From drawing to image
Upload a drawing to ZeroTwo, describe the finished image, and an image model redraws it. Below: a real drawing and its result, an honest read of what carried over and what did not, and how the underlying techniques differ.
Free plan, no card required. Image generation is limited on Free; see pricing.
From ZeroTwo's own image-generation gallery. Which model made it is not recorded there.
Two different jobs hide behind the same search. Decide which one you have before you pick a tool.
The two images above, compared piece by piece. The prompt was:
Create a realistic version of this drawing, paying attention to the details to create a lifelike drawing.
The flat-topped, boxy head, the two big ears at the upper corners, two small eyes, a nose, sets of whiskers and the wide smile all carry over, as does the upright stance.
The tangle of loops became fur with swirl curls, and the model added folded paws and a curling tail. Nothing in the drawing says fur; the model decided that.
The ruled lines, the binder holes and the table edge are gone. The result sits on a plain cream background. That is usually what you want, and it is also a reminder that the model redraws the whole image rather than tracing the lines.
Red crayon became warm orange-red with a soft pencil texture, so the result reads as a rendered illustration, not a photograph. The prompt only asked for a realistic version. The frame also changed shape.
Search results mix these together. They give different amounts of control, so the row you are in decides what to expect.
| Method | What the model is given | What tends to survive | Where you meet it |
|---|---|---|---|
| Reference image plus text promptThe route on this page | Your drawing as an input image, plus instructions in plain language. | Subject and composition. Exact stroke positions are not guaranteed, and you steer by wording (what to keep, what to add). | Image-edit models, including those listed further down this page. |
| Image-to-image (img2img) | Your image, encoded and partly noised, then denoised under the prompt. | Set by the strength parameter: lower values stay closer to the original, and 1.0 mostly ignores it. | Open-source pipelines such as Hugging Face Diffusers. Not claimed for ZeroTwo on this page. |
| ControlNet with a conditioning map | A separate map made from your drawing (scribble, Canny edges, depth, segmentation or pose) next to the prompt. | The information in the map. This is the route to keep lines in place. | Self-run tools built on the ControlNet code and models. Not claimed for ZeroTwo on this page. |
| Text prompt only | Words and no drawing. | Nothing from a drawing, because none is supplied. | A different job. See AI image generator from text. |
Technique descriptions come from Hugging Face Diffusers: image-to-image guide, ControlNet paper, arXiv:2302.05543, and ControlNet repository README, read on 5 October 2026.
The only result that matters is your drawing's. Use the same sketch for every attempt so the comparison is fair, and see image generation in ZeroTwo for the upload flow. If your drawing is a room plan, the AI interior design page shows how a concept sketch feeds a room redesign.
Photograph it straight on in even light, crop to the artwork and raise the contrast so dark lines read clearly against the paper. Pencil and pale pen fade first, so go darker than looks natural.
Name the parts that must not move: the head shape, the roofline, the label positions. A sentence that begins with Keep gives the model a constraint instead of leaving it to guess.
Material, lighting and camera feel are what a drawing cannot hold. If you want a photograph, ask for photographic traits: our example prompt only asked for realistic and came back as an illustration.
Open both at the same scale and flip between them. Check part counts, proportions, where the main lines landed and the frame shape, since the aspect ratio can change.
Edit the wording or switch the model, not both at once, so you can tell which change helped.
Starting points to adapt, not tested outputs. Each one pairs a Keep clause with an Add clause, then names what to check.
Studio product photograph of [PRODUCT], based on the attached drawing. Keep the silhouette, proportions and the position of every labelled part. Add a matte ceramic finish, soft three-point softbox light, a seamless grey backdrop and an 85mm lens look.
Check afterButton, label and logo positions.
Photorealistic exterior of the building in the attached sketch. Keep the massing, the roofline and the window grid exactly. Add polished concrete and brushed steel, golden-hour light and a wide-angle lens look.
Check afterNumber of stories and window bays.
Render the attached line drawing as a lit portrait photograph. Keep the pose, the face shape and the hair silhouette. Add rim light, shallow depth of field and fine film grain.
Check afterFace proportions, hands and any accessories.
Turn this storyboard frame into a cinematic still. Keep the camera angle, the subject placement and the horizon line. Add dramatic backlight and a 2.39:1 frame.
Check afterAspect ratio. You may need to crop.
These accept an input image with your instructions. We have not run one sketch through all of them, so we do not rank them: try your drawing on two and keep the better reading. Which ones your plan includes is on the pricing page, and the wider catalog is under models.
It is an AI image tool that takes a drawing as input and returns a more finished image that follows its subject and composition. How closely it follows your lines depends on the method: a reference image with a prompt follows loosely, while a conditioning map such as a ControlNet edge or scribble map is built to hold structure in place.
In image-to-image generation the drawing is encoded, noise is added to it, and the model removes that noise while following your prompt. A strength setting controls how far the result may drift: lower stays closer to the original and 1.0 more or less ignores it. ControlNet adds a separate conditioning image, such as a canny edge map, a depth map or a scribble, and the generated image preserves the information in it.
This page does not claim that. What ZeroTwo's own image-generation examples show is a drawing uploaded as an image with a text instruction, handled by an image-edit model. If you need explicit edge or depth control, ControlNet is open-source software that you run yourself.
Not for a hosted tool: the provider runs the model on its own servers. Running ControlNet yourself does need a GPU; the repository documents a low-VRAM mode for 8 GB cards. Check the hardware notes for the specific model you pick.
Yes, as an uploaded image. Practical advice rather than a tested rule: photograph it straight on, crop to the artwork and increase the contrast first, since faint pencil is the first thing to disappear. Then say in the prompt which parts must stay put.
That depends on the model provider's current terms and on your plan, and this page does not summarize licenses. Read the provider's terms before you publish client or brand work, and keep a record of which model made the image.
Image generation is limited on the Free plan ($0, no card required) and is included on Plus ($14.99/month) and Pro ($29.99/month, $26.99/month billed annually). Which image models each plan includes, and the credit allowances, are on the pricing page.