Google Gemini

This is the current Google Gemini text node - the replacement for the deprecated first-gen Gemini node, and the one you should be reaching for. Same family job (prompt in, STRING out, multimodal context allowed) but the internals are a generation newer: a dynamic model dropdown that reveals per-model controls, thinking levels, and a token budget big enough to actually matter. If you've got questions like "what does this render look like" or "draft a prompt that keeps the composition," this is the natural home.

The mechanism

Prompt plus optional media goes to Gemini's Generate Content API through Comfy's proxy, billed through your Comfy account in credits - no Google key of your own, price badge on the node, cloud-only. Pick a model in the dynamic combo and the node grows that model's inputs. The current lineup: Gemini 3.5 Flash, Gemini 3.1 Pro, and Gemini 3.1 Flash-Lite. The headline control is thinking_level - how hard the model reasons internally before answering. MINIMAL/LOW on Flash-Lite for cheap fast stuff, HIGH on Pro for hard problems. HIGH is smarter and it's also slower and pricier, because thinking burns tokens.

The inputs that matter

• prompt - the text, multiline.
• model - dynamic combo, revealing per-model inputs: images (up to 16), audio (one clip), video (one clip), files (from Gemini Input Files), plus thinking_level, temperature (default 1.0), top_p (0.95), and max_output_tokens (default 32768, up to 65536).
• seed - default 42; the tooltip is refreshingly blunt: set 0 for a random seed, and deterministic output isn't guaranteed.
• system_prompt - optional foundational instructions.

Output is a single STRING, wired anywhere text goes.

The one trap that bites everyone

max_output_tokens includes the model's thinking tokens. Run thinking_level HIGH with a low cap and the model spends the whole budget reasoning - you get back an empty or truncated answer. The node's own tooltip tells you the fix: raise max_output_tokens with HIGH thinking, and don't sweat a big cap on short replies because the model stops early when it's done. That's the difference between this node feeling broken and feeling great.

Beyond that, it's the family's standard caveats: needs a Comfy account with credits, needs internet, non-deterministic, and 16 images plus one audio and one video clip is your multimodal ceiling. If you need a different media mix, the older deprecated node still runs - but for images-and-text, this is the one. It shipped as the current version of the Gemini text node, replacing V1, and it's where the new models keep landing.