ComfyUI-Gemini-Direct

ComfyUI Gemini Direct

Direct Google Gemini image generation, video generation (Gemini Omni), and AI prompt enhancement for ComfyUI — bypass the credit system and use your own API key.

A drop-in replacement for ComfyUI's built-in "Nano Banana Pro (Google Gemini Image)" node, plus an AI-powered prompt engineering tool. Instead of paying through ComfyUI's opaque credit system, this pack calls the Google Gemini API directly with your own key and shows you the real USD cost per generation.

Nodes

1. Gemini Image Generate (Direct API)

Direct image generation with 3 model tiers and real cost transparency.

| Input | Type | Description | |-------|------|-------------| | prompt | STRING | Text prompt for image generation | | model | COMBO | Model tier selection (Pro, 3.1 Flash, 2.5 Flash) | | seed | INT | Seed for reproducibility | | aspect_ratio | COMBO | auto, 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9 | | resolution | COMBO | 512, 1K, 2K, or 4K (falls back if unsupported by model) | | response_modalities | COMBO | IMAGE+TEXT or IMAGE only | | images | IMAGE (optional) | Batched reference images as visual context | | system_prompt | STRING (optional) | System instruction (default forces image output) | | api_key | STRING (optional) | Google AI API key |

Outputs: images (IMAGE), text (STRING), cost_info (STRING), cache_key (STRING)

Cost per Image

| Model | Resolution | Cost per Image | Best For | |-------|-----------|---------------|----------| | Gemini 3 Pro | 1K-2K | ~$0.13 | Best quality | | Gemini 3 Pro | 4K | ~$0.24 | High-res best quality | | Gemini 3.1 Flash | 512-1K | $0.05-0.07 | Balanced quality/cost | | Gemini 3.1 Flash | 2K-4K | $0.10-0.15 | Balanced, higher res | | Gemini 2.5 Flash | 1K | ~$0.04 | Cheapest option |

ComfyUI's built-in Gemini node charges credits: 211 credits = $1 USD, with no visibility into what each generation actually costs.

Reference Images

Batch multiple reference images and feed them as visual context. For example, batch a background image with character photos, then prompt "integrate these people into the background scene."

Cache Key Output

The cache_key output encodes all generation parameters (model, seed, aspect ratio, resolution, prompt) into a deterministic string. Wire it to the ComfyUI API Optimizer's Hash Vault so any parameter change triggers a fresh generation, while identical runs hit cache at $0.

---

2. Prompt Studio (AI Enhancer)

AI-powered prompt engineering using Gemini text models. Domain-adaptive with a built-in anti-AI realism mandate.

| Input | Type | Description | |-------|------|-------------| | brief | STRING | Creative direction (short brief or full prompt) | | mode | COMBO | Expand, Refine, or Edit | | model | COMBO | Gemini 3 Flash recommended (~$0.001/call) | | images | IMAGE (optional) | Reference images for visual context analysis | | previous_prompt | STRING (optional) | For Edit mode: wire enhanced_prompt back here | | feedback | STRING (optional) | For Edit mode: what to change | | custom_instructions | STRING (optional) | Extra directives for any mode | | api_key | STRING (optional) | Google AI API key |

Outputs: enhanced_prompt (STRING), analysis (STRING)

Three Modes

Expand — Short brief to full structured prompt. The LLM detects the creative domain (architecture, fashion, portrait, product, fine art) and generates appropriate section structures automatically.

"brutalist concrete interior, Tadao Ando, morning light, ArchDigest editorial"
  --> Full structured prompt with Spatial Composition, Material & Texture,
      Light Study, Architectural Context, Editorial Style sections

Refine — Takes your existing full prompt and surgically strengthens weak areas, adds missing technical detail, and enforces the realism mandate. Preserves your intent and structure.

Edit — Iterative adjustment loop. Wire the previous enhanced_prompt back into previous_prompt, write your feedback, and only the sections you mention get changed.

Prompt Studio (Expand) --> enhanced_prompt --> Preview
                                  |
                 Not happy?       |  Wire it back:
                                  v
                Prompt Studio (Edit) <-- previous_prompt
                     feedback: "change lens to 24mm, make it golden hour"
                          |
                          v
                    edited_prompt --> Gemini Direct

Anti-AI Realism Mandate

Every prompt — regardless of domain — is engineered to avoid AI-generated aesthetics:

• Real camera body + lens combinations (Leica M6, Hasselblad 500C/M, etc.)
• Specific film stocks or sensor characteristics (Kodachrome 64, Portra 400)
• Optical imperfections: chromatic aberration, vignetting, halation
• Physical grain structure (silver halide, sensor noise)
• Micro-texture on all surfaces: pores, weave, wood grain, concrete
• No plastic skin, no AI sheen, no impossibly clean environments

Domain-Adaptive Structure

The LLM automatically detects the creative domain from your brief and generates the most appropriate section structure:

• Portrait/Street — Camera, Subject Lock, Tonal Refinement, Environment, Photographic Style
• Architecture/Interior — Spatial Composition, Material & Texture, Light Study, Architectural Context
• Fashion/Editorial — Model Direction, Wardrobe & Styling, Lighting Design, Set Design
• Product/Commercial — Product Detail, Lighting Setup, Surface & Background
• Fine Art — Composition & Form, Palette & Tonality, Medium Reference

---

3. Prompt Studio Settings (v1.2.0)

Companion node that owns the style dropdown for Prompt Studio. Emits the selected style as a wildcard STRING output. Wire that single output to BOTH Prompt Studio's style input (right-click on the style widget → Convert Widget to Input) AND to a Hash Vault any_input_N slot.

| Input | Type | Description | |-------|------|-------------| | style | COMBO | Master photographer DNA prompt (or None) |

Outputs: style (wildcard STRING)

This is the single-source-of-truth pattern: picking the photographer in one place automatically drives both the prompt enhancement AND the cache key, so cache invalidation can never desync from the style actually applied. Mirrors the Gemini Style Transfer Settings companion node from the Style Transfer workflow.

A demo workflow showing this wiring is at ComfyUI-API-Optimizer/workflows/prompt_studio_styled_generation.json.

---

4. Gemini Video Omni (Direct API)

Video generation and conversational video editing with Google's Gemini Omni model, using your own API key. Text, reference images, or a video go in; a 3-10 second 720p/24fps clip with audio comes out.

| Input | Type | Description | |-------|------|-------------| | prompt | STRING | Describe the scene, dialogue and sound. State the duration (3-10s) in the prompt | | model | COMBO | Gemini Omni Flash (preview) | | aspect_ratio | COMBO | 16:9 or 9:16 | | seed | INT | Re-run trigger only, output is non-deterministic | | images | IMAGE (optional) | Up to 14 batched reference images (subjects, style, first frame) | | video | VIDEO (optional) | A video to edit (max 10s). See the region note below | | interaction_id | STRING (optional) | Continue a previous turn: wire from another Video Omni node's output | | store | BOOLEAN (optional) | Keep the turn server-side so it can be chained (default: on) | | api_key | STRING (optional) | Google AI API key | | timeout_sec | INT (optional) | HTTP timeout, default 480s |

Outputs: video (VIDEO), text (STRING), interaction_id (STRING), cost_info (STRING), cache_key (STRING)

Conversational editing (the reason this node exists)

Gemini Omni is stateful. Each generation returns an interaction_id; wire it into a second Video Omni node and the second prompt edits the first video server-side, keeping the scene consistent:

Video Omni ("A 5-second clip of a woman playing violin in a park, 16:9")
     | interaction_id
     v
Video Omni ("Same video, but at sunset")
     | interaction_id
     v
Video Omni ("Replace the violin with a cello")

ComfyUI's caching makes this cheap to iterate: change only the last node's prompt and the upstream turns are cache hits at $0. Only the edit you changed is billed.

Cost per video

Billed by output tokens (about $0.10 per second of 720p video, plus a small input cost):

| Clip length | Approx. cost | |-------------|--------------| | 3s | ~$0.31 | | 6s | ~$0.61 | | 10s | ~$1.01 |

ComfyUI's built-in Gemini Video Omni node bills through the credit system at $0.146 per second. Direct API is roughly 30% cheaper, and cost_info shows the real USD amount from the API's own token counts.

Notes and limits

• Duration, dialogue and sound effects are controlled from the prompt. There is no duration widget; the model follows "a 6-second clip of ..." reliably.
• Output is fixed at 720p, 24 fps, with audio. Need frames or the audio track separately? Wire the video output into ComfyUI's core Get Video Components node.
• Region limit (EEA / Switzerland / UK): editing an uploaded video is blocked by Google in these regions. Editing a generated video through interaction_id chaining works everywhere and is the better workflow anyway.
• Reference videos shorter than 3 seconds may be silently misprocessed by the API (their limitation, the node warns in the console).
• Chained turns require store to stay on; stored interactions expire server-side after roughly 55 days.
• On first use the google-genai SDK prints a one-time "interactions API is experimental" warning. It is harmless.

---

API Key Setup

Provide your Google AI API key via one of these methods (checked in order):

1. Key file — create a gemini_api_key.txt file in your ComfyUI root directory (recommended)
2. Environment variable — set GEMINI_API_KEY in your system environment
3. Direct input — paste into the api_key field on the node (not recommended — visible in workflow JSON)

Get an API key at Google AI Studio. Billing must be enabled for image generation.

Installation

Clone this repository into your ComfyUI/custom_nodes/ directory:

cd ComfyUI/custom_nodes/
git clone https://github.com/jeremieLouvaert/ComfyUI-Gemini-Direct.git
pip install -r ComfyUI-Gemini-Direct/requirements.txt

Restart ComfyUI. Nodes appear under the Gemini Direct category.

Dependencies

• google-genai >= 2.0.0 — Google's Gen AI Python SDK. The video Interactions API requires 2.0+ (Google removed the legacy schema server-side in June 2026); if video generation returns a "legacy Interactions API schema" error, run pip install -U google-genai
• Pillow — image conversion
• PyTorch — already present in any ComfyUI installation

License

MIT