Most losses in SLYGEN are not related to the model itself, but to how the user formulates the query and selects the source data. According to internal statistics, up to 80% of the first generations end up being reprocessed or fail to deliver the desired result precisely due to input errors.
Below are the 8 most common reasons why coins are spent inefficiently, and how to structure your process so that the result is more stable from the very first attempt.
Mistake #1. Poor Source Image
The most frequent cause of losses is a blurry or overloaded source photo.
What happens: the model "fills in" details instead of relying on the reality of the frame.
What it looks like: artifacts on the face, distortion of features, loss of identity.
How to fix it:
- a clear face occupies most of the frame
- even lighting without harsh shadows
- simple composition (portrait or half-body shot)
- high resolution (1024 px and above)
Mistake #2. Too Long a Prompt
The more text there is, the higher the chance that some parameters will be ignored.
The problem: the model loses priorities and mixes details.
How to fix it: Use a short structure: character → BREAK → pose → BREAK → light/scene.
Example:
1 girl, blonde, blue eyes, BREAK, full body, sitting in a chair, BREAK, soft light
Instead of: a long description where everything is mixed without structure.
Mistake #3. Missing BREAK
BREAK is not "decoration," but a separator of semantic blocks.
If it's missing:
- the pose may conflict with the lighting
- the face loses priority
- the scene becomes chaotic
Correct approach: each block is responsible for one level:
- character
- composition
- light / atmosphere
Mistake #4. Incorrect Word Order
The model "listens" most strongly to the beginning of the query.
Common error: details first, then the character.
Correct logic:
- who is depicted
- key features
- body / accents
- pose
- light and atmosphere
This is critical for a stable result.
Mistake #5. Ignoring Built-in Tags
Tags are not an add-on, but a way to set the "quality level."
What using them provides:
- a clearer face
- stable anatomy
- improved detail
Mistake #6. Overloading with Details
An overly complex query reduces control over the result.
Symptom: the scene becomes random, elements "drift."
Working rule: Maximum 40–60 words for the main prompt.
Important: it's not the number of details, but their hierarchy.
Mistake #7. Ignoring the Gallery
The gallery is not just examples, but ready-made visual solutions.
How to use it correctly:
- find a similar pose
- fix the composition
- adapt it to your character
This sharply reduces the number of iterations.
Mistake #8. Complex Formulations
Too "literary" language worsens recognition.
Why: the model works better with direct and concrete words.
Bad: "an elegant figure with a deep gaze"
Good: "1 girl, blue eyes, portrait"
How to Build a Stable Result
A basic system that reduces coin losses:
1. Source
- clear face
- good lighting
- simple scene
2. Prompt
- short
- structured
- using BREAK
3. Tags
- score_9
- detailed face
- natural lighting
4. Rely on the Gallery
- don't invent a pose from scratch
Most losses in SLYGEN are related not to generation quality, but to the lack of query structure. When the input becomes simple, separated, and predictable, the model starts working stably from the very first attempt, without unnecessary iterations and coin consumption.