Reference images: upload, compression, previews, and originals

Choose an entry for your purpose

Purpose	Recommended action	Result
Place images in Canvas before deciding their use	Canvas Upload or image paste.	Each file becomes a standalone image node. Upload alone creates no prompt, output, or connection.
Generate immediately with several references	Card input picker or canvas multi-reference node.	An ordered input. Confirming selection is not generation; enter a prompt and submit.
Develop an existing result	Continue creating from that result.	Retains its source relationship without downloading and re-uploading.
Organize regularly used materials	Add to Personal materials in Asset Library.	An independently organized material to select later, not a new generation task.

Selection still needs confirmation before upload

Open the picker and choose device images, history, or personal materials. Similar thumbnails are not necessarily the same source.
Local selection first displays a device preview. Remove or reselect as needed. A preview is not an upload; check estimated storage and current capacity.
Accept or decline compression when prompted. Upload starts only after selection confirmation. Wait for all local files; retry or remove failures.
After success, check order, actual reference count, and model compatibility. A previously accepted count may not suit the current model.

**Image selection before confirmation**
Local preview is not upload. Check estimated usage and available capacity and remove unwanted files first.
Actual interface · Tutorial example, not your account

File size, image count, and storage are separate limits

Each file must be a decodable image. Renaming an extension does not convert a file. Corrupt or excessively large pixel dimensions can be rejected.
Static JPEG, PNG, and WebP are suitable common inputs. Export unsupported device-photo formats first, and select a static frame from animation rather than treat image upload as video input.
Byte size differs from image dimensions and supported model specifications. A small file can contain many pixels; successful upload does not guarantee compatible format, size, or reference count.
Reference limits depend on model and creation entitlements and use expanded image count, not slots or nodes. Images on the whole canvas are not the same as references per request.
Sufficient account storage is also required, including space reserved for ongoing tasks. Shared upload and generation use the owner's workspace capacity even when a participant pays.

Loading current rules…

Why compression is offered and what becomes the original

Large files normally prompt for compression above 7 MiB; a lower configured upload limit can prompt sooner. Compression runs in your browser to reduce upload size while retaining usable detail. It does not silently replace an already saved original.

Compression may re-encode, change format, or reduce dimensions. Opaque images can become JPEG; transparent images use a transparency-capable format. Fine text, texture, and edges can change. Keep the device original and inspect saved dimensions and detail.

After accepting compression and uploading, BANANAI saves the processed file. 'Download original' and generation then use that saved file, not your device's pre-compression original. Declining neither bypasses limits nor completes upload.

Compression does not resolve corruption, insufficient storage, or model incompatibility. For typography, detailed comparisons, or print, inspect the saved original rather than just a small preview.

Thumbnail, preview, and saved original

A thumbnail may appear while its preview loads and is replaced on success. Preview failure does not mean the saved image or task is lost; check connectivity and retry.

Normal generation uses saved originals, not list thumbnails. Local repaint combines the original and your regions for model input. Screen markings indicate edits, not secretly using a low-resolution preview. The model may still process or resize inputs internally.

For deliverable files, inspect the preview, download the original, and verify readability and dimensions. A screenshot only saves the display, not the original file.

Version	Where it is used	What it does not mean
Thumbnail	Dense asset lists and reference strips, reducing network and memory use.	AI is not limited to receiving that thumbnail, and the original has not been reduced to thumbnail resolution.
Preview	Loaded when opening images and on demand in visible canvas areas. Scrolling an asset gallery does not preload every original.	A screen-viewing version, not necessarily suitable for printing or pixel-level inspection.
Saved original	Original downloads, formal editing, and normal generation references.	Not the pre-compression file still on your device and not the image's on-screen size.

**Small images for efficient asset browsing**
These are not generation reference files. Open a clearer preview and download the original for delivery.
Actual interface · Tutorial example, not your account

**Shared enlarged-image viewer**
The image is still a screen preview. Dimensions and Download original help verify the actual file; thumbnail, preview, and downloaded original differ.
Actual interface · Tutorial example, not your account

Give every reference a role

For a character poster, image 1 might provide face and clothing, image 2 composition, and image 3 colors. Order them first, then state each role. Conflicting requirements for one object may be blended by the model.

Continuing from a node means following that creative step. Independently selecting the same file from history means reusing material. Identical images can have different creative relationships; continue from or explicitly reference the current result to retain its lineage.

Repeating an image does not merge its uses or guarantee more weight. It consumes reference slots. Keep duplicates only when separate regions or purposes are genuinely needed and explain each slot.