BananaBanana Image, Video & Speech Generation

Previous value: -"nano-banana-2"New value: +"nano-banana-2-lite"

Previous value: -"nano-banana-2-lite: cheapest, 1K only. nano-banana-2: balanced default. nano-banana-pro: top quality, up to 4K."New value: +"nano-banana-2-lite: cheapest default, 1024 only. nano-banana-2: choose for 512, 2048 or 4096 output. nano-banana-pro: top quality, up to 4K."

Previous value: -"512 only on nano-banana-2; 1024 only on lite; 4096 costs the most."New value: +"Lite requires 1024. Choose nano-banana-2 for 512, 2048 or 4096; 4096 costs the most."

Added value: +{
+  "default": false,
+  "description": "Switch this call to Google's own permissive presets: safety thresholds OFF and person generation allowed for adults (safetySettings and personGeneration=ALLOW_ADULT — documented Vertex AI parameters, not a way around Google's rules). Google filters in two independent stages and this flag only affects the first: (1) a check on the request before anything is rendered — this is what refuses legitimate prompts about real-looking people, swimwear or sportswear, medical and anatomical subjects, mild fictional violence and edgy artwork (upstream_reason SAFETY_BLOCK), and it is the stage relaxed_filter relaxes; (2) a classifier that scores the finished image, which no parameter can switch off (upstream_reason IMAGE_SAFETY). Stage 2 judges the actual pixels and every run renders a different image, so a borderline but legitimate subject rejected there often passes on a plain retry or with slightly less ambiguous wording — retrying is reasonable (failures are fully refunded, so an attempt costs only time), just don't expect the flag itself to bypass that stage. What Google refuses outright — sexual content, minors, real public figures — stays blocked in every attempt. Both stages surface as error_code SAFETY_FILTERED; upstream_reason in get_result tells them apart.",
+  "type": "boolean"
+}

Added value: +{
+  "default": false,
+  "description": "Switch this call to Google's own permissive presets: safety thresholds OFF and person generation allowed for adults (safetySettings and personGeneration=ALLOW_ADULT — documented Vertex AI parameters, not a way around Google's rules). Google filters in two independent stages and this flag only affects the first: (1) a check on the request before anything is rendered — this is what refuses legitimate prompts about real-looking people, swimwear or sportswear, medical and anatomical subjects, mild fictional violence and edgy artwork (upstream_reason SAFETY_BLOCK), and it is the stage relaxed_filter relaxes; (2) a classifier that scores the finished image, which no parameter can switch off (upstream_reason IMAGE_SAFETY). Stage 2 judges the actual pixels and every run renders a different image, so a borderline but legitimate subject rejected there often passes on a plain retry or with slightly less ambiguous wording — retrying is reasonable (failures are fully refunded, so an attempt costs only time), just don't expect the flag itself to bypass that stage. What Google refuses outright — sexual content, minors, real public figures — stays blocked in every attempt. Both stages surface as error_code SAFETY_FILTERED; upstream_reason in get_result tells them apart.",
+  "type": "boolean"
+}

Added value: +{
+  "default": false,
+  "description": "Veo only (omni-flash ignores it), and much weaker than its image counterpart: Google's video API exposes no configurable safety settings at all, so this flag can do only one thing — pin personGeneration=allow_adult, which is already the default for Veo 3.1 here. Expect it to change little. Video is filtered both before generation and again on the finished clip (rai_media_filtered_reasons), and neither check can be turned off. When a clip is refused, the levers that actually work are: retry (each run renders different footage, and failures are fully refunded), reword the people-related part of the prompt, or move an omni-flash idea to veo-3.1-fast, whose filter is noticeably looser.",
+  "type": "boolean"
+}

Added value: +{
+  "description": "Actual visual references for the generated image. Each item is a job_id of a completed image generation on this account, a public http(s) image URL, or an inline data:image/png|jpeg|webp;base64,... URL (max 10 MB each). For a local file, read and base64-encode its bytes into a data URL; a bare local path such as C:\\\\photo.jpg cannot be read by this remote server.",
+  "items": {
+    "type": "string"
+  },
+  "maxItems": 14,
+  "type": "array"
+}

Previous value: -"Start the clip from a still image, which is then animated. Either a job_id of a completed image generation on this account (see list_generations) or a public http(s) image URL — including a signed url returned by get_result, which is how you pick one variant of a multi-image generation. On omni-flash it can be combined with reference_images; on Veo it cannot."New value: +"Start the clip from a still image, which is then animated. Either a job_id of a completed image generation on this account (see list_generations), a public http(s) image URL, or an inline data:image/png|jpeg|webp;base64,... URL for a local file (max 10 MB; a bare local filesystem path cannot be read by this remote server). This includes a signed URL returned by get_result, which is how you pick one variant of a multi-image generation. On omni-flash it can be combined with reference_images; on Veo it cannot."

Previous value: -"Reference images that keep a subject, character or style consistent (they are NOT used as literal frames). Each item is a job_id of a completed image generation on this account or a public http(s) image URL. omni-flash: up to 10 images in total together with first_frame. Veo: at most 3, only with duration 8 and without first_frame."New value: +"Reference images that keep a subject, character or style consistent (they are NOT used as literal frames). Each item is a job_id of a completed image generation on this account, a public http(s) image URL, or an inline data:image/png|jpeg|webp;base64,... URL. For a local file, pass a base64 data URL rather than its filesystem path. omni-flash: up to 10 images in total together with first_frame. Veo: at most 3, only with duration 8 and without first_frame."

Removed value: -8

Previous value: -"Length of the edited clip in seconds ($0.10/s)."New value: +"Optional: trim the source to the first N seconds and edit only that part ($0.10/s). Values above the source length are ignored — the clip cannot be made longer."

Added value: +{
+  "description": "Returned by the quote when video_url is used: pass it back with confirm_cost to reuse the already downloaded clip instead of downloading it again.",
+  "type": "string"
+}

Previous value: -"omni-flash only: job_id of a completed omni video to refine conversationally; prompt describes the changes."New value: +"omni-flash only: job_id of a completed omni video to refine conversationally; prompt describes the changes. Duration, aspect ratio and the scene are inherited from that clip — duration is ignored here (use edit_video to shorten a clip)."

Added value: +{
+  "description": "Start the clip from a still image, which is then animated. Either a job_id of a completed image generation on this account (see list_generations) or a public http(s) image URL — including a signed url returned by get_result, which is how you pick one variant of a multi-image generation. On omni-flash it can be combined with reference_images; on Veo it cannot.",
+  "type": "string"
+}

Added value: +{
+  "description": "Reference images that keep a subject, character or style consistent (they are NOT used as literal frames). Each item is a job_id of a completed image generation on this account or a public http(s) image URL. omni-flash: up to 10 images in total together with first_frame. Veo: at most 3, only with duration 8 and without first_frame.",
+  "items": {
+    "type": "string"
+  },
+  "maxItems": 10,
+  "type": "array"
+}

Added value: +{
+  "default": 8,
+  "description": "Length of the edited clip in seconds ($0.10/s).",
+  "enum": [
+    3,
+    4,
+    5,
+    6,
+    7,
+    8,
+    9,
+    10
+  ],
+  "type": "integer"
+}

Previous value: -"Public http(s) URL of the source video (mp4/mov/webm/mkv, max 200 MB). Use instead of source_generation_id."New value: +"Public http(s) URL of the source video (mp4/mov/webm/mkv/avi/wmv/flv/3gpp, max 200 MB). Use instead of source_generation_id."

Previous value: -"Seconds. Ignored by omni-flash (the model picks 3–10 s itself)."New value: +"Clip length in seconds. Veo accepts only 4, 6 or 8; omni-flash accepts any value from 3 to 10."

Previous value: -[
-  4,
-  6,
-  8
-]New value: +[
+  3,
+  4,
+  5,
+  6,
+  7,
+  8,
+  9,
+  10
+]

Added value: +"What to avoid. On omni-flash it is appended to the prompt as plain text (the model has no separate negative field)."