When a business creates videos on a schedule β rather than once a quarter β the question inevitably arises: where to source images for each scene. Generating from scratch seems like an obvious solution: describe it in a prompt, and receive an image. However, it has a systemic problem that doesn't manifest immediately, but rather after a dozen or so videos β the frames end up featuring faces that don't exist, interiors that aren't in your office, and products that look almost like yours, but not quite. At Telematic, we developed a separate function specifically for this situation β stylizing your own photos β and in this article, we will explore how it differs from regular generation and when to choose which approach.
Stylization is not a filter over a photo or generation "inspired by" your shot. It involves applying one of the 29 visual styles of the project to a real image that you uploaded, while preserving the composition, faces, objects, and details of the frame. The result is a recognizable image in a unified visual language of the video, rather than an abstract illustration with a similar mood.
How Stylization Differs Fundamentally from Generation from Scratch
In generation from scratch, the model receives a text prompt and draws the scene "from a blank slate": the speaker's face, the store's interior, the product's shape β all of this is interpretation, not fact. For videos on abstract topics, this works excellently: the viewer doesn't care whether the abstract "manager at a laptop" in the illustration is real. But as soon as a recognizable element must be in the frame β the company founder, a specific room, a specific product model β generation from scratch begins to "fill in" what should be a fact.
Stylization addresses this differently: it takes your real photograph as a basis and applies a visual style β color, light, texture, character of processing β while recognizable elements of the frame remain themselves. The founder's face remains their face, just in a different color palette and with a different drawing style. The interior remains your interior, not an abstract "similar" space.
When Recognizability is More Important than Beauty
There is a category of content where the loss of recognizability is not a minor flaw, but a failure of the entire idea of the video:
- Personal Brand β if the audience is subscribed to a specific person rather than an abstract expert, the face in the frame must be theirs, not a generated double.
- Real Estate β showcasing a specific property, not any similar apartment with a similar renovation.
- B2B Agencies and Contractors β cases with real client interiors, equipment, and teams.
- E-commerce β the product card must show exactly this model, not something similar.
In all these scenarios, generation from scratch becomes a risk: a viewer familiar with the original immediately detects the substitution, and trust in the video decreases β even if the image is objectively beautiful.
How It Works in Practice
The process is simple: you upload a photo β your portrait, an office photo, a product shot, a frame from a previous shoot β and choose the visual style of the project or specific scene. The system then applies processing so that the result fits into the overall visual line of the video, but does not lose the original identity of the frame.
An important point: stylization works at the scene level, not for the entire video as a whole. This means that in one video, you can mix stylized real photos and generated illustrations from scratch β where recognizability is not critical, you can leave the generation, and where the fact is important, stylize your own. Such mixing is standard practice, not a compromise: most videos should not consist of frames of one type.
A Practical Example from Our Practice
When we tested the function on our own channel, the difference was particularly noticeable in videos formatted as "case studies." Previously, in such videos, a generated illustration of "the team at work" was included for a couple of seconds β it looked decent, but was clearly not our team. After switching to stylization, we began uploading real photos from internal meetings and processing them under the visual style of the video. The result β the same neat visual line, but the viewer sees real people, not a stock illustration with an unclear face.
One of our clients, who runs a channel about room renovations, found that stylization addressed a specific pain point: previously, in "before/after" videos, the object after generation looked more aesthetically pleasing than it actually was, leading to questions from the audience in the comments β "is this really the same apartment?" After switching to stylizing their own photos, the questions ceased: the recognizability of the object was preserved, while the visual presentation still matched the overall style of the channel.
When Generation from Scratch is Still Better
Stylization is not a universal substitute for generation, and it is important to understand the boundaries of applicability:
- If you physically do not have the original photo of the required quality or angle β there is nothing to stylize, and generation from scratch will fill the gap faster.
- For abstract, illustrative scenes (metaphors, section covers, background frames), recognizability is not needed in principle β there, generation from scratch provides more artistic freedom.
- If the photo is of low quality, blurry, or heavily cropped, the result of stylization will be limited by the source β in such cases, it is sometimes more honest to generate an alternative illustration than to try to salvage a poor shot.
Practical Tips for Uploading Photos for Stylization
To ensure that stylization yields predictable results, it is advisable to follow a few simple rules. Upload photos with sufficient resolution β the more detailed the source, the more accurately the style will apply. Try to ensure that the main object β face, product, interior β occupies a significant part of the frame, rather than getting lost in the background. If the project specifies a unified visual style, apply it consistently to all stylized scenes β this way, the transition between stylized and generated frames will be less noticeable.
A separate tip for teams that run several channels with different visual identities: create a library of "reference" photos β employee portraits, interiors, key products β and reuse them across videos. This way, each new visual style is immediately applied to a tested set of frames, rather than being assembled anew.
In conclusion, the logic is simple: where a fact must be in the frame β a specific person, a specific place, a specific product β use stylization of your own photos. Where an illustration of an idea is needed, rather than documentary accuracy, generation from scratch remains a fast and flexible tool. The combination of both approaches within a single video is not a compromise, but a normal working process.