Gemini returned no image — the five `finishReason` values that mean your picture never arrived

"candidates": [{"finishReason": "IMAGE_SAFETY"}] <-- no parts[], no inlineData, no image

You asked Gemini for an image. What came back is a response object with a candidate in it and no image bytes in the candidate — nothing in an inlineData part — and one bare string in finishReason. Five values of that enum are about images: IMAGE_SAFETY, IMAGE_PROHIBITED_CONTENT, IMAGE_OTHER, NO_IMAGE and IMAGE_RECITATION. This page quotes all five rows verbatim out of a copy of the API reference that was already on this lab’s disk, says plainly that four of the five tell you nothing their own names do not, and then settles the one question the string can actually settle: whether what failed was your prompt or what the model produced.

In ten seconds. All five are FinishReason values, and that enum’s own one-line intro is “Defines the reason why the model stopped generating tokens.” — the model, not your prompt. Two of the rows say outright that the generated images were the problem, in those words, and a third says image generation stopped. NO_IMAGE says no image was produced at all. IMAGE_OTHER says there was an issue.

So none of the five is a prompt-side value, and if you were looking for the one that means “Gemini refused my request”, it is not in this group. That case does exist, but it is not a different value — it is the same string on a different field: IMAGE_SAFETY is also a BlockReason value, where its row reads “Candidates blocked due to unsafe image generation content.” Our arithmetic on Google’s rows: of the five, only IMAGE_SAFETY appears twice in the reference. The other four appear once each, so they can only ever have come off candidates[0].finishReason.

And the honest part. Four of these five sentences restate their own names. “Image generation stopped due to recitation.” is the whole documentation for IMAGE_RECITATION. “Image generation stopped because of other miscellaneous issue.” is the whole documentation for IMAGE_OTHER. We quote them and stop there: this page does not invent a cause, a threshold, a policy name or a retry rule for any of the five, because the reference contains none, and a page that supplied one would be making it up.

The five rows, verbatim, in Google’s own order

These are read out of a stored copy of the Gemini API reference for generateContent. They are not five rows gathered from five places in the file: they are one contiguous block of the FinishReason enum, 408 characters long with nothing else between them, which is the only sense in which Google treats them as a family. The row immediately before the block is MALFORMED_FUNCTION_CALL and the row immediately after it is UNEXPECTED_TOOL_CALL. Google’s order, unchanged:

IMAGE_SAFETY Token generation stopped because generated images contain safety violations.

IMAGE_PROHIBITED_CONTENT Image generation stopped because generated images has other prohibited content.

IMAGE_OTHER Image generation stopped because of other miscellaneous issue.

NO_IMAGE The model was expected to generate an image, but none was generated.

IMAGE_RECITATION Image generation stopped due to recitation.

Two observations about that block, before anything else. First, Google’s order is not a grouping: IMAGE_OTHER, the row that says the least, sits in the middle, between the second output-side row and NO_IMAGE. Any tidier arrangement of these five — including the one further down this page — is ours, not Google’s. Second, the IMAGE_PROHIBITED_CONTENT row reads “generated images has other prohibited content”. That is what the file says; we have quoted it unaltered rather than correcting it, and we mention it only because it is the kind of detail that tells you whether a page like this one read the bytes or recited them from memory.

Prompt or output? The field answers it; four of the five values cannot be on the wrong field

This is the part worth ten seconds, because the fix is different in each case. Gemini has two stop enums in this one file: FinishReason, whose intro is quoted above, and BlockReason, whose intro is:

Specifies the reason why the prompt was blocked.

And promptFeedback.blockReason, the field that carries it, is described in full as:

Optional. If set, the prompt was blocked and no candidates are returned. Rephrase the prompt.

IMAGE_SAFETY is a value of both enums. Its BlockReason row is:

IMAGE_SAFETY Candidates blocked due to unsafe image generation content.

So the same twelve characters mean two different things depending on which field you read them off, and the quoted field description settles which: if promptFeedback.blockReason is set, no candidates came back at all and the instruction is “Rephrase the prompt.” If the string came off candidates[0].finishReason, a candidate existed and its generated image was rejected. Our reading of those two rows is that there is no rephrasing instruction anywhere on the candidate side, for any of the five. The dispatch table, which is our arrangement of Google’s rows and not a structure the file states:

What you sent — one value, one field. promptFeedback.blockReason = IMAGE_SAFETY. The only prompt-side member of this family, and it is a field, not a sixth value. Google’s own instruction for it is the quoted “Rephrase the prompt.”
What came back — three values. IMAGE_SAFETY, IMAGE_PROHIBITED_CONTENT and IMAGE_RECITATION on finishReason. Two of those three rows say generated images in so many words; the third says “Image generation stopped”. Something was produced and then judged — our inference, from the word generated appearing in the rows, is that the prompt got far enough to produce pixels.
Nothing came back — one value. NO_IMAGE, and it is the only one of the five whose sentence carries information its name does not: “The model was expected to generate an image, but none was generated.” — expected. Our reading: the expectation is the one you set in the request, which makes this the only one of the five with a documented connection to anything you control.
No information — one value. IMAGE_OTHER. The row is “Image generation stopped because of other miscellaneous issue.”. We have nothing to add and will not pretend otherwise.

The expectation named by NO_IMAGE is set by the request, and the field is documented:

Optional. The requested modalities of the response. Represents the set of modalities that the model can return, and should be expected in the response. This is an exact match to the modalities of the response.

An empty list is equivalent to requesting only text.

The Modality value you set to ask for pixels has the row “Indicates the model should return images.”, and there is a second knob beside it:

Optional. Config for image generation. An error will be returned if this field is set for models that don't support these config options.

What we will not claim from those three quotes is a procedure. The reference nowhere connects responseModalities or imageConfig to NO_IMAGE; we are putting them next to each other because the NO_IMAGE row contains the word expected and this is the field that expresses an expectation. That join is ours. Note also that Modality is itself defined twice in this one file — once as “Content Part modality” with five values, once as the request field’s enum with three — which is the same documentation habit that produces the value collision above.

Four of these five sentences restate their own names

This is the finding, not a complaint, and it is the reason this page is short. Hold each row against the name it documents:

IMAGE_RECITATION — “Image generation stopped due to recitation.” The sentence is the name with the words reordered. It does not say what recitation is, what was recited, what it was recited from, or what threshold was crossed.
IMAGE_OTHER — “Image generation stopped because of other miscellaneous issue.” A shrug, and an honest one. Compare the plain OTHER row in the same enum, “Unknown reason.”, which is the same admission in three words.
IMAGE_SAFETY — “Token generation stopped because generated images contain safety violations.” Names no policy, no harm category, no threshold and no field to read next. Our check, and it is the one that surprised us: IMAGE_SAFETY appears zero times in the whole 122,408-byte safety-settings guide — the document that configures the filters, and the only one of the five files we hold that explains how to read a safety outcome off a response at all. The guide’s instruction to inspect safetyRatings is written against the plain SAFETY value, not this one.
IMAGE_PROHIBITED_CONTENT — “Image generation stopped because generated images has other prohibited content.” “other prohibited content” is defined by neither row; the word other implies a first category that the enum does not name.
NO_IMAGE — “The model was expected to generate an image, but none was generated.” The exception. This one states a fact about the response that the name alone does not: an image was expected, and the expectation was not met. It is the only one of the five we would call documented.

The one field that may say more: `finishMessage`

If you are holding one of the four uninformative values, there is exactly one place in the response where the API is permitted to elaborate, and it is the last field of the candidate object:

Optional. Output only. Details the reason why the model stopped generating tokens. This is populated only when finishReason is set.

Read it, and log it. We cannot tell you what it contains for any of these five values — we have no captured response, only the reference — and the row promises only that it is populated when finishReason is set, not that it is informative. But it is the only channel the documented schema has for the detail these rows omit, and code that reads finishReason and drops finishMessage is throwing away the only sentence that might not be circular. The prompt side has no equivalent: the string blockReasonMessage appears zero times in all 2,729,724 bytes we searched.

cand = (resp.get("candidates") or [{}])[0] parts = ((cand.get("content") or {}).get("parts") or []) if not any("inlineData" in p for p in parts): log.warning("gemini: no image. field=candidates[0].finishReason value=%s message=%s", cand.get("finishReason"), cand.get("finishMessage"))

That block is ours. No file we read contains a code sample that inspects a finishReason, a logging recommendation, or any guidance on handling a missing image. It records the field name as well as the value for the reason the two-enums page sets out: IMAGE_SAFETY logged bare is ambiguous between the two fields, and the ambiguity is created by the log line, not by the API.

Where the image would have been

Worth knowing if you are writing the check above. A candidate’s content holds an array:

Ordered Parts that constitute a single message. Parts may have different MIME types.

and a single element of that array is one-of:

A Part can only contain one of the accepted types in Part.data.

The following is a list of mutually exclusive fields. At most one of the fields will be set in a response:

— a list on which text and inlineData are two of the alternatives. So the bytes answer part of the co-occurrence question and not the rest, and here is exactly where the line falls. A single Part cannot be both text and image bytes: the reference says at most one field of Part.data is set in a response. A candidate plainly can hold several parts of different types — “Parts may have different MIME types.” What we did not find anywhere in the five files is a statement about whether Gemini actually does return a text part alongside one of these five finishReason values. The page cannot answer that, and the structural quotes above must not be read as answering it: they say such a response is representable, not that it occurs. Write the check so that it tolerates both, which is what the any() above does.

On retrying, which the reference does not discuss at all

The word retry appears zero times in the 1,232,462-byte reference. It appears eight times in the troubleshooting guide, and that guide has a section on recitation. Here it is, complete:

If you see the model stops generating output due to the RECITATION reason, this means the model output may resemble certain data. To fix this, try to make prompt / context as unique as possible and use a higher temperature.

We are not going to hand you that as the fix for IMAGE_RECITATION, and the reason is the point of this section. That advice is written against the plain RECITATION value, about model output that “may resemble certain data”. The string IMAGE_RECITATION appears zero times in the troubleshooting guide. The word image occurs in that guide exactly three times, and all three are inside the page’s own navigation menu — they sit within 24 characters of each other and 4,083 characters before the first occurrence of temperature. The prose of that guide is about text. The step from “raise the temperature to fix text recitation” to “raise the temperature to fix image recitation” is a guess, it would be an attractive and plausible-sounding guess, and we are declining to make it. If you try it, you are running an experiment, not following documentation — and the same guide warns in the next breath that changing temperature away from its default on recent models “can cause unexpected behavior”.

Three questions this page cannot answer

Is a partial image ever returned? We cannot tell you, and neither can these files — which is not the same as “no”, and we are not going to round it to one. The word partial occurs once in the entire 1,232,462-byte reference, and it is not about images at all — it is about the code-execution tool: “Code execution ran for too long, and was cancelled. There may or may not be a partial output present.” Nothing in the five files we hold says whether a blocked or stopped image generation returns truncated bytes, a placeholder, or nothing. We looked, and the search is the only reason we can say this at all; we are not asserting it is undocumented elsewhere.
Which policy, and what threshold? Not in these files, for any of the five. The safety guide, which lists harm categories and settable thresholds, does not contain the string IMAGE_SAFETY or IMAGE_PROHIBITED_CONTENT.
Which models produce which of the five? The rows name no model. The imageConfig row quoted above is the only place the reference admits that image support varies by model, and it says only that setting the field on a model without support returns an error.

What this page does and does not claim

Claimed, from quoted rows: the five values exist in FinishReason and form one contiguous block of it; IMAGE_SAFETY is also a BlockReason value and the other four are not; the candidate object has a finishMessage and the prompt side has no message field; a Part holds one data type and a candidate may hold parts of several.
Claimed, as ours and labelled in the sentence: the prompt / output / nothing split; that the words generated images, in the two rows that contain them, imply pixels were produced; that responseModalities is the expectation NO_IMAGE refers to; the logging block.
Not claimed: any cause, threshold, policy name, harm category or retry rule for any of the five. Four of these rows do not support one and we have not supplied one from elsewhere.
Not claimed: that the text-recitation advice applies to IMAGE_RECITATION.
Not claimed: anything about SDK behaviour. We read no SDK source in any language, so whether a client library raises, returns an empty part list, or exposes finishMessage at all is outside everything we checked.

Provenance

Every sentence inside a quote block above was read out of a copy of the document it is attributed to that was already stored in this lab’s repository before this page was written. Nothing was fetched to write this page — not one request left the machine — and nothing here is quoted from memory. The copies, with the byte count of the file actually read:

Gemini API reference, generateContent — 1,232,462 bytes, fetched 4 October 2026 at HTTP 200, carrying its own date Last updated 2026-09-23 UTC. The source of all five rows, both enums, and every field description above.
Gemini API docs, Troubleshooting — 105,225 bytes, same fetch, Last updated 2026-10-01 UTC. The source of the recitation advice we declined to transplant.
Gemini API docs, Safety settings — 122,408 bytes, same fetch, Last updated 2026-09-17 UTC.
Gemini API docs, Thinking — 286,274 bytes, same fetch, Last updated 2026-09-25 UTC.
Gemini API docs, Function calling — 983,355 bytes, same fetch, Last updated 2026-09-23 UTC.

The reference page and the safety-settings page each carry the same licence line: “Except as otherwise noted, the content of this page is licensed under the Creative Commons Attribution 4.0 License”.

The counts. Searched across those five files, tags stripped, whitespace collapsed, whole-token matches only — so IMAGE_SAFETY is not counted as an occurrence of SAFETY, and partial does not count partially. Columns are reference, safety-settings, troubleshooting, thinking, function-calling. The first five lines are the whole argument of this page: four of the five values occur exactly once in the reference and IMAGE_SAFETY occurs twice, which is both the collision and the proof that no row we quoted has a longer version elsewhere in the file.

ref saf tro thi fun NO_IMAGE 1 0 0 0 0 IMAGE_SAFETY 2 0 0 0 0 IMAGE_PROHIBITED_CONTENT 1 0 0 0 0 IMAGE_RECITATION 1 0 0 0 0 IMAGE_OTHER 1 0 0 0 0 RECITATION 1 0 1 0 0 finishReason 6 2 0 0 0 finishMessage 2 0 0 0 0 blockReason 2 1 0 0 0 blockReasonMessage 0 0 0 0 0 promptFeedback 3 2 0 0 0 inlineData 9 0 0 0 0 responseModalities 2 0 0 0 0 imageConfig 2 0 0 0 0 safetyRatings 11 2 0 0 0 partial 1 0 0 0 1 retry 0 0 8 0 0 Modality 6 0 0 0 0

Read off that table: all five values appear zero times in four of the five files — they are documented in exactly one place, the enum. blockReasonMessage is zero everywhere, which is the evidence for the no-prompt-side-message claim, while finishMessage is 2 in the reference. retry is zero in the reference. partial is 1 there, and that one occurrence is quoted above and is about code execution. inlineData is 9 in the reference and zero in the other four, so the field your image should have arrived in is described in exactly one of the documents that might describe it.

Those counts are the evidence for every claim of absence on this page. We did not read the SDK source for any language, we captured no live API response, and nothing above describes either. We did not follow a single link out of these five files. A claim on this page about what Google does not say is a claim about these five files on the date they were fetched, and is not a claim about every document Google has published. Where this page reasons past the documentation it says so in the sentence that does it.

There is a written guide: the full finishReason and BlockReason dispatch as one table you can hold against your own response handler, with the image values, the field that disambiguates IMAGE_SAFETY, and the handler above in a form you can paste.

No page on this site has a checkout widget of its own. There is a written guide behind this host and it is on sale at $19 on a storefront that delivers the files automatically and carries a 30-day money-back guarantee: buy it there; the guide can also be paid for with 19 USDC on Base at the payment page, where delivery is by hand as a reply to your email. Every page on this site, including this one, is free to read in full, with no sign-up and nothing gated.

The short version: IMAGE_SAFETY, IMAGE_PROHIBITED_CONTENT, IMAGE_OTHER, NO_IMAGE and IMAGE_RECITATION are one contiguous block of Gemini’s FinishReason enum, so every one of them is about the model stopping rather than your prompt being refused — except that IMAGE_SAFETY is also a BlockReason value, which makes the field it arrived on, not the string, the thing that tells you whether to rephrase. Three of the rows say the generated images were rejected, NO_IMAGE says none was produced, and IMAGE_OTHER says “Image generation stopped because of other miscellaneous issue.”. Four of the five sentences restate their own names, so the only elaboration the documented schema offers is finishMessage on the candidate — log it, and log which field the value came off.

Nearby

blockReason vs finishReason: five values in both enums — the field-versus-value problem in general. IMAGE_SAFETY is one of the five names it is about, and this page is the image half of it.
finishReason: SAFETY, RECITATION, OTHER — the text-side values with the same defect: rows that explain a value using the value’s own name.
finishReason: MALFORMED_FUNCTION_CALL — the row immediately before this block in the enum, and the clearest row in it: a condition you can reproduce.
finishReason: UNEXPECTED_TOOL_CALL — the row immediately after this block, and one that points at your request rather than at the output.
stop_reason: max_tokens — the other way a response arrives incomplete, at another vendor, where the row does explain itself.

This page counts anonymous readership, and it carries one counter rather than the two that the older pages on this host carry. Each load sends the page path, the address of the page you came from, how long the page was open, whether you scrolled, and any campaign or outreach code in the link you followed; a second “engaged” event is sent once, ten seconds after the page opens — whether or not the tab is in front of you — or as soon as you scroll a quarter of it. The campaign codes from the first link you arrived on are kept in this browser’s local storage, and a later visit that arrives with no codes of its own is counted against them; a link carrying its own codes is used for that visit, and the stored first touch is never replaced. An outreach code is removed from the address bar after it is read. Because this page sends one view event rather than two, a view count taken from it is directly comparable to a load, which is not true of the sixteen pages on this host that send two — any rate measured against those reads half its true value. No name is attached to any of this: the only identifier the code can send is an outreach token minted per recipient, and no link carrying one has ever been sent for this page. No cookie; the local storage above does that job. The page also asks this domain for an analytics script at /_vercel/insights/script.js; on 26 September 2026 that address returned HTTP 404 on every host we publish, so no script from another company was served or ran — the page goes on asking, so this stops being true the moment that address starts answering, without a byte of this page changing. Your browser and the network attach things the page does not send: the identification string your browser gives, your IP address and the time of the request. The host that serves this page keeps its own request logs; those are its record and not ours.

Gemini returned no image — the five finishReason values that mean your picture never arrived