1
Upload and validation
First, the browser prepares a preview, while the server receives the image and validates file type, size and technical integrity. Empty, broken or unsuitable images do not enter processing.
How generation works
From photo to result: what DejaNu does under the hood
From the outside, it looks like uploading one photo. Inside, the image follows a clear path: file validation, temporary storage, queueing, a GPU worker, generation and result storage in history. Below, we walk through that path in plain language, without making it feel like you opened a neural-network textbook halfway through.
From upload to result
What happens to a photo inside the service
In the real service, a generation is a job with statuses: after the photo is accepted, a record is created, credits are charged, the job enters the queue, a worker processes it and the result is returned to account history.
2
Safety rules
Before processing, the task is checked against age restrictions, consent requirements, prohibited scenarios and technical limits.
3
Generation queue
The task receives an identifier and the Queued status. A worker picks the next job, moves it to Processing and starts the run. If infrastructure is temporarily unavailable, the job is returned to the queue instead of disappearing.
4
Encoding into tensors
The worker reads the source from temporary storage and sends it into the generative pipeline. There, the photo becomes numerical feature maps: edges, light, pose, textures and composition become data the model can transform.
5
Generation steps
The selected quality level affects the number of generation steps. Each step is a small refinement: the model reduces uncertainty, aligns details and moves the internal representation closer to a finished image.
6
Result and history
When the generative pipeline returns an image, the service stores the result, marks the job as Done and shows the finished file in account history.
The AI part
What happens inside generation
After the queue, the most interesting part begins: the photo enters an image-to-image workflow. The model does not “look” at the JPEG like a person and does not redraw it by hand. It converts the image into a compact internal representation, refines it step by step and then turns it back into a normal image.
Important: AI does not recover hidden reality
The model does not know what was really behind covered regions or outside the frame. It builds a plausible visual result from visible cues: pose, lighting, edges, composition and learned visual patterns. That is why a good source photo matters more than any magic mode.
01
Tensors
A tensor is a multidimensional table of numbers. In images, those numbers can describe not only pixels, but also features: edges, light, color transitions, depth and relationships between regions.
02
Latent space
Latent space is a compressed internal representation of the image. The model does not edit every pixel directly; it works with a compact feature map, and a decoder turns that map back into an image.
03
Generation steps
Generation does not happen in one jump. At each step, the model refines the internal representation: reducing noise, aligning shape, light and details. More steps usually give more room for refinement, but require more compute.
04
Quality level
In DejaNu, the quality level is tied to the number of generation steps and compute cost. A higher level can improve stability, but it cannot fully fix poor lighting, heavy blur or bad framing.
05
LoRA adapters
LoRA does not store finished results and does not replace the base model. It is a small low-rank weight adaptation: it gently changes the model’s behavior in a visual area, like an extra tuning layer on top of the main engine.
06
Post-processing
After generation, the result is prepared for delivery: storage, account linking, status updates and access through history.
Input quality
Why the source photo matters so much
Generation is not a reconstruction of hidden reality. The model creates a plausible visual interpretation from the input image, so lighting, pose, sharpness and framing directly affect result stability.
Sharpness
Heavy blur and compression make boundaries and small details harder for the model to understand.
Lighting
Balanced light helps preserve volume, skin tone, shadows and natural contrast.
Pose
A clear pose and silhouette reduce ambiguity during generation.
Framing
Cropped body parts, covered regions and unusual angles increase the chance of artifacts.
One main subject
Photos with multiple people or overlapping objects are harder to process consistently.
Quality level
A higher level adds compute headroom, but the result still depends on the source image.
Privacy and safety
Files are not public and are not stored forever
Uploads and generated results are available in the user's account. Materials can be deleted manually from history, and older files are removed according to retention rules. DejaNu also applies safety rules for adult-only content and prohibited scenarios.
Quick answers
Common questions about the AI process
Does DejaNu reconstruct reality? +
No. AI does not know what was outside the frame or behind covered regions. It creates a plausible result from visible cues, learned patterns and the selected quality settings.
Why can the same photo produce different results? +
Generative models use a probabilistic process. Even with the same photo, small differences in noise, queue state, resources and settings can change result details.
What does LoRA do in generation? +
LoRA slightly changes the behavior of the base model through small additional weights. It helps the model work more confidently with certain visual features, but it does not guarantee a perfect result and does not replace source photo quality.
Why are some photos rejected? +
The service applies safety rules and technical limits. Some images are unsuitable because of content, quality, format or intended scenario.
Can I delete a result? +
Yes. Materials can be deleted from account history, and older files are also removed according to retention rules.
Want to see the difference on a real example?
Before and after demo photos are available on the homepage. You can review the examples first, then upload your own photo and choose the generation quality level.