → https://muryou-aigazou.com/ja/gallery/7z3xv41qs542)
However, I couldn't find any posts that broken down the detailed differences between each model's reference feature...
So today, I decided to test 7 free models to see how closely their outputs stick to the original reference image when generating anime-style illustrations!
[Test Setup]
- Using the drawing below as the reference image, I generated an image with each model using the exact same prompt and seed value.
(I couldn't use AI-generated art to keep things fair across models, and couldn't use other people's artwork due to copyright, so I used my own hand-drawn sketch 😅)

-
Shared prompt: "anime style, girl, short wavy blonde hair, headband with blue ribbon, blue eyes, winking, white cape, brown coat, yellow sash belt, selfie angle". I kept it minimal, just capturing the key features of the original sketch.
-
I rated the fidelity (how close it is to the original) on a 3-point scale across three criteria: Silhouette/Contour, Details, and Color Palette.
(Note: Since you can only upload up to 5 images per post, some of the images below are combined into pairs)
[Results]
① Image V3 (Silhouette: 3/3, Details: 3/3, Colors: 3/3)
② Anim4gine (Silhouette: 2/3, Details: 1/3, Colors: 1/3)

Image V3 stayed incredibly close to the original in terms of pose, outfit details, and color scheme! It gave the sketch a nice polish while preserving all the key features. The random gibberish text and the extra mini-girl on the side are pretty hilarious too 😆
Anim4gine had some issues with the outfit coherence, turned the signature into a pattern, deepened the ribbon and coat colors, and changed the girl's facial expression. That said, if you're a fan of retro 90s anime styles like Sailor Moon, you might really dig this vibe 📺️
③ Animij (Silhouette: 2/3, Details: 1/3, Colors: 2/3)
④ WAI (Silhouette: 2/3, Details: 2/3, Colors: 2/3)

Animij added a bandana, tweaked the ribbon shape, and blended the signature into the background. While putting its own fantasy-inspired spin on the design, it kept the girl's original expression and overall composition intact 🏰
WAI slightly altered the ribbon's shape and color, as well as the girl's mouth, but there are no glaring glitches. The overall framing, composition, and palette are almost identical to the original! The signature got turned into a cool pattern ✨
⑤ Neta Yume (Silhouette: 1/3, Details: 1/3, Colors: 2/3)

Here, the ribbon shape/color and mouth changed, the signature morphed into part of the cape, and the yellow sash became a diagonal belt. However, the bangs and eyes are drawn in much finer detail compared to the original sketch 🔍️
⑥ Image V2 (Silhouette: 3/3, Details: 3/3, Colors: 3/3)
⑦ SDXL (Silhouette: 1/3, Details: 1/3, Colors: 2/3)

Image V2 is almost an exact match to the original, capturing the overall composition, hair/ribbon outlines, and even the subtle color contrast between the ribbon and headband—all while adding a lovely hand-drawn feel! The nonsensical text is funny, and the extra chibi character is super cute 😂
SDXL, on the other hand, turned the signature into lighting effects, shifted the eye color from blue to green, and bizarrely swapped the brown and white parts of the outfit! The white section became the sleeves, and the brown turned into a cape, making the clothing structure look pretty messy 😥
[Summary]
Ranking the models from highest to lowest fidelity based on these results:
| Model Name | Silhouette | Details | Colors | Total |
|---|---|---|---|---|
| ① Image V3 | 3 | 3 | 3 | 9 |
| ⑥ Image V2 | 3 | 3 | 3 | 9 |
| ④ WAI | 2 | 2 | 2 | 6 |
| ③ Animij | 2 | 1 | 2 | 5 |
| ② Anim4gine | 2 | 1 | 1 | 4 |
| ⑤ Neta Yume | 1 | 1 | 2 | 4 |
| ⑦ SDXL | 1 | 1 | 2 | 4 |
(Table formatted in Markdown)
In this test, the facial expressions varied quite a bit across models. It seems that without explicit prompts like "smiling," facial expressions can be a bit of a wildcard.
(On the flip side, that means expressions are likely easy to steer with prompts!)
Since I used an image with no background and a signature instead, models handled it very differently—some generated random text, while others turned it into scenery or patterns. This suggests that background fidelity also varies significantly depending on the model 🤔
While results will naturally change depending on your prompt style and the type of reference image you use, if you want to stay faithful to your original art, models like Image V3 are the way to go. If you want creative reinterpretation, models like Anim4gine shine!
Definitely worth experimenting with 🔍️
[Bonus Note]
Fun fact: I drew the original sketch using a basic built-in drawing app on my laptop, using nothing but my finger on a tiny trackpad (I wanted to see if I could do it, but wow, it was hard to draw 😅).
Which output is your favorite? If I'm judging them purely as standalone illustrations, Image V3 and Animij are my top picks!