Dreamina is a practical first choice for a free image-to-image workflow when you want to upload a reference, explore a new version and continue refining it in one workspace. For other priorities, Gemini offers conversational edits, Leonardo lists limited reference access with Auto selection, and ComfyUI offers deeper local control when you are willing to configure the workflow.
Image-to-image AI uses an existing image to guide a generated result. That can mean turning a sketch into an illustration, changing a visual style or developing a new setting around a subject. It does not automatically mean that every original detail remains unchanged. Choose the tool by the transformation you want—and by how much deviation you can accept.
- Six image-to-image routes compared
- 1. Dreamina: from a reference to a developed visual
- 2. Gemini: for an edit that is easier to explain than configure
- 3. Leonardo.Ai: for a limited public reference edit with Auto
- 4. Adobe Firefly: for guided reference-based transformations
- 5. Krea: for exploring a direction quickly
- 6. ComfyUI: for explicit control over transformation strength
- Why a prompt saying “keep everything else unchanged” can fail
- A short workflow that makes the comparison useful
- Which free image-to-image tool should you start with?
Six image-to-image routes compared
The access information was reviewed in October 2026. This is a workflow comparison supported by product documentation and a limited editing example, not a six-tool preservation benchmark.
1. Dreamina: from a reference to a developed visual
When to use it. Dreamina fits a creator who wants to transform a reference into a useful visual without immediately building a technical workflow. You might use a rough composition to guide an illustration, explore another setting or develop a concept for a social image.
The Dreamina image-to-image generator is the relevant starting point. Give the source image a defined role: subject identity, composition, color or style. Then describe the new result. A reference is more useful when the instruction explains which of its properties matter.
How to control the change. Keep the request focused. “Use this layout but make it a watercolor scene” is a different task from “change only the background while preserving this real product.” The second has a much stricter preservation requirement. If the visible controls support a more targeted edit, use them; do not assume a broad reference-generation route enforces a pixel-level boundary.
What is free. Dreamina offers daily free credits, currently 120 per day in the United States. Actual quantities are subject to the final output because model and settings affect consumption. Seedream 5.0 Lite is a documented free model option. A free reference-generation route does not establish that every Pro model or advanced selection tool is included.
What a free reference edit can get wrong
A controlled October 27 edit used a fictional tote-bag image. The instruction asked for only the surrounding background to become black and white, while preserving the colored bag. The free Seedream 5.0 Lite route returned a downloaded image, but the cream fabric and brass-colored button also became desaturated.
The result preserved the broad subject and composition, but not all the details that the instruction protected. For a restyled concept, that degree of change may be acceptable. For a real product listing, it is a reason to preserve the subject as an original layer and edit the background separately. A successful download and a successful edit are different outcomes.
2. Gemini: for an edit that is easier to explain than configure
When to use it. Gemini offers a direct conversational route: upload the image, describe a change and review the result. It is useful for casual restyling and simple scene variations without learning denoise or node controls.
How to control the change. Google's current documentation distinguishes Nano Banana 2 Lite from Nano Banana 2 and the paid Pro redo route. Lite is optimized for quick ideas and is not designed for multiple references or repeated edits. That matters when you want to combine a character reference, a separate environment and a style image over many revisions.
What is free. Free image use has limits and supported-account conditions. Editing requires an adult account. Google lists 1K downloads without an AI plan and 2K with a Google AI plan. Choose the actual route available to you rather than assuming a paid demonstration describes the free model.
3. Leonardo.Ai: for a limited public reference edit with Auto
When to use it. Leonardo's current Free plan explicitly lists limited Image Guidance, along with 150 tokens per day, basic quality and Auto Only model access. That supports trying a simple public reference-led edit; it does not make the whole model library or every guidance mode freely selectable.
How to begin. The official Image Guidance guide starts from the home prompt bar: click the Image Guidance icon, select an image reference, upload the source, describe the intended transformation and generate. On Free, remain within the limited reference functions available with Auto. The plan does not promise the six-reference allowance shown for paid tiers, and specialist style, content or character controls have their own model and plan conditions.
What is free. The daily 150-token allowance resets every 24 hours, and unused tokens do not accumulate. Actions and settings affect the cost, so this is not a fixed number of completed edits. Free creations are public; do not upload confidential client material to this route.
Choose it for a small public experiment when automatic model selection is acceptable. If your requirement is a specific model or a particular reference-control mode, the Free plan is not a substitute for the relevant paid access.
4. Adobe Firefly: for guided reference-based transformations
When to use it. Firefly is useful when you want a visible, guided image-to-image process rather than a pure chat interface. Its product documentation describes uploading a reference, entering a prompt, selecting a model and adjusting relevant strength and style controls.
The official product-page screenshot below shows the reference-upload entry and an illustrative transformation. It helps identify the workflow, rather than proving the result of a comparison test. Captured October 28, 2026.
How to control the change. Decide whether the reference should guide structure, style or a more literal transformation, then use the applicable controls for the selected model. A strength setting makes the degree of influence more explicit, but it is not a guarantee that a particular label or face will remain exact.
What is free. Adobe provides limited daily generations without a paid plan. Its own Firefly models and partner-model actions can have different access and credit conditions. The free entry point is useful for trying the workflow; it is not one fixed image allowance covering every model and resolution.
5. Krea: for exploring a direction quickly
When to use it. Krea combines real-time creation, image editing and enhancement. It is useful when you want to move through visual directions and compare how a reference can be interpreted.
How to control the change. Treat the output as a creative version unless its important details survive inspection. An appealing texture or lighting change can disguise a different face, object proportion or surface pattern. Decide which changes are welcome before adding more enhancement.
What is free. The current Free plan lists 100 compute units per day. Different operations consume different amounts. Commercial licensing is a paid-plan condition, so the free route is a place to experiment rather than assume permission for a client deliverable.
Choose Krea when the feedback loop helps you find a look. Choose a more controlled mask or local workflow when the task is repeating an exact change across a catalog.
6. ComfyUI: for explicit control over transformation strength
When to use it. ComfyUI is free workflow software that can run compatible models locally. It suits someone who wants to retain settings, repeat a sequence and understand what each part of the process does.
How to control the change. In its conventional image-to-image example, the source is encoded and sampled with a denoise value below 1.0. Lower denoise adds less noise, so the image generally changes less. That is a practical control over the transformation, though it does not by itself guarantee identity or text preservation.
Masks and model-specific conditioning can provide further control. Keep the model version, workflow and settings together so a useful result can be reproduced. A saved image alone is not a reusable production process.
A concrete starting model. FLUX.2 [klein] 4B supports reference-image editing, and ComfyUI's official guide provides a 4B Image Edit Distilled template. Use that template with the matching 4B model, text encoder and VAE; put your reference into its image-input node and describe the change. This is a model-specific reference-editing route, so do not assume that every control is identical to the conventional denoise example above.
What is free. The 4B model is released under Apache 2.0, allowing commercial use subject to the license, and can run with free ComfyUI software. Black Forest Labs estimates about 13 GB of graphics memory for its stated model configuration; ComfyUI's optimized templates can use a different amount. Hardware, storage and setup are still your responsibility. The 9B versions use different, non-commercial licensing, so the 4B permission must not be applied to the entire FLUX family or other checkpoints. A workflow calling a hosted API also has that service's costs and conditions.
Why a prompt saying “keep everything else unchanged” can fail
A reference-guided generator produces a new image conditioned on the input and instruction. It may preserve the broad layout while reconstructing materials, edges or small details. That is why a bag can remain recognizable while its cream fabric becomes white.
A mask-based composite solves a different problem. It can retain original pixels outside the edited region, then blend the change at the boundary. When the subject must remain accurate, that distinction is more useful than simply adding stronger preservation language to the prompt.
For style exploration, some reconstruction is the point. For product documentation, identity-sensitive portraits or evidence, preservation usually has priority. Choose the editing method accordingly.
For a selective correction around the generated image, free photo-editing tools with layers can be more useful than another full-frame transformation.
A short workflow that makes the comparison useful
- 1
- Define the source's role. Is it the subject, layout, style or a reference for one element? 2
- Name one transformation. State what should change and identify the details that determine whether it worked. 3
- Choose suitable controls. Use a reference-strength, denoise or mask route when it serves that requirement; do not assume all tools expose the same controls. 4
- Review the whole result. Check both the edited region and the important details that should have remained stable. 5
- Download and inspect. Confirm dimensions, visible branding and whether the final file fits the intended use.
Keep a copy of the original and the successful settings. If an exact-preservation task continues to change the subject, switch to a layered edit instead of using the rest of your free allowance on the same unsuitable route.
Compare one controlled change and its usable cost
Use one source file and a short written brief for the first comparison. For example, ask two suitable routes to turn the same fictional house image into a watercolor illustration while keeping the roof line and viewpoint recognizable. This proposed exercise allows stylistic change. It is different from the tote-bag task above, where material colors had to survive exactly. Do not judge both against the same tolerance for invention.
Save the chosen model, reference role, strength or denoise setting where exposed, aspect ratio and displayed action cost. Change one variable in the next attempt. If the style is too weak, alter its strength while holding the composition brief stable; if the roof is distorted, reduce transformation or improve the input instead of also adding a new camera angle.
Allocate the allowance before the first submission: available balance minus a revision reserve minus any quoted finishing or enlargement cost leaves the initial-generation budget. Those units belong to the selected platform; another tool's tokens are not interchangeable. If a free route publishes no dependable amount, plan one available attempt and a download check rather than promising a whole pack.
A few repeats can be useful when consistency is the question, but one successful sample cannot establish a success rate. The tote-bag edit proves a narrower distinction: a free generation and download succeeded while the requested preservation failed. If repeated exact-preservation attempts keep changing the subject, spend the remaining effort on a mask-based edit of the original instead of buying more of the same unsuitable transformation.
Which free image-to-image tool should you start with?
Begin with Dreamina when you want a browser workflow that connects reference-led creation with refinement. Use Gemini for a simple conversational change, Leonardo for a limited public reference edit with Auto, Firefly for guided reference controls, Krea for creative iteration, or ComfyUI for local technical control.
The best choice is not the tool that changes the image most dramatically. It is the one that changes the right things while preserving enough of the original for your purpose.