An unlabeled image asks the model to guess what should transfer
A portrait contains identity, pose, wardrobe, lighting, background, palette, crop, and mood at the same time. A style sample contains subjects and objects in addition to rendering choices. Without a declared role, any reference can compete with the prompt or leak irrelevant structure into the draft.
Multiple references amplify that ambiguity. Numbering and role labels let the workflow distinguish the intended signal from details that should be ignored, while specialized preprocessing gives face and color references narrower paths before the model request is assembled.
Best for
Separating facial identity cues from style, layout, lighting, and background sources.
Combining several references when each image has one explainable contribution.
Reordering references so their numbered directives match the intended priority and discussion order.
Not designed for
Recreating a source image, transferring every visible detail, or guaranteeing exact likeness.
Using images without the rights or permission required for the intended project.
Assuming every account or generation path has the same reference entitlement or fixed cost.
How labels change reference preparation
Studio infers an intent from the saved type, label, or description. Preset labels map to face, style, composition, color, lighting, pose, background, texture, or mood. A missing or unrecognized label becomes generic inspiration with an instruction not to recreate specific identity, composition, objects, or text. Dragging changes the array order and therefore the reference number used in role directives.
Before generation, a Face intent attempts a centered crop from the upper portion of the source and resizes that crop for model use; if preprocessing fails, the original remains available. A Color Palette intent extracts and sanitizes palette values, contributes them to brand-color guidance, and is excluded from model image inputs. Remaining model-bound references carry explicit instructions about what to use and what to ignore. The Studio cost breakdown should be reviewed because model-bound references can add cost.
What each part does
These are the practical decisions available in the current TubeBoosts workflow.
Upload, drop, or paste
Adds one validated and compressed image at a time until the eight-image maximum is reached.
Choose the clearest source for one role instead of uploading several near-duplicates.
Reference label
Stores free text and infers a recognized role when the wording matches a supported intent.
Use the provided suggestions when possible so the backend receives an unambiguous intent.
Label suggestions
Offers Face, Character, Portrait, Style Reference, Composition, Color Palette, Lighting, Character Pose, Background, Texture, and Mood choices.
Character and Portrait map to the same face intent as Face.
Drag handle
Reorders the reference list and its displayed index numbers.
Review labels after reordering because backend directives refer to each image by its new position.
Remove
Deletes the reference from the current list and attempts to clear its cached image.
Remove references whose role duplicates stronger evidence or increases cost without adding useful direction.
Estimated Cost breakdown
Shows the current generation estimate, including reference-related cost when the calculation applies it.
Check the live estimate rather than relying on a fixed entitlement or price statement from this guide.
How to use AI Thumbnail Reference Image Roles
A short process that keeps the tool inside a real thumbnail decision.
Assign one job before upload
Decide whether each candidate image exists for face, style, composition, color, lighting, pose, background, texture, or mood.
Check: Can you finish the sentence 'Use this image only for...' with one role?
Upload and apply a recognized label
Add the image, choose the closest preset label, and avoid combining several intents in one free-text phrase.
Check: Does the selected label describe what should transfer and imply what should not?
Order and prune the set
Drag references into a deliberate sequence and remove duplicates, weak sources, or images whose rights are uncertain.
Check: Does every numbered image contribute information absent from the others?
Review preprocessing and cost implications
Expect face-role cropping and color-role palette extraction, then inspect the current Studio estimate before generating.
Check: Are you comfortable with the transformed role and the displayed cost for this exact request?
Three references with three non-overlapping jobs
Setup
A hypothetical creator has a portrait of the host, a product photo with useful diagonal placement, and a poster whose teal-coral color relationship fits the series.
Action
They label the images Face, Composition, and Color Palette, then order them intentionally. The face image follows the face-crop path, the product image remains model-bound with composition-only direction, and the poster contributes extracted colors instead of being sent as a general reference.
Practical result
The hypothetical request tells the workflow which evidence to use from each source while limiting transfer of wardrobe, objects, text, and unrelated style.
Lesson
Role labeling reduces ambiguity; it does not promise exact identity, geometry, palette placement, or output quality.
What this tool does not promise
Studio accepts up to eight references, but fewer well-labeled sources can provide a clearer brief.
Face preprocessing uses a fixed geometric crop rather than semantic face detection and can fall back to the original image if processing fails.
Color-role images become extracted palette guidance and are excluded from ordinary model image references.
Reference availability and cost can vary by active plan, model path, resolution, and request; use the live Studio state as the authority.
Questions about AI Thumbnail Reference Image Roles
What happens if I leave a reference unlabeled?
It is treated as generic inspiration. The backend adds a directive not to recreate specific identity, composition, objects, or text from that source.
Does Face use the entire uploaded image?
The preparation path first attempts a centered upper-area crop and resize. If that operation fails or the image is too small for the normal path, the original can remain in use.
Is a Color Palette reference sent to the model?
Not as an ordinary reference image. Its colors are extracted into palette guidance, and the image is filtered out of the model-reference list.
Do reference images cost more?
They may add generation cost. Studio calculates the current request and displays a reference-related breakdown when applicable, so check that estimate before submission.
Try it in the real workflow
Tool pages explain the current TubeBoosts controls. The Studio is where the tool works with your own prompt, references, settings, and generated image.
Label References in Studio